Method, device and equipment for evaluating real works and storage medium

Abstract

The application discloses a real work evaluation method and device, equipment and a storage medium, and relates to the technical field of computers. The method is applied to a real evaluation device, and the real evaluation device comprises a first camera unit, a second camera unit, a processing unit and a display unit. The method comprises the following steps: the first camera unit collects at least two images of a first real work; the second camera unit collects three-dimensional geometric information of the first real work; the processing unit determines a matching degree between the first real work and a second real work in a historical evaluation library according to the at least two images of the first real work and the three-dimensional geometric information of the first real work; and an evaluation result of the first real work is determined according to the matching degree between the first real work and the second real work; and the display unit displays the evaluation result of the first real work. The above method improves the evaluation effect of real works.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, device and storage medium for evaluating physical works. Background Technology

[0002] In creative competitions, participants submit physical works. Generally, the organizers need to conduct independent reviews of the submitted physical works to determine whether they are plagiarized, whether they are truly creative, and so on.

[0003] In related technologies, participating users upload photos of their physical works to the competition system, and the organizers compare the uploaded images with works in the database. If the similarity is too high, the work is generally considered to be plagiarized.

[0004] However, this method of independent judging based on photos of physical works uploaded by participating users has relatively poor judging results. Summary of the Invention

[0005] This application provides a method, apparatus, device, and storage medium for evaluating physical works, which can improve the evaluation effect of physical works. The technical solution provided by this application includes the following aspects.

[0006] According to one aspect of the embodiments of this application, a method for evaluating physical works is provided, applied to a physical evaluation device, the physical evaluation device comprising: a first camera unit, a second camera unit, a processing unit, and a display unit, the method comprising: The first camera unit acquires at least two images of the first physical artwork, wherein the different images are obtained from shooting angles of the first physical artwork; The second camera unit acquires the three-dimensional geometric information of the first physical work, and the three-dimensional geometric information of the first physical work is used to indicate the geometric shape of the first physical work in three-dimensional space; The processing unit determines the matching degree between the first physical work and a second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work; and determines the review result of the first physical work based on the matching degree between the first physical work and the second physical work. The display unit displays the review results of the first physical artwork.

[0007] According to one aspect of the embodiments of this application, a physical object evaluation device is provided, the physical object evaluation device comprising: a first camera unit, a second camera unit, a processing unit, and a display unit; The first camera unit is used to capture at least two images of the first physical work, wherein the different images are obtained from shooting angles of the first physical work; The second camera unit is used to acquire the three-dimensional geometric information of the first physical work, and the three-dimensional geometric information of the first physical work is used to indicate the geometric shape of the first physical work in three-dimensional space. The processing unit is configured to determine the matching degree between the first physical work and a second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work; and to determine the review result of the first physical work based on the matching degree between the first physical work and the second physical work. The display unit is used to display the review results of the first physical work.

[0008] According to one aspect of the embodiments of this application, a judging device for physical works is provided, the device comprising: The data acquisition module is used to acquire at least two images of the first physical work and the three-dimensional geometric information of the first physical work. The different images are obtained from shooting angles of the first physical work. The three-dimensional geometric information of the first physical work is used to indicate the geometric shape of the first physical work in three-dimensional space. The matching degree determination module is used to determine the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work. The result determination module is used to determine the review result of the first physical work based on the matching degree between the first physical work and the second physical work.

[0009] According to one aspect of the embodiments of this application, a computer device is provided, the computer device including a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the steps performed by the processing unit in the above-described method for reviewing physical works.

[0010] According to one aspect of the present application, a computer-readable storage medium is provided, wherein a computer program is stored in the storage medium, the computer program being loaded and executed by a processor to implement the steps performed by the processing unit in the above-described method for reviewing physical works.

[0011] According to one aspect of the embodiments of this application, a computer program product is provided, the computer program product including a computer program, which is loaded and executed by a processor to implement the steps performed by the processing unit in the above-described method for reviewing physical works.

[0012] The technical solutions provided in this application embodiment may have the following beneficial effects: First, this application acquires at least two images of the first physical work to be reviewed using a first camera unit, and acquires the three-dimensional geometric information of the first physical work using a second camera unit. The processing unit determines the matching degree between the first physical work and the reviewed second physical work based on the images and geometric information, thereby determining and displaying the review result of the first physical work. Compared to related technologies that rely solely on photos uploaded by participating users to determine the independence of a work, this application acquires images and three-dimensional geometric information of the physical work itself. Because the acquired data is multi-dimensional (images taken from at least two shooting angles and three-dimensional geometric information indicating the geometric shape of the physical work), the rich data source improves the effectiveness of the review of the physical work, rather than limiting the review to a single image, as in related technologies.

[0013] Furthermore, the method for evaluating physical works proposed in this application integrates the processes of image acquisition, information collection, evaluation, and result display into a single physical evaluation device. This eliminates the need for users to upload their own work, improving both the evaluation effectiveness and efficiency. Moreover, this integrated evaluation method also ensures the security and privacy of the evaluation process for physical works. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of a computer system provided in one embodiment of this application; Figure 2 This is a schematic diagram of a physical evaluation device and a physical work evaluation method provided in one embodiment of this application; Figure 3 This is a flowchart of a method for evaluating physical works provided in one embodiment of this application; Figure 4 This is a flowchart of a method for evaluating physical works provided in another embodiment of this application; Figure 5 This is a flowchart of a method for evaluating physical works provided in another embodiment of this application; Figure 6 This is a flowchart of a method for evaluating physical works provided in another embodiment of this application; Figure 7 This is an architecture diagram of a physical evaluation system provided in one embodiment of this application; Figure 8 This is a schematic diagram of the processing flow of module M1 provided in one embodiment of this application; Figure 9 This is a schematic diagram of the processing flow of module M2 provided in one embodiment of this application; Figure 10 This is a schematic diagram of the processing flow of module M3 provided in one embodiment of this application; Figure 11 This is a schematic diagram of the processing flow of modules M4 and M5 provided in one embodiment of this application; Figure 12 This is a schematic diagram of the processing flow of module M6 provided in one embodiment of this application; Figure 13 This is a block diagram of a physical evaluation apparatus provided in one embodiment of this application; Figure 14 This is a block diagram of a physical work judging device provided in one embodiment of this application; Figure 15 This is a structural block diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0016] Please refer to Figure 1 This illustration shows a schematic diagram of a computer system provided in one embodiment of this application. The computer system may include: a terminal device 10 and a server 20.

[0017] Terminal device 10 includes, but is not limited to, mobile phones, tablets, smart voice interaction devices, game consoles, wearable devices, multimedia playback devices, PCs (Personal Computers), in-vehicle terminals, smart home appliances, and other electronic devices. A client application for the target application can be installed on terminal device 10. Optionally, the target application can be an application that requires downloading and installation, or it can be an application that can be used instantly; this embodiment of the application does not limit this.

[0018] In this application embodiment, the target application is an application capable of reviewing physical works. Exemplarily, the target application may be a review application, a judgment application, a large-scale model dialogue application, etc., and this application does not limit it to this.

[0019] In some embodiments, the target application runs on a processing unit on the terminal device 10. That is, a client running the target application runs on the processing unit. Optionally, the target application is used to acquire at least two images of the first physical work and the three-dimensional geometric information of the first physical work, wherein the different images are taken from different shooting angles of the first physical work, and the three-dimensional geometric information of the first physical work is used to indicate the geometric shape of the first physical work in three-dimensional space; based on the at least two images of the first physical work and the three-dimensional geometric information of the first physical work, determine the degree of matching between the first physical work and a second physical work in the historical review database; and based on the degree of matching between the first physical work and the second physical work, determine the review result of the first physical work.

[0020] Server 20 is used to provide backend services for the client of the target application in terminal device 10. For example, server 20 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, but it is not limited to these.

[0021] Terminal device 10 and server 20 can communicate with each other via a network. This network can be a wired network or a wireless network.

[0022] The method provided in this application embodiment can be executed by a computer device, which refers to an electronic device capable of data calculation, processing, and storage. The computer device can be... Figure 1 The terminal device 10 can also be a server 20.

[0023] In some embodiments, the computer device includes a physical evaluation apparatus, which includes a first camera unit, a second camera unit, a processing unit, and a display unit.

[0024] Optionally, the first camera unit is used to acquire at least two images of the first physical artwork. Optionally, the second camera unit is used to acquire three-dimensional geometric information of the first physical artwork.

[0025] Optionally, the processing unit is configured to acquire at least two images of the first physical work captured by the first camera unit, and acquire three-dimensional geometric information of the first physical work captured by the second camera unit; determine the matching degree between the first physical work and the second physical work in the historical review database based on the at least two images of the first physical work and the three-dimensional geometric information of the first physical work; and determine the review result of the first physical work based on the matching degree between the first physical work and the second physical work.

[0026] Optionally, a display unit is provided to display the judging results of the first physical work.

[0027] The method for evaluating physical works proposed in this application is applied to a physical evaluation device, but the core evaluation method (i.e., the steps related to determining the evaluation result) is mainly executed by the processing unit in the physical evaluation device. In the following embodiments, the evaluation method for physical works is explained as follows, with each unit in the physical evaluation device as the executing entity. Of course, the executing entity of the evaluation method for physical works in this application may also include only the processing unit in the physical evaluation device. This application does not limit the executing entity of each step.

[0028] In some embodiments, such as Figure 2 As shown in sub-figure a, this is a structural schematic diagram of a physical evaluation device. Figure 2 As shown, the physical evaluation device includes the following modules: a top camera, a bottom camera, a left camera, a right camera, a rear camera, a front camera (not shown in the figure), a display screen, and a hexagonal chamber. Optionally, the physical evaluation device also includes a motorized door and a planar structured light projector.

[0029] The top camera, bottom camera, left camera, right camera, rear camera, and front camera are used to capture images of the physical artwork from six shooting angles, resulting in a six-view image. Optionally, the top camera, bottom camera, left camera, right camera, rear camera, and front camera constitute a camera array. The first camera unit mentioned in other embodiments of this application can also be considered as the camera array here, and the second camera unit can also be considered as including a planar structured light projector and two cameras in the camera array (such as the top camera and the bottom camera).

[0030] The display screen is used to show the judging results of the physical artwork. The hexagonal chamber is used to place the physical artwork. Optionally, the hexagonal chamber can be opened via a single door. After the physical artwork is placed in the hexagonal chamber, the door can be closed, and the judging of the physical artwork can begin. The display unit mentioned in other embodiments of this application can also be considered as the display screen here.

[0031] The processing units mentioned in other embodiments of this application are not shown. Figure 2 In sub-figure a, the processing unit can be placed below the hexagonal chamber included in the physical review device.

[0032] In some embodiments, such as Figure 2 Sub-figure b illustrates a physical object review method. A first camera unit acquires at least two images of the physical object (e.g., six-view images), and a second camera unit acquires the 3D reconstruction information of the physical object (e.g., a 3D point cloud, or a global 3D point cloud). Optionally, the processing unit calculates the matching degree between the two images based on previously reviewed physical objects in a historical review database to review the physical object. Optionally, the review result includes work judgment information. Work judgment information includes three categories: passed, reviewed again, and failed.

[0033] If the matching degree is high (greater than or equal to the first threshold), the physical work is deemed unacceptable. If the matching degree is relatively low (greater than or equal to the second threshold, but less than the first threshold), the physical work is deemed to require review. If the matching degree is low (less than the second threshold), the physical work is deemed acceptable.

[0034] If the physical artwork is deemed to have passed the evaluation, the historical evaluation database is updated using the physical artwork, such as updating the corresponding evaluation graph. If the physical artwork is deemed to have passed the evaluation, the artwork scoring model is used to score the physical artwork using AI (Artificial Intelligence).

[0035] like Figure 2 The physical review device shown in sub-figure a is an integrated offline review and scoring cabin, designed for scenarios such as offline creative work self-service review stations for game applications, game marketing event pop-ups, offline art and design competition self-service review, and collection figure appraisal.

[0036] Optionally, the physical judging device has a transparent acrylic shell, with optional dimensions of 60cm×60cm×50cm. The top is equipped with a 15.6-inch touch screen for displaying the judgment results, a five-dimensional scoring radar chart, and historical similar samples. Both the front and back are equipped with electric doors, one side for players to put in their works and the other side to open for retrieval after the judgment and scoring. The design of the chamber is compatible with multiple scenarios such as the Golden Shovel competition, handicraft competitions, and art creation judging.

[0037] Optionally, the physical evaluation device has a camera array, specifically, one 1200W pixel industrial camera is deployed on each of the six sides (top, bottom, left, right, front, and back) to simultaneously capture images of the physical object from six perspectives; the bottom camera shoots upwards through the transparent tempered glass support platform to completely eliminate the blind spot on the bottom surface; another set of structured light projectors and corresponding stereo cameras are set at the top to generate a global 3D (Three Dimensions) point cloud of the work.

[0038] Optionally, participants place their work on the platform inside the cabin → the cabin door closes → the physical judging device sequentially triggers six-view simultaneous shooting and structured light scanning (total approximately 1.5 seconds) → the edge computing module completes multi-dimensional feature extraction, lineage map retrieval, and grading judgment within 3 seconds → for works deemed passing, downstream five-dimensional AI scoring is performed within ≤2 seconds → the screen simultaneously displays three conclusions: "Pass / Manual Review / Rejection," similar historical samples, and a five-dimensional scoring radar chart; if passing, the work's multi-dimensional features and five-dimensional score are entered into the database and simultaneously added to the weakly labeled sample pool of the aesthetic alignment dataset. Optionally, "Rejection" means failing the assessment.

[0039] Optionally, for works judged as "passing", the screen simultaneously displays their ancestor nodes, similar sibling nodes, and a five-dimensional scoring radar chart (creativity / thematic expression / visual expression / craftsmanship / fun) in the graph, enhancing the interpretability of the review and scoring; works judged as "manual review" are automatically entered into the offline judges' review queue and the top-3 similar samples are displayed simultaneously; for works judged as "rejected", the cabin door remains closed, the cabin voice broadcast prompts and notifies the on-site staff for handling.

[0040] Optionally, the physical judging process based on the physical judging device is completed entirely offline within the all-in-one machine, without the need for network uploads, protecting participant privacy and ensuring compatibility with offline scenarios. Furthermore, using "six-view images + 3D point clouds" as identification credentials naturally adapts to physical works. Moreover, the "database expands upon entry" mechanism allows the image database and aesthetic anchor pool to continuously grow with each competition, evolving the system's ability to identify independence and align with the competition's aesthetic sensibilities. Finally, this physical judging device can be extended to other scenarios such as art appraisal, intangible cultural heritage crafts, and marketing creativity competitions.

[0041] For details, please refer to the explanations in the following embodiments, which will not be repeated here.

[0042] Please refer to Figure 3 The diagram illustrates a flowchart of a method for evaluating physical works according to an embodiment of this application. This method may include at least one of the following steps (310-340). The method is applied to a physical evaluation device, which includes a first camera unit, a second camera unit, a processing unit, and a display unit.

[0043] Step 310: The first camera unit acquires at least two images of the first physical work, the different images being taken from different shooting angles of the first physical work.

[0044] First, the physical judging device will be explained. This physical judging device is a real, physical device used to judge physical works. Optionally, the physical judging device includes at least one of the following units: a first camera unit, a second camera unit, a processing unit, and a display unit. The first camera unit is used to acquire images of the physical work to be judged; the second camera unit is used to acquire the three-dimensional geometric information of the physical work to be judged; the processing unit is used to compare the similarity between the physical work to be judged and previously judged physical works based on the acquired images and three-dimensional geometric information, determine the matching degree between the two, and finally judge the physical work to be judged based on the matching degree, obtaining the judging result of the physical work. The display unit is used to display the judging result of the physical work.

[0045] Optionally, the physical judging device can consist entirely of four units: a first camera unit, a second camera unit, a processing unit, and a display unit, enabling integrated judging. Optionally, the physical judging device can be placed at a fixed location during offline competitions, allowing participating users to come to the competition venue and place their physical works within the device for judging. Optionally, each participating user is provided with a physical judging device, allowing them to judge their own physical works at home.

[0046] Optionally, the physical evaluation device can be divided into two parts (i.e., a non-integrated device). The first part consists of a first camera unit, a second camera unit, and a display unit, while the second part is a processing unit. The first part can be deployed on the user side (the participant's side), and the second part can be deployed on the competition side (or also on the user's side). That is, the participant can use the first camera unit to capture images of the first physical artwork and the second camera unit to capture the three-dimensional geometric information of the first physical artwork. Optionally, the first camera unit sends the captured images to the processing unit, and the second camera unit also sends the captured three-dimensional combined information to the processing unit. The processing unit determines the evaluation result based on the images and the three-dimensional geometric information and then sends the evaluation result to the display module. Finally, the participant can see the displayed evaluation result through the display unit. Based on this, participants can conduct physical evaluations at home, which also improves the efficiency of physical artwork evaluation. Optionally, the aforementioned first and second parts can be a parent-child device, with each first part connected to the second part. This connection can be wired or wireless, and this application does not limit this.

[0047] The specific steps and procedures are explained below.

[0048] In some embodiments, the first camera unit acquires at least two images of the first physical work, the different images being taken from different shooting angles. The first physical work is the physical work to be reviewed. Here, the physical work can be considered a real, existing work, which can be a work printed using 3D printing, a work handmade by the user, or a work created using tools. Optionally, this application does not limit the specific type of the physical work; the type of the physical work can be at least one of the following: sculpture, paper-cutting, toy, clothing, etc.

[0049] Optionally, the first camera unit is a unit for image acquisition. Optionally, the first camera unit includes at least two cameras, each camera being used to photograph the first physical work to obtain an image of the physical work. Optionally, each camera is deployed at different locations to photograph the physical work from different shooting angles. Optionally, the first camera unit includes six cameras, respectively deployed above, below, left, right, front, and back of the physical work, to capture a top view, bottom view, left view, right view, front view, and rear view of the first physical work, that is, to obtain six images of the physical work. This application does not limit the specific number of cameras included in the first camera unit, such as the first camera unit including one camera or two cameras, the first camera unit including at least one camera, or the first camera unit including at least two cameras.

[0050] In some embodiments, the physical judging device further includes a physical judging chamber for placing the physical artwork to be judged. Optionally, the first camera unit is fixed according to the position of the physical judging chamber. Optionally, the first camera unit includes six cameras deployed at six points: above, below, left, right, front, and back of the physical judging chamber. Since the position of the physical judging chamber is fixed, that is, the position of the physical artwork is fixed, six images of the physical artwork from different shooting angles can be captured by the cameras at these six points. Optionally, the physical judging chamber corresponds to a cuboid device, with a camera placed on different faces of the cuboid to form the first camera unit. When the participating user places the first physical artwork in the physical judging chamber, at least two images are captured by the first camera unit. Of course, the first camera unit may also include only one camera, which is a movable camera, and by moving the camera, at least two images of the physical artwork taken from different shooting angles can be captured.

[0051] Step 320: The second camera unit acquires the three-dimensional geometric information of the first physical artwork. The three-dimensional geometric information of the first physical artwork is used to indicate the geometric shape of the first physical artwork in three-dimensional space.

[0052] In some embodiments, the second camera unit is a unit for acquiring three-dimensional geometric information. Optionally, the second camera unit includes at least one camera. Optionally, the camera included in the first camera unit and the camera included in the second camera unit can be the same camera or different cameras; this application does not limit this. Optionally, taking an example where the first camera unit includes six cameras, at least one of the six cameras can also be used as the second camera unit to acquire the three-dimensional geometric information of the first physical work. That is, for the same camera, both image acquisition and three-dimensional geometric information acquisition are achieved, realizing dual-purpose use of one device, thereby reducing device design costs and improving device utilization.

[0053] In some embodiments, three-dimensional geometric information is information used to indicate the geometric shape of a physical work in three-dimensional space. Optionally, the information type of the three-dimensional geometric information includes at least one of the following: point cloud information, texture information, and depth information. The three-dimensional geometric information in this application should be interpreted broadly; any information used to determine the geometric shape of a physical work in three-dimensional space can be considered as three-dimensional geometric information.

[0054] Optionally, the point cloud information of the first physical object is a set of points used to represent the first physical object in three-dimensional space. Optionally, the texture information of the first physical object is a texture map (such as a 3D texture map) of the first physical object used to represent the surface of the object in three-dimensional space. Optionally, the depth information of the first physical object is a combination of information reflecting the distances from various positions in the first physical object in three-dimensional space to the second camera unit. Based on this depth information of the first physical object, the geometry of the first physical object in three-dimensional space can be calculated.

[0055] Optionally, the second camera unit includes at least one of the following types of cameras: a stereo vision camera, a structured light camera, or a laser line scan camera. Based on these cameras, the three-dimensional geometric information of the first physical artwork can be directly acquired. The specific acquisition process will not be elaborated upon.

[0056] Step 330: The processing unit determines the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work; and determines the review result of the first physical work based on the matching degree between the first physical work and the second physical work.

[0057] In some embodiments, the first camera unit sends at least two images of the first physical work to the processing unit, and the processing unit receives the at least two images of the first physical work.

[0058] In some embodiments, the second camera unit sends the acquired three-dimensional geometric information of the first physical work to the processing unit, and the processing unit receives the three-dimensional geometric information of the first physical work.

[0059] In some embodiments, the processing unit is a module in the physical review apparatus used to implement specific review steps. Optionally, the processing unit runs a client of the aforementioned target application. Optionally, the processing unit has a pre-set review algorithm (designed by the developer). Optionally, the processing unit deploys a neural network model for implementing the review. Optionally, a first camera unit is connected to the processing unit (wired or wireless connection). Optionally, a second camera unit is connected to the processing unit. Optionally, the processing unit is connected to a display unit.

[0060] The specific review process will be explained below.

[0061] First, the processing unit determines the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work.

[0062] Optionally, the historical review database can also be considered a database used to store relevant data of reviewed physical works. The historical review database includes a collection of reviewed physical works. Optionally, the historical review database includes at least one reviewed physical work. Optionally, a second physical work is also a reviewed physical work. In some embodiments, the historical review database can be collected based on the review results of historical review activities (such as online competitions, offline competitions, creative activities, scoring activities, etc.). Optionally, the historical review database stores at least one of the following: images and three-dimensional geometric information of each reviewed physical work, and the review results for that reviewed physical work.

[0063] Optionally, the processing unit determines the matching degree between the first physical work and the second physical work based on at least two images and three-dimensional geometric information of the first physical work. Optionally, for each reviewed physical work (second physical work) in the historical review database, the matching degree between the first physical work and that reviewed physical work is calculated separately. Optionally, a portion of reviewed physical works (such as the N second physical works mentioned below; the specific selection process can be referred to in the following embodiments) are selected from the historical review database, and the matching degree between the first physical work and the reviewed physical works is calculated separately for this portion of reviewed physical works.

[0064] Optionally, the matching degree between the first physical work and the second physical work can also be considered as an indication of the matching result between the first physical work and the second physical work. Optionally, the matching degree between the first physical work and the second physical work is used to indicate the degree of similarity (similarity) between the first physical work and the second physical work. Optionally, the higher the similarity between the first physical work and the second physical work, the more matched they are. Conversely, the lower the similarity between the first physical work and the second physical work, the less matched they are.

[0065] Optionally, the processing unit determines the matching degree between the first physical artwork and the second physical artwork directly using a matching classification model, based on at least two images and 3D geometric information of the first physical artwork and at least two images and 3D geometric information of the second physical artwork. Optionally, the matching classification model is a trained neural network model used to determine whether they match. Optionally, the input to the matching classification model includes at least two images and 3D geometric information of the first physical artwork and at least two images and 3D geometric information of the second physical artwork, and the output of the matching classification model includes the matching degree between the first physical artwork and the second physical artwork. Optionally, the matching classification model can directly determine the matching degree using a binary classification method, that is, the matching degree can include: matching and non-matching.

[0066] Optionally, the processing unit determines the matching degree between the first physical work and the second physical work based on at least two images and three-dimensional geometric information of the first physical work and at least two images and three-dimensional geometric information of the second physical work. Optionally, the processing unit determines the similarity between the first physical work and the second physical work based on at least two images and three-dimensional geometric information of the first physical work and at least two images and three-dimensional geometric information of the second physical work; and determines the matching degree between the first physical work and the second physical work based on the similarity between them.

[0067] Optionally, the processing unit determines the image similarity between the first physical work and the second physical work based on at least two images of the first physical work and at least two images of the second physical work. Optionally, the processing unit determines the shape similarity between the first physical work and the second physical work based on the three-dimensional geometric information of the first physical work and the three-dimensional geometric information of the second physical work. Optionally, the similarity between the first physical work and the second physical work is determined based on the image similarity and shape similarity; and the matching degree between the first physical work and the second physical work is determined based on the similarity between them.

[0068] For example, at least two images of the first physical work are directly image-coded to obtain the image features of the first physical work, and at least two images of the second physical work are directly image-coded to obtain the image features of the second physical work. The similarity between the image features of the first physical work and the image features of the second physical work is calculated to obtain the image similarity between the first physical work and the second physical work.

[0069] For example, the three-dimensional geometric information of the first physical work is directly encoded to obtain the information encoding features of the first physical work, and the three-dimensional geometric information of the second physical work is directly encoded to obtain the information encoding features of the second physical work. The similarity between the information encoding features of the first physical work and the information encoding features of the second physical work is calculated to obtain the shape similarity between the first physical work and the second physical work.

[0070] For example, the similarity between the first physical work and the second physical work can be obtained by directly weighting and summing the image similarity and shape similarity.

[0071] For example, the similarity between the first physical work and the second physical work can be directly used as the matching degree between the first physical work and the second physical work.

[0072] For details on the specific calculation method of the matching degree, please refer to the explanation of the following embodiments, which will not be repeated here.

[0073] Secondly, the processing unit determines the review result of the first physical work based on the matching degree between the first physical work and the second physical work.

[0074] Optionally, if the matching degree is numerical, it can be divided according to a threshold. If the matching degree is greater than or equal to the threshold, it is considered to be a high matching degree. If the first physical work and the second physical work are similar, the review result includes a judgment of failure (or rejection, i.e., suspected plagiarism). If the matching degree is less than the threshold, it is considered to be a low matching degree. If the first physical work and the second physical work are not similar, the review result includes a judgment of approval (or independence).

[0075] For example, if a second physical work similar to the first physical work exists in the historical review database, the review result is determined to include a rejection. For example, if no second physical work similar to the first physical work exists in the historical review database, the review result is determined to include a pass.

[0076] Optionally, if the matching degree indicates that the first physical work and the second physical work are a match, the review result includes a rejection. If the matching degree indicates that the first physical work and the second physical work are not a match, the review result includes a pass.

[0077] For example, if a second physical work matching the first physical work exists in the historical review database, the review result is determined to include a rejection. For example, if no second physical work matching the first physical work exists in the historical review database, the review result is determined to include a pass.

[0078] For details on the specific method for determining the review results, please refer to the explanation of the following embodiments, which will not be repeated here.

[0079] Step 340: The display unit shows the review results of the first physical work.

[0080] In some embodiments, the processing unit sends the review result to the display unit, which receives and displays the review result. In some embodiments, the display unit is a module in the physical review apparatus used for display. Optionally, the display unit may be a display screen.

[0081] First, this application acquires at least two images of the first physical work to be reviewed using a first camera unit, and acquires the three-dimensional geometric information of the first physical work using a second camera unit. The processing unit determines the matching degree between the first physical work and the reviewed second physical work based on the images and geometric information, thereby determining and displaying the review result of the first physical work. Compared to related technologies that rely solely on photos uploaded by participating users to determine the independence of a work, this application acquires images and three-dimensional geometric information of the physical work itself. Because the acquired data is multi-dimensional (images taken from at least two shooting angles and three-dimensional geometric information indicating the geometric shape of the physical work), the rich data source improves the effectiveness of the review of the physical work, rather than limiting the review to a single image, as in related technologies.

[0082] Furthermore, the method for evaluating physical works proposed in this application integrates image acquisition, information collection, evaluation, and result display into a single physical evaluation device. This eliminates the need for users to upload their work manually, improving the evaluation effect while saving unnecessary steps and increasing efficiency. Moreover, this integrated evaluation method also ensures the security and privacy of the evaluation process for physical works.

[0083] The following explanation addresses the determination of the matching degree and other related matters.

[0084] Please refer to Figure 4The diagram illustrates a flowchart of a method for evaluating physical works according to another embodiment of this application. This method may include at least one of the following steps (410-460). The method is applied to a physical evaluation apparatus, which includes a first camera unit, a second camera unit, a processing unit, and a display unit.

[0085] Step 410: The first camera unit acquires at least two images of the first physical work, the different images being taken from different shooting angles of the first physical work.

[0086] Step 420: The second camera unit acquires the three-dimensional geometric information of the first physical artwork. The three-dimensional geometric information of the first physical artwork is used to indicate the geometric shape of the first physical artwork in three-dimensional space.

[0087] Step 430: The processing unit determines the geometric similarity, appearance similarity, and style similarity between the first physical work and the second physical work based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work.

[0088] Among them, geometric similarity is used to indicate the similarity between the first physical work and the second physical work in the geometric dimension, appearance similarity is used to indicate the similarity between the first physical work and the second physical work in the appearance dimension, and style similarity is used to indicate the similarity between the first physical work and the second physical work in the style dimension.

[0089] In some embodiments, a multi-output similarity calculation model is pre-trained. Optionally, the input to the similarity calculation model is at least two images and 3D geometric information of a first physical artwork and at least two images and 3D geometric information of a second physical artwork, and the output is the geometric similarity, appearance similarity, and style similarity between the first physical artwork and the second physical artwork. Optionally, the similarity calculation model is a neural network model for calculating geometric similarity, appearance similarity, and style similarity. Optionally, the similarity calculation model includes a first feature encoding layer, a second feature encoding layer, a feature fusion layer, a first feature decoding layer, a second feature decoding layer, and a third feature decoding layer.

[0090] Optionally, the first feature encoding layer is used to perform image encoding on at least two images of the first physical artwork to obtain first image features. Optionally, the first feature encoding layer is used to perform image encoding on at least two images of the second physical artwork to obtain second image features.

[0091] Optionally, the second feature encoding layer is used to encode the three-dimensional geometric information of the first physical artwork to obtain a first information feature. Optionally, the second feature encoding layer is used to encode the three-dimensional geometric information of the second physical artwork to obtain a second information feature.

[0092] Optionally, the feature fusion layer is used to fuse the first image features, the second image features, the first information features, and the second information features to obtain fused features.

[0093] Optionally, the first feature decoding layer is used to perform geometric decoding on the fused features to obtain the geometric similarity between the first physical work and the second physical work.

[0094] Optionally, the second feature decoding layer is used to perform appearance decoding on the fused features to obtain the appearance similarity between the first physical work and the second physical work.

[0095] Optionally, the third feature decoding layer is used to perform style decoding on the fused features to obtain the style similarity between the first physical work and the second physical work.

[0096] Of course, geometric similarity, appearance similarity and style similarity can also be calculated in the manner described in the following embodiments. This application does not limit the specific calculation method of geometric similarity, appearance similarity and style similarity.

[0097] Step 440: The processing unit determines the matching degree between the first physical work and the second physical work based on geometric similarity, appearance similarity and style similarity.

[0098] In some embodiments, the processing unit performs a weighted summation of geometric similarity, appearance similarity, and style similarity to determine the matching degree between the first physical work and the second physical work.

[0099] In some embodiments, the processing unit performs weighted summation after normalizing geometric similarity, appearance similarity, and style similarity to determine the matching degree between the first physical work and the second physical work.

[0100] In some embodiments, the processing unit determines the weights of geometric similarity, appearance similarity, and style similarity based on their respective priorities. Optionally, a weighted summation is performed on the geometric similarity, appearance similarity, and style similarity according to their respective weights to obtain the matching degree between the first physical work and the second physical work.

[0101] In some embodiments, the processing unit determines the matching degree between the first physical work and the second physical work by taking the maximum value among geometric similarity, appearance similarity, and style similarity.

[0102] Of course, the matching degree between the first physical work and the second physical work can also be determined in accordance with the method in the following embodiments. This application does not limit the specific calculation method of the matching degree.

[0103] Step 450: The processing unit determines the review result of the first physical work based on the matching degree between the first physical work and the second physical work.

[0104] Step 460: The display unit shows the review results of the first physical work.

[0105] For technical details not mentioned in the above steps, please refer to the explanations of other embodiments; they will not be repeated here.

[0106] The technical solution provided in this application determines the matching degree between a first physical work and a second physical work by calculating geometric similarity, appearance similarity, and style similarity respectively. Based on this, the determination of the matching degree is no longer limited to a single dimension, but rather combines similarity judgments across all dimensions. Therefore, this application improves the accuracy of the determined matching degree, thereby enhancing the review effect.

[0107] like Figure 7 As shown in 700, the overall technical architecture of this application consists of five layers: an acquisition layer 710, a feature layer 720, a map layer 730, a decision layer 740, and an evaluation layer 750. Optionally, the overall technical architecture of this application includes module M1, used to realize six-view + structured light multi-dimensional acquisition. Optionally, module M1 corresponds to the first camera unit and the second camera unit in the aforementioned physical evaluation device.

[0108] like Figure 8 As shown in Figure 800, module M1 (not shown) is responsible for the hardware-level synchronous acquisition of multi-dimensional data within the cabin. Six industrial cameras are uniformly driven by the same FPGA (Field Programmable Gate Array) trigger board, with a hardware trigger error of <1ms. Six-view image acquisition is completed within one frame after the artwork is stationary, with a resolution of 4000×3000 and JPEG (Joint Photographic Experts Group) output. A structured light projector continuously projects seven frames of Gray-Code coded stripes within 1.6ms. A stereo camera simultaneously captures and reconstructs the image, obtaining a dense 3D point cloud of the artwork with a depth accuracy ≤0.1mm. The six-view RGB images and the 3D point cloud are jointly registered and aligned using ICP + color consistency, and finally output in the form of "6 RGB images + 1 PLY (Polygon File Format) point cloud + acquisition timestamp," serving as the artwork's identity credential and shared by modules M2 and M6.

[0109] This corresponds to the aforementioned process where the first camera unit acquires at least two images (i.e., six-view images) of the first physical artwork, and the second camera unit acquires the three-dimensional geometric information (i.e., 3D point cloud) of the first physical artwork. Optionally, the overall technical architecture of this application includes module M2 for implementing multi-path independent similarity extraction. Optionally, module M2 corresponds to the processing unit in the physical object review device.

[0110] like Figure 9 As shown, module M2 (not shown in the figure) extracts three independent similarity signals from the multidimensional feature set of each artwork. Geometric similarity is determined via geometric path 910, appearance similarity via appearance path 920, and style similarity via style path 930. That is, the processing unit determines the matching degree between the first physical artwork and the second physical artwork based on at least two images and three-dimensional geometric information of the first physical artwork. Optionally, the processing unit first determines the geometric similarity, appearance similarity, and style similarity between the first physical artwork and the second physical artwork based on at least two images and three-dimensional geometric information of the first physical artwork. Optionally, the processing unit then determines the matching degree between the first physical artwork and the second physical artwork based on the geometric similarity, appearance similarity, and style similarity.

[0111] Optionally, the input to module M2 is the raw data collected by module M1—point cloud. With a collection of six-view RGB images The unified preprocessing is as follows: H (height) = W (width) = 224 (ImageNet mean and variance normalization). N represents the total number of points in the point cloud.

[0112] Optionally, module M2 outputs: three atomic signals. The candidate works and the existing nodes n in the lineage graph are measured from the dimensions of geometry, appearance, and style, respectively. The degree of similarity is used by downstream modules to determine independence. For geometric similarity, For appearance similarity, For style similarity.

[0113] The following is an explanation of how similarity is calculated.

[0114] In some embodiments, the processing unit performs voxelization processing on the first physical work based on the three-dimensional geometric information of the first physical work to obtain the voxelization processing result of the first physical work; and determines the geometric similarity between the first physical work and the second physical work based on the voxelization processing result of the first physical work and the voxelization processing result of the second physical work.

[0115] Optionally, voxelization is the process of converting a 3D object or 3D space into a series of small cubic units (i.e., voxels). Optionally, the 3D reconstruction information includes point cloud information (or 3D point cloud information). The processing unit performs voxelization on the first physical object based on the point cloud information of the first physical object, obtaining the voxelization result of the first physical object.

[0116] Optionally, the voxelization result of the second physical work can be stored in the historical review database in advance, or it can be calculated in real time; this application does not impose any restrictions on this.

[0117] In one possible implementation, the processing unit determines the geometric similarity between the first physical work and the second physical work based on the voxelization results of the first physical work and the second physical work. Optionally, the processing unit performs similarity calculations on the voxelization results of the first physical work and the second physical work to determine the geometric similarity between the two physical works.

[0118] In another possible implementation, the processing unit calculates the similarity between the voxelization results of the first physical work and the second physical work to obtain a first geometric similarity between the two physical works. Optionally, the processing unit calculates the similarity between the point cloud information of the first physical work and the point cloud information of the second physical work to obtain a second geometric similarity between the two physical works. Optionally, the processing unit performs a weighted sum of the first and second geometric similarities to obtain a geometric similarity between the first and second physical works.

[0119] In some embodiments, the processing unit directly calculates the Hausdorff distance, chamfer distance, etc. between the first physical work and the second physical work based on the three-dimensional geometric information (such as point cloud information) of the objects of the first physical work and the second physical work, respectively, to determine the geometric similarity between the first physical work and the second physical work.

[0120] (1) Geometric similarity The calculation process (purely geometric calculation, without involving neural networks) is as follows: After normalizing the point cloud P to a unit cube, voxelize it at a resolution of R=128 to obtain the voxelized result:

[0121] by voxels of candidate nodes (candidate nodes in the review graph, equivalent to the second physical node in the historical review database mentioned above) Calculate Voxel-IoU as follows:

[0122] Simultaneously calculate the bidirectional Chamfer Distance (CD) of the original point cloud:

[0123] in, This is equivalent to the three-dimensional geometric information of the first physical work mentioned above. This is equivalent to the three-dimensional geometric information of the second physical work mentioned above.

[0124] Ultimately, for and After weighting, the geometric similarity is obtained as follows:

[0125] in, The function restricts the calculated result to the interval [0,1]. If the result is less than 0, the output is 0; if it is greater than 1, the output is 1. This ensures that the final similarity value is a valid score normalized to the interval [0,1]. The function is a hyperbolic tangent function, used to non-linearly map the chamfer distance (CD) value to the interval [0,1]. This prevents excessively large CD values ​​from affecting the results, making... It is more sensitive to changes in CD values ​​within a certain range, but less sensitive to large changes outside the range. It is a normalization constant, such as 0.05 (an empirical scale under a unit cube).

[0126] In some embodiments, the processing unit employs an appearance encoder to perform feature encoding on at least two images of the first physical work to obtain appearance encoding features of the first physical work; and determines the appearance similarity between the first physical work and the second physical work based on the appearance encoding features of the first physical work and the appearance encoding features of the second physical work; wherein, the appearance encoder is a neural network model used to encode appearance features of images.

[0127] Optionally, the appearance encoder is a neural network model used to encode appearance features of an image. Optionally, the appearance encoder includes an image encoder.

[0128] Optionally, the processing unit calculates the similarity between the first physical work and the second physical work based on the appearance coding features of the first physical work and the second physical work. The similarity metrics involved in this application include, but are not limited to: cosine similarity, Euclidean distance, dot product, and Manhattan distance.

[0129] Optionally, for each image taken from a different angle, the appearance similarity (i.e., single-view similarity) between the first physical work and the second physical work is calculated. Optionally, the appearance similarity of at least two photos is weighted and summed to determine the appearance similarity between the first physical work and the second physical work.

[0130] Taking one of the shooting angles as an example, the processing unit uses an appearance encoder to perform feature encoding on the image of the first physical work corresponding to that shooting angle, thereby obtaining the appearance encoding features of the first physical work; based on the appearance encoding features of the first physical work and the appearance encoding features of the second physical work, the appearance similarity between the first physical work and the second physical work is determined.

[0131] Optionally, the appearance coding features of the second physical work are obtained by feature coding at least two images of the second physical work. Optionally, the appearance coding features of the second physical work can be stored in advance in a historical review database, or they can be calculated in real time; this application does not impose any limitation on this.

[0132] In some embodiments, the processing unit directly calculates the mean square error between the images, the structural similarity between the images, the histogram difference between the images, etc., based on at least two images of the objects of the first physical work and the second physical work, respectively, to determine the appearance similarity between the first physical work and the second physical work.

[0133] Below is about appearance similarity. The calculation process of (feature extraction based on pre-trained deep convolutional neural networks) is explained as follows.

[0134] A ResNet-50 deep residual convolutional neural network, pre-trained under supervision on the ImageNet-1k dataset, is used as the backbone for visual feature extraction. The model structure is a 50-layer residual network, specifically consisting of a 7×7 convolutional backbone (Conv1, stride 2), a max-pooling layer, four residual stages (Conv2_x containing 3 Bottleneck residual blocks, Conv3_x containing 4, Conv4_x containing 6, and Conv5_x containing 3, for a total of 16 Bottleneck residual blocks, each containing 1×1, 3×3, and 1×1 convolutional layers and identity residual connections), a global average pooling layer (GAP), and a classification head. This application removes the final fully connected classification layer, retaining only the GAP output as the visual feature extractor. Its output dimension is 2048.

[0135] Optionally, by reasoning forward through each of the six viewpoint images, six sets of features are obtained:

[0136] Optionally, This represents the image features from a single viewpoint. Optionally, these features are concatenated in viewpoint order to form a global appearance vector. :

[0137] The PCA (Principal Component Analysis) projection matrix, pre-fitted onto a feature library of historical works, is used. Dimensionality reduced to 512:

[0138] in, The 512-dimensional feature vector is obtained by PCA dimensionality reduction of the global appearance vector z of the first physical work. To train the feature library mean, perform the same operation on the candidate nodes. , The global appearance vector of the candidate node (i.e., the second physical work) is reduced to a 512-dimensional feature vector by PCA.

[0139] Optionally, calculate the maximum cosine similarity from a single viewpoint:

[0140] in, This can be considered as the appearance coding feature of the single-view image of the first physical work mentioned above. This represents the appearance coding features of the single-view image of the first physical work obtained after PCA dimensionality reduction. It can be considered as the appearance coding feature of the single-view image of the second physical work mentioned above. This represents the appearance coding features of a single-view image of the second physical artwork obtained after PCA dimensionality reduction.

[0141] Here, we will not retrain or fine-tune ResNet-50, but will directly use its ImageNet pre-trained weights for inference, because ImageNet pre-trained features have been proven to have good discriminative and transferability for general visual objects, which is sufficient to support the appearance differentiation between similar creative physical works in this scenario.

[0142] In some embodiments, the processing unit employs a style encoder to perform feature encoding on at least two images of the first physical work to obtain style encoding features of the first physical work; and determines the style similarity between the first physical work and the second physical work based on the style encoding features of the first physical work and the style encoding features of the second physical work; wherein, the style encoder is a neural network model used to encode style features of images.

[0143] Optionally, the style encoder is a neural network model used to encode style features of an image. Optionally, the style encoder includes an image encoder.

[0144] Optionally, the style features include depth style features and outline style features. Optionally, based on the style coding features of the first physical work and the style coding features of the second physical work, the depth style features and outline style features between the first physical work and the second physical work are determined. Optionally, the depth style features and outline style features are weighted and summed to obtain the style similarity between the first physical work and the second physical work.

[0145] Optionally, the style coding features of the second physical work are obtained by feature coding at least two images of the second physical work. Optionally, the style coding features of the second physical work can be stored in advance in a historical review database, or they can be calculated in real time; this application does not impose any limitation on this.

[0146] In some embodiments, the processing unit directly calculates the Gram matrix of the first physical work and the second physical work, style feature embedding based on a pre-trained network, style classification model, etc., based on at least two images of the objects of the first physical work and the second physical work, respectively, to determine the style similarity between the first physical work and the second physical work.

[0147] Below is about style similarity. The calculation process of (Gram style features + contour shape descriptor based on pre-trained deep convolutional neural network) is explained as follows.

[0148] (1) Deep Style Features: A VGG-19 deep convolutional neural network pre-trained on ImageNet-1k is used as the backbone for style feature extraction. VGG-19 consists of 16 3×3 convolutional layers and 5 max pooling layers connected in series, divided into 5 convolutional segments according to the pooling layers. In this application, the ReLU activation output of the first convolutional layer of the 5 segments is selected, i.e. A total of 5 layers serve as multi-scale style representation layers. The number of channels corresponding to the 5 layers is as follows: The spatial size is halved with each pooling stage.

[0149] For images No. Layer activation (Already flattened by channel and space), its Gram matrix is ​​defined as:

[0150] in, For the number of channels, The height of the feature map, This represents the width of the feature map.

[0151] Gram matrix corresponding to candidate nodes The Frobenius distance is:

[0152] Five-layer distance weighted normalization ( For example, set as weight. , (This is a normalization constant, representing the 95th percentile value in the feature library).

[0153] The similarity of the deep style features is obtained by averaging the results from the six perspectives as follows:

[0154] (2) Contour style features: Extract the binary contour of the work from each viewpoint image, and then perform contour analysis according to curve parameters. After sampling, the Elliptic Fourier Descriptor (EFD) expansion is performed, and the first 16 harmonic coefficients are taken.

[0155] right After scaling and rotation normalization, the vectors are concatenated into a 64-dimensional EFD vector. .in, , This represents the Fourier coefficients of the k-th harmonic in the x-direction. This represents the Fourier coefficients of the k-th harmonic in the y-direction. It is related to the candidate node EFD vector. Calculating cosine similarity and averaging across six perspectives yields the following similarity scores for contour style features:

[0156] (3) The similarity of deep style features and the similarity of contour style features are weighted and fused to obtain the following style similarity:

[0157] It should be noted that there are three signals. All values ​​are normalized to [0,1] and used as input for the next module M3. Based on this, the downstream module retrieves Top-K candidate nodes in the graph and performs multi-dimensional independence decisions to achieve quantitative evaluation of the independence of candidate works and historically entered works, thus completing the closed-loop application in the offline physical creative work independence evaluation scenario.

[0158] The technical solutions provided in this application, which calculate geometric similarity, appearance similarity, and style similarity based on specific coding features, are relatively accurate. These similarities effectively reflect the degree of similarity between the first and second physical works in terms of geometry, appearance, and style, thus improving the accuracy of the determined matching scores. Furthermore, directly using a model for similarity calculation reduces computational complexity and increases calculation efficiency.

[0159] For example, the training samples for the appearance encoder and style encoder include: sample pairs consisting of physical works that passed the manual review and physical works that failed the review in the same historical review activity.

[0160] Optionally, the process for manual review can be found in the explanation of the following embodiments, which will not be repeated here.

[0161] Optionally, for the same historical judging activity (e.g., the same competition), sample pairs consisting of works that passed and those that failed the manual review are used as training samples to train the appearance encoder and style encoder. Since these are historical judging activities, this training can be considered pre-training. Alternatively, it can be considered that after each judging activity, sample pairs consisting of works that passed and those that failed the manual review in that activity are used to fine-tune the appearance encoder and style encoder.

[0162] The technical solution provided in this application uses sample pairs consisting of physical works that pass the test and those that fail to pass the test to train the appearance encoder and style encoder. This enables the appearance encoder and style encoder to better distinguish different physical works and better capture the differences between them, thereby improving the encoding effect and increasing the accuracy of determining appearance similarity and geometric similarity.

[0163] The following is an explanation of how the matching degree is calculated.

[0164] In some embodiments, the processing unit normalizes the geometric similarity, appearance similarity, and style similarity respectively to obtain the normalized processing results corresponding to each of the geometric similarity, appearance similarity, and style similarity; the processing unit performs a weighted summation of the normalized processing results corresponding to each of the geometric similarity, appearance similarity, and style similarity to obtain the matching degree between the first physical work and the second physical work.

[0165] Optionally, the processing unit normalizes the geometric similarity to obtain the normalized result corresponding to the geometric similarity.

[0166] Optionally, the processing unit normalizes the appearance similarity to obtain the normalized result corresponding to the appearance similarity.

[0167] Optionally, the processing unit normalizes the style similarity to obtain the normalized result corresponding to the style similarity.

[0168] Optionally, the processing unit performs a weighted summation of the normalized processing results corresponding to geometric similarity, appearance similarity, and style similarity to obtain the matching degree between the first physical work and the second physical work.

[0169] Optionally, the weights of each of the three dimensions are pre-set. For example, the weights of the three dimensions are 0.45 (corresponding to the geometric dimension), 0.35 (corresponding to the appearance dimension), and 0.20 (corresponding to the style dimension). Optionally, the weights of the three dimensions are fixed or may be dynamically adjusted as the review process progresses.

[0170] The technical solution provided in this application calculates the matching degree by normalizing and weighted summing. More important dimensions can be assigned higher weights, while relatively unimportant dimensions can be assigned lower weights. Based on this, a method for determining the matching degree based on specific dimensions and dimension weights is realized, improving the indicative effect of the matching degree on similarity. In addition, the normalization method can reduce the difficulty of data processing to a certain extent, thereby reducing hardware processing costs.

[0171] The following is for a set of similar works (used to reduce the number of second physical works that need to be calculated for matching scores). The following explanation is provided.

[0172] In some embodiments, the processing unit determines a first index value for the first physical work based on the voxelization result of the first physical work determined by the three-dimensional geometric information of the first physical work and the voxelization result of each second physical work in the historical review database. The first index value of the first physical work is used to indicate a first physical work set, which includes at least one second physical work that is similar to the first physical work in geometric dimensions.

[0173] For example, the first index value can be a MinHash LSH (Locality Sensitive Hashing) index. MinHash LSH is specifically designed for quickly finding objects with high Jaccard similarity on high-dimensional sparse datasets. Calculation process: The point cloud 64... 64 The voxel coordinates occupied in a 64-voxel raster are considered as set A (typically 5000~20000 voxels). The minimum hash value of each set is obtained using k=128 independent hash functions, resulting in a 128-dimensional signature vector sig(A). The proportion of identical bits between two signatures is approximately equal to the Jaccard similarity J(A,B)=|A∩B| / |A∪B|. The 128-dimensional signature is then divided into b=32 bands, each band having r=4 dimensions. Any band with identical dimensions is considered a candidate pair. The hit probability P=1-(1-J^r)^b, and P≈0.97 when J=0.6. N The N Jaccard computations are reduced to hash bucket lookups, and the time complexity of geometric path recall is reduced from O(N²) to approximately O(N).

[0174] That is, based on the voxelization results of the first physical work and the voxelization results of each second physical work in the historical review database, hash buckets are classified, and the first index value is determined according to the hash bucket where the appearance coding features of the first physical work are located.

[0175] In some embodiments, the processing unit determines a second index value for the first physical work based on the appearance coding features of the first physical work determined from at least two images of the first physical work and the appearance coding features of each second physical work in the historical review database. The second index value of the first physical work is used to indicate a set of second physical works, which includes at least one second physical work that is similar to the first physical work in appearance dimension.

[0176] For example, the second index value can be FAISS IndexIVF (Inverted File Index). Optionally, FAISS is a classic approximate nearest neighbor data structure provided by the Facebook AI Similarity Search library. The core idea of ​​IndexIVF is to first use K-Means to cluster all historical vectors into nlist clusters (e.g., nlist=1024), with each vector attached to the nearest cluster center; during a query, the distance between the query vector and the 1024 cluster centers is calculated first, and only the nearest nprobe clusters (e.g., nprobe=16) are expanded for precise comparison, while all other clusters are skipped.

[0177] That is, based on the appearance coding features of the first physical work and the appearance coding features of each second physical work in the historical review database, the second index value is determined according to the vector cluster in which the appearance coding features of the first physical work belong.

[0178] In some embodiments, the processing unit determines a third index value for the first physical work based on the style coding features of the first physical work determined from at least two images of the first physical work and the style coding features of each second physical work in the historical review database. The third index of the first physical work is used to indicate a third set of physical works, which includes at least one second physical work that is similar to the first physical work in terms of style.

[0179] For example, the third index value can be HNSW (Hierarchical Navigable Small World). Optionally, HNSW is a multi-level graph index structure. When building the index, M=16 nearest neighbor connections are inserted into the multi-level graph for each vector, forming a navigation network that is sparse at the top and dense at the bottom; during a query, a greedy descent is performed from the top-level entry point, and a priority queue is used to expand efSearch=64 neighbors to the bottom level.

[0180] That is, based on the style coding features of the first physical work and the vector similarity between the style coding features of each second physical work in the historical review database, the nearest neighbors associated with the appearance coding features of the first physical work are determined. The third index value is then determined based on the nearest neighbors associated with the appearance coding features of the first physical work.

[0181] In some embodiments, the processing unit performs a step of determining the degree of matching between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work for the second physical work in the similar works set.

[0182] The set of similar works includes the first set of physical works, the second set of physical works, and the third set of physical works. The set of similar works includes N second physical works, where N is a positive integer.

[0183] Optionally, the union of the first physical portfolio recalled based on the geometric dimension, the second physical portfolio recalled based on the appearance dimension, and the third physical portfolio recalled based on the style dimension is used to obtain a similar portfolio.

[0184] Optionally, the method of recalling the first physical portfolio by geometric dimension, the second physical portfolio by appearance dimension, and the third physical portfolio by style dimension includes, but is not limited to, the method of determining the first index value, the second index value, and the third index value mentioned above. Other similarity calculation methods can also be used. The core is to find the second physical portfolio that is similar to the first physical portfolio in each dimension. The specific determination method is not limited.

[0185] Optionally, the aforementioned three-way ANN recall refers to running three sets of indexes—FAISS IVFPQ (appearance path), MinHash LSH (geometric path), and HNSW (style path)—in parallel. Each path independently returns the 50 candidate works with the highest similarity, denoted as C_app (corresponding to the second physical work set mentioned above), C_geo (corresponding to the first physical work set mentioned above), and C_sty (corresponding to the third physical work set mentioned above). The union C (corresponding to the set of similar works mentioned above) = C_app ∪ C_geo ∪ C_sty, and the candidate size after deduplication does not exceed 150. The reason for parallel operation of the three paths is that a single-path index is only sensitive to its own strengths (the appearance path is sensitive to color composition, the geometric path is sensitive to shape structure, and the style path is sensitive to texture and brushstrokes). Parallel operation can complement each other in recall, avoiding missed recalls by a single path.

[0186] Optionally, on the candidate set recalled by the ANN, the exact calculation of the three-way similarity is strictly performed (no longer using approximate hashing or quantization compression), and then fused according to fixed weights:

[0187] Finally, output the first K=10 final matches in descending order of S.

[0188] The technical solution provided in this application only performs further refined calculations of the matching degree for the similar works selected above before conducting physical review, thereby reducing the computational load of matching degree calculations and reducing hardware resource consumption. Therefore, this method reduces the review cost of physical works and improves the review efficiency of physical works.

[0189] The following explanations are provided regarding the display of the review results.

[0190] Please refer to Figure 5 The diagram illustrates a flowchart of a method for evaluating physical works according to another embodiment of this application. This method may include at least one of the following steps (510-560). The method is applied to a physical evaluation apparatus, which includes a first camera unit, a second camera unit, a processing unit, and a display unit.

[0191] Step 510: The first camera unit acquires at least two images of the first physical work, the different images being taken from different shooting angles of the first physical work.

[0192] Step 520: The second camera unit acquires the three-dimensional geometric information of the first physical artwork. The three-dimensional geometric information of the first physical artwork is used to indicate the geometric shape of the first physical artwork in three-dimensional space.

[0193] Step 530: The processing unit determines the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work; and determines the review result of the first physical work based on the matching degree between the first physical work and the second physical work.

[0194] In some embodiments, the review results include work judgment information, which is one of the following: passed, subject to review, or failed.

[0195] Optionally, the work evaluation information refers to the result of the evaluation of the independence of the physical work. If the work evaluation information indicates that the evaluation is passed, it means that the first physical work does not involve plagiarism, is an independent work, and has creativity. If the work evaluation information indicates that the evaluation is failed, it means that the first physical work involves plagiarism, is not an independent work, and lacks creativity. If the work evaluation information indicates that the evaluation is subject to review, it means that the first physical work may involve plagiarism and needs to be further determined through manual review.

[0196] In some embodiments, the display unit displays the review results of the first physical work, including judgment information.

[0197] If the work is deemed to have passed the initial assessment, the review results also include the work score information. The work score information is determined by scoring the first physical work using a work scoring model, which is a neural network model used to score physical works.

[0198] Optionally, scoring the first physical artwork using an artwork scoring model can also be called AI scoring. Optionally, when the artwork scoring information includes scores in five dimensions, this method is also called 5D AI scoring. For the specific steps and procedures of AI scoring, please refer to the explanations in the following embodiments, which will not be repeated here.

[0199] Optionally, for the first physical work that passes the initial assessment, an AI scoring step needs to be performed. That is, the first physical work is scored using a work scoring model to obtain work score information. Optionally, this work score information is also displayed on the display unit as part of the review result.

[0200] When the work's judgment information indicates a review, the review result also includes at least one similar physical work. The at least one similar physical work is a second physical work that is similar to the first physical work and is determined from the historical review database. The first physical work and at least one similar physical work are used to review the first physical work manually.

[0201] Optionally, for the first physical work to be reviewed, a manual review process needs to be performed. Before conducting the manual review, at least one similar physical work needs to be identified. Optionally, based on the matching degree between the first physical work and each second physical work in the historical review database, the top K second physical works with the highest matching degree are selected as at least one similar physical work. Optionally, K is a positive integer.

[0202] Optionally, the first physical work and at least one similar physical work are added to the review queue for manual review, resulting in a manual review result. The specific process for manual review is not detailed here. Optionally, the manual review result may be either "pass" or "fail".

[0203] In some embodiments, if the manual review result of the first physical work, in conjunction with at least one similar physical work, is deemed to have passed, the processing unit scores the first physical work using a work scoring model to determine the work scoring information of the first work.

[0204] Optionally, for the first physical work that passes the manual review, the AI ​​scoring step can continue to be executed to obtain the work's score information. Optionally, for the first physical work that fails the manual review, the final work judgment information for the first physical work is determined to be "failed".

[0205] Step 540: The display unit displays the work judgment information of the first physical work, and if the work judgment information is deemed passed, displays the work score information of the first physical work in a graphical form; wherein, displaying the work score information in a graphical form includes: displaying the scores of multiple dimensions included in the work score information in the form of a radar chart.

[0206] Optionally, the work's rating information includes scores for at least one dimension. Optionally, the work's rating information includes scores for five dimensions. Optionally, at least one dimension includes: creativity, thematic expression, visual expression, craftsmanship, and entertainment value. Optionally, the work's rating information is obtained by scoring at least one dimension using a work rating model.

[0207] Optionally, if the work's evaluation information indicates that it has passed, the work score information for the first physical work can be displayed in radar chart format. Optionally, if the work's evaluation information indicates that it has passed, the scores for the first physical work in five dimensions can be displayed in radar chart format.

[0208] Alternatively, in addition to radar charts, the scoring information can also be displayed in tables, histograms, bar charts, etc. This application does not limit the graphical representation method.

[0209] Step 550: The display unit displays the work determination information of the first physical work, and if the work determination information is a review of the determination, it displays at least one similar physical work of the first physical work.

[0210] Optionally, in the event of a review, at least one similar physical work of the first physical work may also be displayed. Optionally, images of each of the at least one similar physical work may be displayed.

[0211] Step 560: The display unit displays the artwork judgment information of the first physical artwork, and the artwork judgment information is "judged as not passed".

[0212] Optionally, if the judgment fails, the judgment of the first physical work is directly displayed as "failed".

[0213] Technical details not mentioned in the above steps can be found in the explanations of other embodiments and will not be repeated here.

[0214] The technical solution provided in this application firstly triggers different review results based on different judgment situations of the works. Taking a re-review as an example, it is necessary to further identify similar physical works to support subsequent manual review. Taking a pass as an example, AI scoring is also used to display the work's scoring information. Based on this, the review of physical works is no longer simply limited to pass or fail. By displaying the work's scoring information, users can intuitively understand why the work was judged to pass, and by displaying similar physical works, users can also understand why the work was judged to be re-reviewed. Based on this, this application enriches the visual elements and improves the display effect of the review results. In addition, this method can attract users to actively participate in the physical review, thereby promoting human-computer interaction.

[0215] In some embodiments, if the work is deemed to have passed the review, the review result may also include: Ancestor-similar physical works are used to indicate the second physical work in the historical review database that meets the first similarity relationship with the first physical work and was reviewed earliest.

[0216] The first similarity relation here can be that the matching degree belongs to the first preset matching degree interval. The first similarity relation here can be that the matching degree is the maximum. The first similarity relation here can also be the top K matching degrees. K is a positive integer.

[0217] "Sibling similar physical works" is used to indicate a second physical work in the historical review database that meets the second similarity relationship with the first physical work and is located in the same review period as the first physical work.

[0218] The second similarity relationship here can be that the matching degree belongs to a second preset matching degree interval. The second similarity relationship here can be that the matching degree is the maximum. The second similarity relationship here can also be that the matching degree is among the top K.

[0219] For example, if the first physical work passes the evaluation, the ancestral similar physical works and sibling similar physical works of the first physical work are further found and displayed to the user as the review results.

[0220] For specific steps on determining similar physical works by ancestors and similar physical works by brothers, please refer to the explanations in the following embodiments, which will not be repeated here.

[0221] In this embodiment, if the work passes the initial evaluation, similar physical works by ancestors and similar physical works by siblings are further displayed. Similarly, the "blood relationship" between the current first physical work and physical works in the historical artifact review database is determined by matching degree. This method enriches the expression of the review results and improves the review effect. Furthermore, this method enhances the user experience in participating in the review process, attracting users to actively participate in physical artifact review, thereby promoting human-computer interaction.

[0222] The following explanations are provided regarding the determination of the review results.

[0223] Please refer to Figure 6 The diagram illustrates a flowchart of a method for evaluating physical works according to another embodiment of this application. This method may include at least one of the following steps (610-670). The method is applied to a physical evaluation device, which includes a first camera unit, a second camera unit, a processing unit, and a display unit.

[0224] Step 610: The first camera unit acquires at least two images of the first physical work, the different images being taken from different shooting angles of the first physical work.

[0225] Step 620: The second camera unit acquires the three-dimensional geometric information of the first physical artwork. The three-dimensional geometric information of the first physical artwork is used to indicate the geometric shape of the first physical artwork in three-dimensional space.

[0226] Step 630: The processing unit determines the first matching degree based on the matching degree between the first physical work and each of the N second physical works; wherein, the first matching degree is the maximum value of the matching degree between the first physical work and each of the N second physical works, and N is a positive integer.

[0227] Optionally, for each of the N second physical works in the historical review database, the matching degree between the first physical work and the second physical work is calculated. Optionally, the N second physical works here are the set of similar works recalled and determined through the similar physical works mentioned above in the three dimensions.

[0228] Optionally, the first matching degree is the maximum value of the matching degrees between the first physical work and each of the N second physical works. That is, the maximum value of the matching degrees between the first physical work and each of the N second physical works is determined as the first matching degree.

[0229] Step 640: If the first matching degree is greater than or equal to the first threshold, the processing unit determines that the work determination information of the first physical work is "determination not passed".

[0230] Optionally, the first threshold is a preset value. Optionally, the first threshold is 0.90. Optionally, the first threshold can be a fixed value or a value that is dynamically adjusted based on historical review data.

[0231] Optionally, if the first matching degree is greater than or equal to the first threshold, it indicates that the first physical work and the second physical work have a high degree of similarity. In this case, the first physical work is judged as unsuccessful, that is, the work judgment information is judged as unsuccessful.

[0232] Step 650: If the first matching degree is less than the first threshold and greater than or equal to the second threshold, the processing unit determines the work judgment information of the first physical work as a judgment review; based on the matching degree between the first physical work and N second physical works, at least one similar physical work is determined.

[0233] Optionally, the second threshold is a preset value. Optionally, the second threshold is 0.75. Optionally, the second threshold can be a fixed value or a value that is dynamically adjusted based on historical review data.

[0234] Optionally, the first and second thresholds are dynamically updated based on a preset number of physical works that have passed the judgment and the results of manual review. Optionally, the preset number is 1000. Optionally, the processing unit performs offline statistics on the historical review results for every 1000 newly added works, and adaptively fine-tunes the first and second thresholds according to the ROC (Receiver Operating Characteristic Curve) curve to avoid drift in the false positive and false negative rates.

[0235] Optionally, if the first matching degree is less than the first threshold and greater than or equal to the second threshold, it indicates that the first physical work and the second physical work have a relatively high similarity. In this case, it is determined that the first physical work needs to be manually reviewed, that is, the work judgment information is the judgment review.

[0236] Optionally, based on the matching degree between the first physical work and each of the N second physical works, at least one similar physical work is determined. Optionally, based on the matching degree between the first physical work and each of the N second physical works, the K second physical works with the highest matching degree are determined as at least one similar physical work. Optionally, K equals 3.

[0237] Optionally, at least two images and three-dimensional geometric information of the first physical work, as well as at least two images and three-dimensional geometric information of at least one similar physical work, are sent to the judges of the review activity for manual review to obtain the manual review result of the first physical work.

[0238] Step 660: If the first matching degree is less than the second threshold, the processing unit determines that the first physical work's work judgment information is qualified; the first physical work is scored by the work scoring model to determine the work scoring information of the first physical work. The work scoring model is a neural network model used to score physical works.

[0239] Optionally, if the first matching degree is less than the second threshold, it indicates that the similarity between the first physical work and the second physical work is low, and the first physical work is judged to be passed, that is, the work judgment information is judged to be passed.

[0240] Optionally, for the first physical work that passes the initial assessment, an AI scoring process is executed. Optionally, the first physical work is scored using a work scoring model to determine the work score information for the first physical work.

[0241] Step 670: The display unit shows the review results of the first physical work.

[0242] Optionally, the review results may include information on the judgment of the works.

[0243] Optionally, the overall technical architecture of this application includes module M4 for implementing hierarchical access decisions. Optionally, module M4 corresponds to the processing unit in the aforementioned physical review device.

[0244] like Figure 11 As shown, module M4 (not shown in the figure) performs a three-level decision based on the maximum comprehensive similarity of 1110 (i.e. S_max, corresponding to the first matching degree mentioned above) among the Top-K similar nodes retrieved and returned by module M3.

[0245] (1) S_max≥0.90: Directly reject, keep the hatch closed, and display the 3 most similar historical works, their submission time, and authors on the screen; do not proceed to downstream AI review. This step corresponds to the first matching degree being greater than or equal to the first threshold mentioned above. That is, if... Figure 11 As shown, for physical works with a maximum comprehensive similarity greater than or equal to the first threshold, the judgment is not passed (1140).

[0246] (2) 0.75≤S_max<0.90: Enter the manual review queue, and simultaneously push the six-view image pairs and 3D point cloud visualizations of the current work and the Top-3 similar samples to the offline judges; works that pass the review will continue to go through the M6 ​​scoring process. This step corresponds to the first matching degree being less than the first threshold and greater than or equal to the second threshold mentioned above. That is, if Figure 11 As shown, for physical works with a maximum comprehensive similarity less than the first threshold and greater than or equal to the second threshold, a review is required (1130).

[0247] (3) S_max < 0.75: The judgment is passed. The work and its complete multidimensional feature set are added to the graph as new nodes and immediately enter the M6 ​​five-dimensional AI review. This step corresponds to the first matching degree being less than the second threshold mentioned above. That is, if... Figure 11 As shown, for physical works with a maximum comprehensive similarity less than the second threshold, the judgment is passed (1120).

[0248] The thresholds of 0.90 and 0.75 are not fixed values. The system will perform offline statistics on the historical review results for every 1,000 newly added works and adaptively fine-tune them within ±0.02 based on the ROC curve to avoid drift in the false positive rate and false negative rate.

[0249] The technical solution provided in this application executes different processes for the first physical work that passes the initial review, undergoes a second review, or fails the initial review. For works that pass the initial review, AI scoring is performed; for works that fail, similar physical works are identified to support the expression of the review result; and for works undergoing a second review, a manual review process is conducted. Based on this, this application adopts a multi-branch review method, improving the review effect. Furthermore, this multi-branch processing flow can also reduce the resource consumption of the review process to a certain extent (for example, works that fail do not need to enter the manual review process, nor do they need to enter the AI ​​scoring process), thereby reducing review costs and improving review efficiency.

[0250] In addition, for physical works that pass the manual review, the processing unit scores the first physical work using a work scoring model to determine its score information. This achieves a perfect closed loop in the review process, ensuring that only the final-passing physical works consistently enter the AI ​​scoring process, reflecting the completeness of the review process. Of course, directly using AI for the final scoring does not incur excessive processing overhead (model-based scoring is less difficult and less expensive).

[0251] The following explanation will first address the work rating mentioned in step 660 above.

[0252] In some embodiments, the work rating information includes scores in at least two of the following dimensions: creativity, theme expression, visual expression, craftsmanship, and entertainment value.

[0253] Optionally, the dimensions here are manually set, and the content of the specific dimensions is not limited to these five types; other dimensions may also be used. This application does not limit the specific dimensions.

[0254] For example, the scoring of the first physical work is determined by using a work scoring model, including: inputting at least two images of the first physical work and the three-dimensional geometric information of the first physical work into the work scoring model, and obtaining scores in at least two dimensions output by the work scoring model.

[0255] For example, the input to the work scoring model is at least two images of the first physical work and the three-dimensional geometric information of the first physical work, and the output of the work scoring model is a score in at least two dimensions.

[0256] For example, after obtaining scores for at least two dimensions, scores for at least two dimensions (or at least one dimension) in a graphical form as mentioned above can be obtained.

[0257] Optionally, the scores of at least two dimensions are weighted and summed to determine the overall score of the first physical work, wherein the overall score of the first physical work is used to determine the ranking of the first physical work in the judging process.

[0258] For example, the result is obtained by weighted summation of scores from at least two dimensions. This serves as the overall score for the first physical work. For example, the higher the overall score, the higher the ranking of the first physical work in the judging process.

[0259] Optionally, the overall technical architecture of this application includes module M6, used to implement downstream five-dimensional AI review and aesthetic alignment training. Optionally, module M6 corresponds to the processing unit in the aforementioned physical review device. That is, the processing unit in the aforementioned physical review device includes modules 2 to 6.

[0260] like Figure 12 As shown, module M6 (not shown in the figure) is the AI ​​scoring chain that follows the independence assessment in this application, and it consists of two parts: a "scoring model" and an "aesthetic alignment dataset". Figure 12 As shown, AI scoring of works is achieved through work scoring model 1210. Optionally, a five-dimensional scoring system can be adopted.

[0261] For each work that passes the independence judgment, a score of 0 to 10 is assigned to each of the following five dimensions, and these scores are weighted and combined to form a comprehensive score: 1. Creativity (weight 0.30): Novelty of the idea and degree of differentiation from historical samples; 2. Theme Expression (weight 0.20): Semantic relevance to the theme description of this competition; 3. Visual Expression (Weight 0.20): Color, Composition, Overall Visual Impact; 4. Craftsmanship (weight 0.20): Surface flatness, joint accuracy, structural stability, and other process indicators that can be inferred from geometry and high-resolution images; 5. Fun factor (weight 0.10): Does it have features such as "gamified easter eggs / interactive elements / secondary creation memes" that make players smile?

[0262] The scores of each dimension are weighted and summed to obtain the comprehensive score as follows: Score = 0.30 Creativity + 0.20 Theme + 0.20 Visual + 0.20 Craftsmanship + 0.10 Fun.

[0263] The technical solution provided in this application introduces a work scoring model to achieve AI scoring for the first physical work that passes the initial assessment. It also assists in determining the ranking of the physical work in the review process by calculating a comprehensive score. Therefore, this method not only improves the efficiency of determining review results but also improves the efficiency of ranking determination. Furthermore, scoring is implemented through a model, which has lower scoring difficulty, lower model processing overhead, and faster processing speed.

[0264] In some embodiments, the work rating model includes a multimodal encoder, a text encoder, and a regression decoder. For example, the work rating model is also called a dual-tower rating model, meaning it implements rating from both multimodal and textual perspectives.

[0265] For example, at least two images of the first physical artwork and its three-dimensional geometric information are encoded using a multimodal encoder to obtain multimodal encoded features. Optionally, the input to the multimodal encoder is at least two images of the first physical artwork and its three-dimensional geometric information, and the output of the multimodal encoder is the multimodal encoded features. A multimodal encoder is an encoder used to implement multimodal encoding.

[0266] For example, a text encoder encodes the descriptive text of the first physical work to obtain text-encoded features. The descriptive text of the first physical work describes the judging activities in which the first physical work participated. Optionally, the input to the text encoder is the descriptive text of the first physical work, and the output of the text encoder is the text-encoded features. A text encoder is an encoder used to implement text encoding.

[0267] For example, a regression decoder decodes multimodal encoded features and text encoded features to obtain scores in at least two dimensions. Optionally, the input to the regression decoder is the multimodal encoded features and the text encoded features, and the output of the text encoder is the score in at least two dimensions. The regression decoder is a decoder used to implement multi-head decoding.

[0268] Optionally, the work scoring model is called the CLIP-style dual-tower scoring model. Optionally, such as... Figure 12 The work scoring model 1210 shown adopts a dual-tower structure: 1. Real-world multimodal tower: The image branch processes six-view RGB images using ViT-B / 16 (each view is encoded separately and then attention pooling is performed to output 768 dimensions), while the point cloud branch processes the point cloud using PointNet++ to output 256 dimensions. The two are then concatenated and projected onto a 512-dimensional shared embedding space via MLP. This corresponds to the multimodal encoder mentioned above.

[0269] 2. Topic Text Tower: The tournament theme description (e.g., "Golden Shovel Spring Tournament Theme: Starry Night Tales") is encoded into a 512-dimensional topic vector using the CLIP text encoder. This corresponds to the text encoder mentioned above.

[0270] 3. Shared embedding: Cosine matching of physical object embedding and theme embedding serves as the basic signal for the theme dimension.

[0271] 4. Five-dimensional regression head: The physical embedding vector is connected to five independent MLP regression heads (each head has 2 layers of 256 units + 10 Sigmoid), which output scores in five dimensions. This corresponds to the regression decoder mentioned above.

[0272] The technical solution provided in this application embodiment achieves score prediction by combining multimodal (image and three-dimensional geometric information) and text (descriptive text) through a work scoring model (i.e., a dual-tower scoring model). The multi-dimensional score prediction achieved by the split-encoding of multi-angle input information is beneficial to improving the accuracy of score prediction.

[0273] In some embodiments, the training process of the work rating model includes a first-stage training process and a second-stage training process. Optionally, the first-stage training process is performed first, and the second-stage training process is performed later. Optionally, the first-stage training is pre-training, and the second-stage training is fine-tuning training.

[0274] Optionally, during the first stage of training, the work scoring model is trained using the first type of training samples, which are static and unchanging training samples.

[0275] For example, static training samples are obtained from historical review activities. Alternatively, the review results of physical works from historical review activities are obtained to determine the first type of training samples.

[0276] Optionally, during the second stage of training, the work scoring model is trained using a second type of training sample, which is a dynamically updated training sample.

[0277] For example, the dynamically updated training samples come from recent judging activities (judging activities within a preset time period). Optionally, the judging results of physical works from recent judging activities are obtained to determine the second type of training samples.

[0278] In some embodiments, the training sample set of the work rating model includes: a first training sample set, a second training sample set, and a third training sample set.

[0279] The training samples in the first training sample set include physical works that were manually reviewed and determined to be of the first type in historical judging activities. Optionally, the first type of physical works are also the outstanding award-winning works, that is, the award-winning works that were rated as outstanding by the judges. Optionally, the first training sample set is also called the L1 gold standard layer dataset. Outstanding award-winning works from previous judging activities (such as competitions) (for example, about 5,000 works, independently scored by senior judges in five dimensions) are used as strongly supervised samples to obtain the first training sample set.

[0280] The training samples in the second training sample set include physical works from previous judging events. Optionally, the second training sample set is also referred to as the L2 weakly labeled layer dataset. Optionally, the second training sample set is obtained by using the final ranking / relative order of previous shortlisted works (e.g., 30,000 works) as paired partial order learning samples.

[0281] The training samples in the third training sample set include: physical works with work determination information indicating a pass in the review activities within a preset time period, and physical works with additional scores manually assigned. Optionally, the third training sample set is also referred to as the dataset of the L3 self-evolution layer. Optionally, the third training sample set is obtained by weekly expanding with the newly stored passed works of each review activity (such as a competition) and the physical works with additional scores given by on-site judges.

[0282] Wherein, the first type of training samples includes the training samples in the first training sample set and the training samples in the second training sample set, and the second type of training samples includes the training samples in the third training sample set. That is, the first training sample set and the second training sample set are used to perform the first-stage training on the work scoring model, and the third training sample set is used to perform the second-stage training on the work scoring model.

[0283] Optionally, as shown in Figure 12 the dataset (i.e., the training sample set) of the work scoring model 1210 is divided into three layers: 1. L1 gold label layer 1220, corresponding to the aforementioned first training sample set: about 5000 excellent award-winning works from previous Golden Shovel events and similar competitions, with independent scores given by senior judges in five dimensions, serving as strongly supervised samples; 2. L2 weak label layer 1230, corresponding to the aforementioned second training sample set: about 30000 shortlisted works from previous events, with only final rankings / relative orders, serving as pairwise partial order learning samples; 3. L3 self-evolution layer 1240, corresponding to the aforementioned third training sample set: the newly stored passed works of each competition plus additional scores from on-site judges, expanded on a weekly basis.

[0284] The training loss is as follows: L = λ1·MSE(score_pred, score_gt)+λ2·PairwiseRank(score_pred_i, score_pred_j | rank_i<rank_j)+ λ3·CLIP-Contrastive(image_emb,theme_text_emb). Wherein, MSE(score_pred, score_gt) is a five-dimensional strong supervision loss, PairwiseRank(score_pred_i, score_pred_j | rank_i<rank_j) is a partial order learning loss, and CLIP-Contrastive(image_emb, theme_text_emb) is a theme alignment loss.

[0285] Wherein, λ1:λ2:λ3=1.0:0.5:0.3.

[0286] Training process: Offline pre-training on L1+L2 → LoRA fine-tuning using L3 incremental data weekly → After fine-tuning, the model is validated offline by ROC / τ-Kendall before being deployed to the offline cabin.

[0287] The technical solution provided in this application improves the training effect of the work scoring model through two-stage training and two types of training samples, thereby enhancing the model's ability to score physical works. Furthermore, using three different sets of training samples from various sources enriches the model's training sample set, further improving the model's training effect.

[0288] The following steps address the updating of the atlas and the identification of ancestral and sibling-like physical works. Explanation.

[0289] In some embodiments, another embodiment of this application provides a method for evaluating physical works, which includes at least one of the following steps (701-704, not shown in the figures). This method is applied to a physical evaluation device, which includes a first camera unit, a second camera unit, a processing unit, and a display unit.

[0290] Step 701: The first camera unit acquires at least two images of the first physical work, the different images being taken from different shooting angles of the first physical work.

[0291] Step 702: The second camera unit acquires the three-dimensional geometric information of the first physical artwork. The three-dimensional geometric information of the first physical artwork is used to indicate the geometric shape of the first physical artwork in three-dimensional space.

[0292] Step 703: The processing unit determines the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work; and determines the review result of the first physical work based on the matching degree between the first physical work and the second physical work.

[0293] In some embodiments, the review results include information on the judgment of the work.

[0294] Step 704: If the work judgment information is that the judgment is passed, the processing unit updates the corresponding judgment map in the historical judgment database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work, and obtains the updated judgment map.

[0295] In this framework, nodes in the review graph indicate reviewed physical works, edge weights between nodes indicate the matching degree between different physical works, and the updated review graph is used to review physical works to be reviewed.

[0296] In some embodiments, if the work is deemed to have passed the evaluation, the first physical work needs to be added to the historical evaluation database. Optionally, the historical evaluation database can be updated by adding the first physical work to the evaluation map.

[0297] In some embodiments, the historical review database is stored in the form of a graph, resulting in a review graph. Optionally, each node in the review graph indicates a reviewed physical work, and the edge weights between nodes are considered to represent the degree of matching between two physical works.

[0298] In some embodiments, each node stores node information. Optionally, the node information is information related to the reviewed physical artwork corresponding to that node. Optionally, the node information includes: at least two images of the physical artwork, three-dimensional geometric information of the physical artwork, and artwork rating information.

[0299] In some embodiments, the processing unit creates nodes in the review graph based on at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork. The processing unit also creates edges in the review graph based on the matching degree between the first and second physical artworks, resulting in an updated review graph.

[0300] In some embodiments, the judging chart includes a second node that indicates a second physical artwork. For example, the second node in the judging chart is the node corresponding to the second physical artwork.

[0301] Optionally, the processing unit creates a first node in the review atlas based on at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork. For example, at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork are stored as node information of the first node.

[0302] Optionally, if the matching degree between the first physical work and the second physical work is greater than or equal to a third threshold, the processing unit adds an edge between the first node and the second node to obtain an updated review graph. Conversely, if the matching degree between the first physical work and the second physical work is less than the third threshold, the processing unit does not add an edge between the first node and the second node. For example, the third threshold is a preset value. For instance, the third threshold is equal to 0.5. That is, when creating an edge for the first node, the edge is created selectively, only between two nodes with relatively high matching degrees. For example, if the matching degree between the first physical work and the second physical work is greater than or equal to 0.5, the processing unit adds an edge between the first node and the second node. Optionally, if the matching degree between the first physical work and the second physical work is less than 0.5, the processing unit does not add an edge between the first node and the second node.

[0303] The edge weight between the first node and the second node is determined based on the matching degree between the first physical artwork and the second physical artwork. Optionally, the edge weight between the first node and the second node is equal to the matching degree between the first physical artwork and the second physical artwork.

[0304] In some embodiments, for each node in the review graph, a matching degree is calculated with the first physical work, and nodes with a matching degree greater than or equal to a third threshold are selected. For these nodes, edges are created between them and the first node.

[0305] Optionally, the overall technical architecture of this application includes module M3, used to construct and retrieve creative lineage maps. Optionally, module M3 corresponds to the processing unit in the aforementioned physical review device.

[0306] like Figure 10 As shown, module M3 (not shown in the figure) maintains a creative lineage graph G=(V,E) composed of all historical works in the database. This creative lineage graph is also the review graph mentioned above. This graph includes the following: (1) Node V 1010: Each work in the database corresponds to a node, which stores its six-view image thumbnail (corresponding to at least two images mentioned above), point cloud hash (corresponding to the three-dimensional geometric information mentioned above), collection time (corresponding to the timestamp mentioned above), submitter ID (corresponding to the user identifier mentioned above), five-dimensional rating vector (corresponding to the work rating information mentioned above), and other attributes.

[0307] (2) Edge E 1020: For newly added nodes, calculate the comprehensive similarity pairwise with all historical nodes:

[0308] Only keep Edges with a value ≥0.50 significantly control graph sparsity.

[0309] The technical solution provided in this application, for physical works that have passed the evaluation, achieves dynamic storage of physical works by updating the evaluation graph. Furthermore, by creating nodes and edges in the evaluation graph, the storage of the first physical work is achieved, and the graph update maintains the lineage relationship (similarity relationship) between works. Additionally, only nodes with a matching degree greater than or equal to a third threshold are retained for edge creation, which improves the efficiency and effectiveness of the graph update. Moreover, each update involves relatively minor changes to the original evaluation graph, making the update easier and reducing the processing overhead (the number of edges and nodes to be created is reduced).

[0310] In some embodiments, the first node in the review map is supplemented with at least one of the following information: at least two images of the first physical work; three-dimensional geometric information of the first physical work, which includes a three-dimensional point cloud of the first physical work; a timestamp of the first physical work to indicate the time when the image of the first physical work was acquired; a user identifier of the first physical work to indicate the user who submitted the first physical work to participate in the review; a session identifier of the first physical work to indicate the session of the first physical work in the review activity; and work scoring information of the first physical work.

[0311] Optionally, for any node in the review map, at least one of the following information is added: at least two images of the physical work corresponding to the node; three-dimensional geometric information of the physical work corresponding to the node; a timestamp of the physical work corresponding to the node, indicating the time when the image of the physical work was acquired; a user identifier of the physical work corresponding to the node, indicating the user who submitted the physical work to participate in the review; a session identifier of the physical work corresponding to the node, indicating the session of the review activity; and the work scoring information of the physical work corresponding to the node.

[0312] The timestamp of the physical artwork, also known as the timestamp corresponding to the node, is used to indicate the time when data (images and three-dimensional geometric information) was collected from the physical artwork.

[0313] The session number identifier for physical works can also be considered a session number tag for physical works. When reviewing similar physical works using the judging chart, one can also filter based on the session number tag, for example, filtering physical works from the same session.

[0314] Optionally, the overall technical architecture of this application includes module M5, used to realize sample retention and continuous evolution of the spectrum. Optionally, module M5 corresponds to the processing unit in the aforementioned physical evaluation device.

[0315] like Figure 11 As shown, module M5 is responsible for "how the approved works become the comparison base library and training samples". For works that are determined to pass, the system enters them into the database according to the following process.

[0316] (1) Feature persistence: Six-view JPEG + 3D PLY + FAISS / LSH index entries + five-dimensional rating are written to local storage, with an average of about 24MB occupied by a single work.

[0317] (2) Graph Update: New nodes are added to the graph and their adjacency lists are maintained. The FAISS and LSH indexes are updated incrementally to avoid a full database reconstruction. The adjacency list is the standard storage structure for a graph, maintaining a list of directly connected neighbors and edge weights for each node. When a new work is added to the database, a new node is added, and the edge between it and the recalled highly similar nodes is written into their respective adjacency lists. The FAISS IVF index supports directly appending new vectors to the corresponding inverted bucket without retraining the quantizer (the quantizer can be trained offline using the historical database). The MinHash signature of the new work is calculated and then appended to the corresponding LSH bucket, which is also an append operation. Therefore, each work added to the database only requires a local append operation instead of a full database reconstruction.

[0318] (3) Evolutionary feedback: For each round of manual review, the appearance / style feature encoder is fine-tuned weekly using Triplet Loss for the sample pairs that are "judged as plagiarism / judged as independent". At the same time, the samples that receive additional high scores from the manual judges in the newly added works are promoted to "gold standard aesthetic anchors" and the aesthetic alignment dataset of module M6 is written back. (4) Cross-event reuse: The graphs of multiple games can be layered by "event tag", which supports strict deduplication within the game and also supports tracing the creative lineage across game dimensions, adapting to the event structure of multiple themes in parallel.

[0319] The nodes in the review graph in this application's embodiments correspond to rich node information, which is beneficial for supporting subsequent review of physical works based on the review graph, as well as the screening and searching of similar physical works. Therefore, this application not only enriches the information indicated by the review graph, but also helps to improve the review effect of subsequent physical works and the efficiency of work search.

[0320] In some embodiments, the processing unit determines the ancestral similar physical works and sibling similar physical works of the first physical work based on the updated review map, and the review results also include the ancestral similar physical works and sibling similar physical works.

[0321] Among them, ancestral similar physical works are used to indicate the second physical works in the historical review database that meet the first similarity relationship with the first physical work and have the earliest review time. Sibling similar physical works are used to indicate the second physical works in the historical review database that meet the second similarity relationship with the first physical work and have the same review time period as the first physical work.

[0322] Optionally, the processing unit, based on the updated review graph, identifies the second physical work in the historical review database that satisfies a first similarity relationship with the first physical work and has the earliest review time as an ancestor-similar physical work. For example, the processing unit, based on the updated review graph, identifies the node in the review graph whose edge weights satisfy the first similarity relationship with the first node and has the earliest timestamp as an ancestor-similar node. The physical work corresponding to this ancestor-similar node is also the ancestor-similar physical work. Optionally, satisfying the first similarity relationship for edge weights includes edge weights greater than or equal to a threshold.

[0323] Optionally, the processing unit, based on the updated review graph, identifies second physical works in the historical review database that satisfy a second similarity relationship with the first physical work and are within the same review period as the first physical work as sibling similar physical works. For example, the processing unit, based on the updated review graph, identifies nodes in the review graph whose edge weights satisfy a second similarity relationship with the first node and whose timestamps are the same as or within the same preset time period as the first node as sibling similar nodes. The physical work corresponding to this sibling similar node is also a sibling similar physical work. Optionally, satisfying the second similarity relationship for edge weights includes edge weights greater than or equal to a threshold.

[0324] In one possible implementation, the processing unit determines, from the updated review graph, at least one node whose edge weight to the first node is greater than or equal to the fourth threshold; and identifies the physical work corresponding to the node with the earliest timestamp among the at least one nodes as an ancestor-similar physical work.

[0325] Optionally, among the nodes connected to the first node, at least one node with an edge weight greater than or equal to a fourth threshold is selected, and the physical work corresponding to the node with the earliest timestamp among the at least one node is identified as the ancestor-similar physical work. Optionally, the timestamp of each of the at least one node is earlier than or equal to the timestamp of the first node.

[0326] In another possible implementation, the processing unit determines from the updated review graph at least one node whose edge weight with the first node is greater than or equal to the fourth threshold; the processing unit determines the physical works corresponding to the at least one node whose timestamp is the same as the first node and whose edge weight with the first node is the largest as sibling similar physical works.

[0327] Optionally, among the nodes connected to the first node, at least one node with an edge weight greater than or equal to the fourth threshold is selected, and the physical works corresponding to the at least one node with the same timestamp as the first node and the largest edge weight between the node and the first node are identified as sibling similar physical works.

[0328] The term "same timestamp" should be interpreted broadly. Same timestamp includes timestamps falling within the same preset time period (e.g., 30 days). The matching degree mentioned in this application is also considered a comprehensive similarity.

[0329] Optionally, appearance features are dimensionality-reduced using PCA and stored in FAISS IndexIVF, point cloud voxel signatures are stored in MinHashLSH, and style Gram matrices are stored in the approximate nearest neighbor index. When querying new works, a three-way ANN (Approximate Nearest Neighbor) is first used to recall the Top-50 candidates, and then the comprehensive similarity is calculated on the candidate set, with an overall query time of <0.8 seconds. Optionally, the appearance path (FAISS IVF): 512-dimensional PCA features → training an IVF quantizer to divide the vector space into nlist clusters (Voronoi units), each feature is recorded in the inverted bucket of the nearest cluster when entering the database; during querying, only the inner product / cosine nearest neighbor search is performed in the nearest nprobe buckets to recall the Top-N. Optionally, the geometric path (MinHash LSH): Treat the set occupied by point cloud voxels as a set, and use k hash functions to take the minimum hash value of each set to form a MinHash signature; the signature is then divided into bands for LSH bucketing, and works with similar signatures (high Jaccard) fall into the same bucket and are recalled. The Jaccard estimated by MinHash is approximately equal to the voxel IoU. Optionally, the style path (ANN index): Build an approximate nearest neighbor index (such as HNSW) for the style vectors after flattening / dimensionality reduction of each layer of Gram matrix, and recall according to the style vector distance. The top-50 candidates are obtained by the union of the three recall candidates, and the comprehensive similarity is then calculated on these 50.

[0330] Optionally, based on the graph edge weights and timestamps, each node can be used to reverse-look up the "earliest highly similar ancestor node" and "similar sibling nodes from the same period," providing a visual basis for manual review and interpretation. Here, "ancestor" corresponds to the aforementioned ancestral similar physical works. "Sibling" corresponds to the aforementioned sibling similar physical works.

[0331] The technical solution provided in this application presents the review results, including similar physical works by siblings and similar physical works by ancestors, to the user. This can also serve as an auxiliary means to determine plagiarism or for subsequent judging, enriching the expression of physical work review and encouraging user participation in physical work review, thus promoting human-computer interaction. Therefore, this application enriches the means of reviewing physical works, and the proposed method for determining similar physical works by siblings and similar physical works by ancestors can shorten search time and improve the efficiency of determining similar physical works by siblings and similar physical works by ancestors.

[0332] The technical solution provided in this application embodiment calls module M6 to complete the five-dimensional scoring on the terminal side for each work that passes module M4; the scoring results and the multimodal features entered into the database by module M5 are stored together; at the same time, the five-dimensional radar chart and the comprehensive score are displayed on the screen to facilitate players' understanding; the scoring results are aggregated weekly and written back to the competition ranking service for award collection and marketing content generation.

[0333] The technical solution provided in this application automatically triggers a dual-track review of "AI scoring + human judges" for samples with low confidence in the score (the confidence in any dimension prediction (which is output synchronously by the work scoring model when outputting a certain dimension score) > the threshold), avoiding inaccuracies of the edge model in rare styles; at the same time, it uses "dimensional radar chart + Top-3 similar award samples" as aesthetic references on the screen to explain the scoring basis to the players.

[0334] The technical solution provided in this application has the following beneficial effects: (1) Truly understand three-dimensional works: Through the collaborative acquisition of six-view images and 3D point clouds, the system can identify three-dimensional plagiarism and micro-modification behaviors that cannot be distinguished by two-dimensional images, making the review results closer to the "work itself" rather than just "a certain angle of the work", significantly improving the authenticity and effectiveness of plagiarism detection; (2) From subjective to objective review: The five-dimensional AI scoring model is continuously trained based on outstanding works from previous years, so that the system's judgment on "what is a good work" gradually approaches the aesthetics of competition experts, alleviating the experience problem of "mechanical, rigid and lacking competition feel" of pure algorithm scoring; (3) The more you use it, the smarter the judging ability becomes: Each entry into the database will feed back into the independence map and aesthetic model. The deeper the competition goes, the more sensitive the system becomes to plagiarism methods and aesthetic trends. It can achieve self-evolution of capabilities without human intervention, and the long-tail operating cost decreases with each competition. (4) One solution can be reused across multiple business lines: The same hardware and algorithm system can be transferred to various offline physical review scenarios such as art appraisal, intangible cultural heritage handicraft competition, campus creative competition, and collection figurine appraisal, so that the R&D investment of a single business line can be transformed into a company-level reusable asset.

[0335] To address the issues of plagiarism / copying independence and subjective human scoring in the review of offline physical UGC (User Generated Content) creative works, this application proposes a closed-loop method covering the entire chain, from physical collection to lineage mapping for plagiarism detection and then to multi-dimensional AI scoring and database entry.

[0336] The core idea of ​​this application is to map each physical artwork as a node in a graph, with its identity credential composed of six-view images and 3D point clouds simultaneously acquired within the machine. Between nodes, a weighted edge is obtained through weighted fusion of independent similarities in geometry, appearance, and style. All nodes and edges are organized into a continuously growing creative lineage graph. After a new artwork enters the system, it first executes a complete acquisition process to form a multi-dimensional feature set, and then searches the graph for its closest ancestor nodes and similar sibling nodes. Based on a graded admission strategy, it gives three conclusions: pass, manual review, or direct rejection. For artworks that pass the independence assessment, they are further sent to the downstream AI review subsystem, where the CLIP-style dual-tower scoring model scores them along five dimensions: creativity, thematic expression, visual expression, craftsmanship, and fun, and then weights and fuses the scores to obtain the artwork's comprehensive competition score. Any artwork that passes, along with its complete multi-dimensional features and five-dimensional score, is written back into the graph and training sample pool, forming the comparison basis for subsequent artworks and the aesthetic anchor point for the model.

[0337] Unlike online image plagiarism detection solutions that can only process two-dimensional images and rely on network uploads, and unlike solutions that only provide a single aesthetic score, this application demonstrates that all calculations are completed on the in-cabin end, eliminating the need for network uploads, supporting physical 3D comparison, and supporting five-dimensional subdivision of competition scores. By aligning the aesthetic dataset with past outstanding works as anchor points, the AI ​​scoring output continuously aligns with the aesthetic preferences of the Golden Shovel and similar competitions. It is suitable for offline scenarios such as handicrafts, art products, and collectible figurines that require strict independent evaluation while also taking into account artistic assessment.

[0338] The following explanation details some possible combinations of steps involved in this application.

[0339] Combination method 1: Six-view image + point cloud information + database entry. This method includes at least one of the following steps P1~P4 (not shown in the figure). Optionally, the physical evaluation device includes: a first camera unit, a second camera unit, and a processing unit.

[0340] Step P1: The first camera unit acquires at least two images of the first physical work, and the different images are obtained from shooting angles of the first physical work.

[0341] In step P2, the second camera unit acquires the three-dimensional geometric information of the first physical artwork. The three-dimensional geometric information of the first physical artwork is used to indicate the geometric shape of the first physical artwork in three-dimensional space.

[0342] In step P3, the processing unit determines the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work; and determines the review result of the first physical work based on the matching degree between the first physical work and the second physical work.

[0343] Step P4: The processing unit updates the historical review database based on the matching degree.

[0344] The processing unit updates the historical review database based on the matching degree, including the following steps: The processing unit determines the first matching degree based on the matching degree between the first physical work and each of the N second physical works; wherein, the first matching degree is the maximum value of the matching degree between the first physical work and each of the N second physical works, and N is a positive integer; When the first matching degree is less than the second threshold, the processing unit updates the review graph corresponding to the historical review database based on the first physical work, thus obtaining the updated review graph. In the review graph, nodes indicate physical works that have already been reviewed, edge weights between nodes indicate the matching degree between different physical works, and the updated review graph is used to review the physical works to be reviewed.

[0345] The processing unit updates the corresponding judging graph in the historical judging database based on the first physical work, resulting in an updated judging graph, including: The processing unit creates a first node in the review map based on at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork; If the matching degree between the first physical work and the second physical work is greater than or equal to the third threshold, the processing unit adds an edge between the first node and the second node to obtain an updated review graph; wherein, the edge weight between the first node and the second node is determined based on the matching degree between the first physical work and the second physical work.

[0346] Optionally, the processing unit determines the review result of the first physical work based on the degree of matching between the first physical work and the second physical work.

[0347] Optionally, the physical judging device also includes a display unit. The display unit displays the judging results of the first physical work.

[0348] Combination method 2: Six-view image + point cloud information + matching degree + AI scoring. This method includes at least one of the following steps Q1~Q4 (not shown in the figure). Optionally, the physical evaluation device includes: a first camera unit, a second camera unit, and a processing unit.

[0349] Step Q1: The first camera unit acquires at least two images of the first physical work, and the different images are obtained from shooting angles of the first physical work.

[0350] In step Q2, the second camera unit acquires the three-dimensional geometric information of the first physical artwork. The three-dimensional geometric information of the first physical artwork is used to indicate the geometric shape of the first physical artwork in three-dimensional space.

[0351] In step Q3, the processing unit determines the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work.

[0352] In step Q4, the processing unit scores the first physical artwork based on the matching degree.

[0353] In some embodiments, the processing unit scores the first physical artwork based on the matching degree, including the following steps: The processing unit determines the review result of the first physical work based on the degree of matching between the first physical work and the second physical work; The processing unit scores the first physical work based on the review results.

[0354] The review results include information on the work's evaluation, which can be categorized as "passed," "subject to review," or "failed." Based on the review results, the processing unit scores the first physical work, including the following steps: If the work is deemed to have passed the judgment, the processing unit scores the first physical work using the work scoring model to determine the work scoring information. If the work's evaluation information indicates a review, and the manual review result is a pass, the processing unit will score the first physical work using the work scoring model to determine the work's scoring information.

[0355] Optionally, if the evaluation information indicates that the work fails, it is not necessary to score the first physical work.

[0356] Optionally, the physical judging device also includes a display unit. The display unit displays the judging results of the first physical work.

[0357] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0358] Please refer to Figure 13The diagram illustrates a block diagram of a physical evaluation apparatus according to an embodiment of this application. This apparatus is used to implement the aforementioned method for evaluating physical works. The apparatus can be the computer device described above, or it can be integrated into a computer device. Figure 13 As shown, the device 1300 may include: a first camera unit 1310, a second camera unit 1320, a processing unit 1330, and a display unit 1340; The first camera unit 1310 is used to acquire at least two images of the first physical work, wherein the different images are obtained from shooting angles of the first physical work; The second camera unit 1320 is used to acquire the three-dimensional geometric information of the first physical work, and the three-dimensional geometric information of the first physical work is used to indicate the geometric shape of the first physical work in three-dimensional space. The processing unit 1330 is configured to determine the matching degree between the first physical work and a second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work; and to determine the review result of the first physical work based on the matching degree between the first physical work and the second physical work. The display unit 1340 is used to display the review results of the first physical work.

[0359] In some embodiments, the processing unit 1330 is configured to determine the geometric similarity, appearance similarity, and style similarity between the first physical work and the second physical work based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work. The processing unit 1330 is used to determine the matching degree between the first physical work and the second physical work based on the geometric similarity, the appearance similarity and the style similarity; Wherein, the geometric similarity is used to indicate the similarity between the first physical work and the second physical work in the geometric dimension, the appearance similarity is used to indicate the similarity between the first physical work and the second physical work in the appearance dimension, and the style similarity is used to indicate the similarity between the first physical work and the second physical work in the style dimension.

[0360] In some embodiments, the processing unit 1330 is configured to perform voxelization processing on the first physical work based on the three-dimensional geometric information of the first physical work to obtain the voxelization processing result of the first physical work; and determine the geometric similarity between the first physical work and the second physical work based on the voxelization processing result of the first physical work and the voxelization processing result of the second physical work. The processing unit 1330 is configured to use an appearance encoder to perform feature encoding on at least two images of the first physical work to obtain appearance encoding features of the first physical work; and to determine the appearance similarity between the first physical work and the second physical work based on the appearance encoding features of the first physical work and the appearance encoding features of the second physical work; wherein the appearance encoder is a neural network model for encoding appearance features of images. The processing unit 1330 is configured to use a style encoder to perform feature encoding on at least two images of the first physical work to obtain style encoding features of the first physical work; and to determine the style similarity between the first physical work and the second physical work based on the style encoding features of the first physical work and the style encoding features of the second physical work; wherein the style encoder is a neural network model for encoding style features of images.

[0361] In some embodiments, the training samples for the appearance encoder and the style encoder include: In the same historical review activity, the results of manual review consist of a sample pair of physical works that passed the review and those that failed the review.

[0362] In some embodiments, the processing unit 1330 is configured to normalize the geometric similarity, the appearance similarity, and the style similarity respectively to obtain the normalization results corresponding to the geometric similarity, the appearance similarity, and the style similarity. The processing unit 1330 is used to perform a weighted summation of the normalized processing results corresponding to the geometric similarity, the appearance similarity, and the style similarity, to obtain the matching degree between the first physical work and the second physical work.

[0363] In some embodiments, the review result includes work judgment information, which is one of judgment passed, judgment for review, or judgment failed; If the work is deemed to have passed the judgment, the review result also includes work scoring information. The work scoring information is determined by scoring the first physical work using a work scoring model, which is a neural network model used to score physical works. When the work judgment information indicates a review, the review result also includes at least one similar physical work. The at least one similar physical work is a second physical work that is similar to the first physical work and is determined from the historical review database. The first physical work and the at least one similar physical work are used to review the first physical work manually.

[0364] In some embodiments, the display unit 1340 is configured to display the work judgment information of the first physical work, and, if the work judgment information indicates a pass, to display the work rating information of the first physical work in a graphical form; wherein, displaying the work rating information in a graphical form includes: displaying the scores of multiple dimensions included in the work rating information in a radar chart format; or, The display unit 1340 is configured to display the work determination information of the first physical work, and, if the work determination information indicates a review, to display at least one similar physical work of the first physical work; or... The display unit 1340 is used to display the work determination information of the first physical work, and the work determination information is that the determination fails.

[0365] In some embodiments, if the work is deemed to have passed the review, the review result further includes: Ancestor-similar physical works are used to indicate the second physical work in the historical review database that satisfies the first similarity relationship with the first physical work and has the earliest review time; The term "sibling similar physical works" is used to indicate a second physical work in the historical review database that satisfies a second similarity relationship with the first physical work and is located in the same review period as the first physical work.

[0366] In some embodiments, the review result includes work judgment information, which is one of judgment passed, judgment for review, or judgment failed; The processing unit 1330 is configured to determine a first matching degree based on the matching degree between the first physical work and each of the N second physical works; wherein, the first matching degree is the maximum value of the matching degree between the first physical work and each of the N second physical works, and N is a positive integer; If the first matching degree is greater than or equal to the first threshold, the processing unit 1330 is used to determine that the work determination information of the first physical work is a failure to pass the determination. If the first matching degree is less than the first threshold and greater than or equal to the second threshold, the processing unit 1330 is used to determine that the work judgment information of the first physical work is a judgment review; and to determine at least one similar physical work based on the matching degree between the first physical work and N second physical works respectively. If the first matching degree is less than the second threshold, the processing unit 1330 is used to determine that the work judgment information of the first physical work is judged as passed; and to score the first physical work through a work scoring model to determine the work scoring information of the first physical work, wherein the work scoring model is a neural network model used to score physical works.

[0367] In some embodiments, the work scoring information includes scores in at least two of the following dimensions: creativity, theme expression, visual expression, craftsmanship, and entertainment value. The processing unit 1330 is used to input at least two images of the first physical work and the three-dimensional geometric information of the first physical work into the work scoring model to obtain the scores of the at least two dimensions output by the work scoring model. The processing unit 1330 is used to perform a weighted summation of the scores of the at least two dimensions to determine the comprehensive score of the first physical work, wherein the comprehensive score of the first physical work is used to determine the ranking of the first physical work in the review activity.

[0368] In some embodiments, the work scoring model includes: a multimodal encoder, a text encoder, and a regression decoder; The processing unit 1330 is used to encode at least two images of the first physical work and the three-dimensional geometric information of the first physical work through the multimodal encoder to obtain multimodal coding features. The processing unit 1330 is used to encode the descriptive text of the first physical work through the text encoder to obtain text encoding features. The descriptive text of the first physical work is used to describe the review activities in which the first physical work participated. The processing unit 1330 is used to decode the multimodal coding features and the text coding features through the regression decoder to obtain scores for the at least two dimensions.

[0369] In some embodiments, the training process of the work scoring model includes: a first-stage training process and a second-stage training process; During the first stage of training, the work scoring model is trained using a first type of training sample, which is a static and unchanging training sample. During the second stage of training, the work scoring model is trained using a second type of training sample, which is a dynamically updated training sample.

[0370] In some embodiments, the training sample set of the work rating model includes: The first training sample set includes physical works that have been identified as Type I physical works by manual review during historical review activities. The second training sample set includes physical works that were shortlisted in the aforementioned historical review activities. The third training sample set includes training samples that, during the review activities within a preset time period, consist of physical works that have passed the review and physical works that have received additional manual scoring. The first type of training samples includes training samples from the first training sample set and training samples from the second training sample set, and the second type of training samples includes training samples from the third training sample set.

[0371] In some embodiments, if the manual review result of the first physical work, in conjunction with the at least one similar physical work, is deemed to have passed, the processing unit 1330 is used to score the first physical work using the work scoring model to determine the work scoring information of the first work.

[0372] In some embodiments, the processing unit 1330 is configured to determine a first index value of the first physical work based on the voxelization processing result of the first physical work determined by the three-dimensional geometric information of the first physical work and the voxelization processing result of each second physical work in the historical review database. The first index value of the first physical work is used to indicate a first physical work set, which includes at least one second physical work that is similar to the first physical work in geometric dimensions. The processing unit 1330 is configured to determine a second index value of the first physical work based on the appearance coding features of the first physical work determined from at least two images of the first physical work and the appearance coding features of each second physical work in the historical review database. The second index value of the first physical work is used to indicate a second physical work set, which includes at least one second physical work that is similar to the first physical work in appearance dimension. The processing unit 1330 is configured to determine a third index value of the first physical work based on the style coding features of the first physical work determined from at least two images of the first physical work and the style coding features of each second physical work in the historical review database. The third index of the first physical work is used to indicate a third physical work set, which includes at least one second physical work that is similar to the first physical work in terms of style. The processing unit 1330 is configured to perform the step of determining the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work for the second physical work in the similar works set. The set of similar works includes the first set of physical works, the second set of physical works, and the third set of physical works. The set of similar works includes N second physical works, where N is a positive integer.

[0373] In some embodiments, the review results include work judgment information; If the work is deemed to have passed the judgment, the processing unit 1330 is used to update the review map corresponding to the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work, so as to obtain the updated review map. The nodes in the review graph are used to indicate the physical works that have been reviewed, the edge weights between the nodes in the review graph are used to indicate the matching degree between different physical works, and the updated review graph is used to review the physical works to be reviewed.

[0374] In some embodiments, the evaluation chart includes a second node, which is used to indicate the second physical work; In some embodiments, the processing unit 1330 is configured to create a first node in the review atlas based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work; If the matching degree between the first physical work and the second physical work is greater than or equal to the third threshold, the processing unit 1330 is used to add an edge between the first node and the second node to obtain the updated review graph. The edge weight between the first node and the second node is determined based on the matching degree between the first physical work and the second physical work.

[0375] In some embodiments, the first node in the review graph is added with at least one of the following information: At least two images of the first physical work; The three-dimensional geometric information of the first physical work includes the three-dimensional point cloud of the first physical work; The timestamp of the first physical work is used to indicate the time when the image of the first physical work was captured; The user identifier of the first physical work is used to indicate the user who submitted the first physical work to participate in the review; The session identifier of the first physical work is used to indicate the session in which the first physical work is judged. The work rating information for the first physical work.

[0376] In some embodiments, the processing unit 1330 is configured to determine, based on the updated review map, ancestral similar physical works and sibling similar physical works of the first physical work, and the review result further includes the ancestral similar physical works and the sibling similar physical works; The ancestral similar physical works are used to indicate the second physical works in the historical review database that satisfy the first similarity relationship with the first physical works and have the earliest review time. The sibling similar physical works are used to indicate the second physical works in the historical review database that satisfy the second similarity relationship with the first physical works and have the same review time period as the first physical works.

[0377] In some embodiments, the processing unit 1330 is configured to determine from the updated review graph at least one node whose edge weight with the first node is greater than or equal to a fourth threshold; and to determine the physical work corresponding to the node with the earliest timestamp among the at least one nodes as the ancestor-similar physical work. The processing unit 1330 is used to identify the physical works corresponding to the nodes in the at least one node whose timestamps are the same as the first node and whose edge weights with the first node are the largest, as the sibling similar physical works.

[0378] First, this application acquires at least two images of the first physical work to be reviewed using a first camera unit, and acquires the three-dimensional geometric information of the first physical work using a second camera unit. The processing unit determines the matching degree between the first physical work and the reviewed second physical work based on the images and geometric information, thereby determining and displaying the review result of the first physical work. Compared to related technologies that rely solely on photos uploaded by participating users to determine the independence of a work, this application acquires images and three-dimensional geometric information of the physical work itself. Because the acquired data is multi-dimensional (images taken from at least two shooting angles and three-dimensional geometric information indicating the geometric shape of the physical work), the rich data source improves the effectiveness of the review of the physical work, rather than limiting the review to a single image, as in related technologies.

[0379] Furthermore, this application integrates the processes of image acquisition, information collection, evaluation, and result display of physical works into a single physical evaluation device, eliminating the need for users to upload their own work. This not only improves the evaluation effect but also increases the efficiency of physical work evaluation. Moreover, this integrated evaluation method also ensures the security and privacy of the evaluation process for physical works.

[0380] Please refer to Figure 14 This diagram illustrates a block diagram of a physical artwork evaluation apparatus according to an embodiment of this application. The apparatus has the function of implementing the steps performed by the processing unit in the aforementioned physical artwork evaluation method. The apparatus can be the processing unit described above, or it can be disposed within the processing unit. Figure 14 As shown, the device 1400 may include: a data acquisition module 1410, a matching degree determination module 1420, and a result determination module 1430.

[0381] The data acquisition module 1410 is used to acquire at least two images of the first physical work and the three-dimensional geometric information of the first physical work. The different images are obtained from shooting angles of the first physical work. The three-dimensional geometric information of the first physical work is used to indicate the geometric shape of the first physical work in three-dimensional space. The matching degree determination module 1420 is used to determine the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work. The result determination module 1430 is used to determine the review result of the first physical work based on the matching degree between the first physical work and the second physical work.

[0382] In some embodiments, the matching degree determination module 1420 is used for: Based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work, determine the geometric similarity, appearance similarity, and style similarity between the first physical work and the second physical work; The matching degree between the first physical artwork and the second physical artwork is determined based on the geometric similarity, the appearance similarity, and the style similarity. Wherein, the geometric similarity is used to indicate the similarity between the first physical work and the second physical work in the geometric dimension, the appearance similarity is used to indicate the similarity between the first physical work and the second physical work in the appearance dimension, and the style similarity is used to indicate the similarity between the first physical work and the second physical work in the style dimension.

[0383] In some embodiments, the matching degree determination module 1420 is used for: Based on the three-dimensional geometric information of the first physical work, the first physical work is voxelized to obtain the voxelization result of the first physical work; based on the voxelization result of the first physical work and the voxelization result of the second physical work, the geometric similarity between the first physical work and the second physical work is determined. An appearance encoder is used to encode features of at least two images of the first physical work to obtain appearance encoding features of the first physical work; based on the appearance encoding features of the first physical work and the appearance encoding features of the second physical work, the appearance similarity between the first physical work and the second physical work is determined; wherein, the appearance encoder is a neural network model for encoding appearance features of images. A style encoder is used to encode features of at least two images of the first physical work to obtain style encoding features of the first physical work; based on the style encoding features of the first physical work and the style encoding features of the second physical work, the style similarity between the first physical work and the second physical work is determined; wherein, the style encoder is a neural network model for encoding style features of images.

[0384] In some embodiments, the training samples for the appearance encoder and the style encoder include: In the same historical review activity, the results of manual review consist of a sample pair of physical works that passed the review and those that failed the review.

[0385] In some embodiments, the matching degree determination module 1420 is used for: The geometric similarity, appearance similarity, and style similarity are normalized respectively to obtain the normalized results corresponding to each of the geometric similarity, appearance similarity, and style similarity. The matching degree between the first physical work and the second physical work is obtained by weighted summing of the normalized results corresponding to the geometric similarity, appearance similarity and style similarity.

[0386] In some embodiments, the review result includes work judgment information, which is one of judgment passed, judgment for review, or judgment failed; If the work is deemed to have passed the judgment, the review result also includes work scoring information. The work scoring information is determined by scoring the first physical work using a work scoring model, which is a neural network model used to score physical works. When the work judgment information indicates a review, the review result also includes at least one similar physical work. The at least one similar physical work is a second physical work that is similar to the first physical work and is determined from the historical review database. The first physical work and the at least one similar physical work are used to review the first physical work manually.

[0387] In some embodiments, if the work is deemed to have passed the review, the review result further includes: Ancestor-similar physical works are used to indicate the second physical work in the historical review database that satisfies the first similarity relationship with the first physical work and has the earliest review time; The term "sibling similar physical works" is used to indicate a second physical work in the historical review database that satisfies a second similarity relationship with the first physical work and is located in the same review period as the first physical work.

[0388] In some embodiments, the review result includes work judgment information, which is one of judgment passed, judgment for review, or judgment failed; In some embodiments, the result determination module 1430 is configured to: A first matching degree is determined based on the matching degree between the first physical work and each of the N second physical works; wherein, the first matching degree is the maximum value of the matching degree between the first physical work and each of the N second physical works, and N is a positive integer; If the first matching degree is greater than or equal to the first threshold, the work determination information of the first physical work is determined to be a failure to pass the determination. If the first matching degree is less than the first threshold and greater than or equal to the second threshold, the work determination information of the first physical work is determined to be subject to review; based on the matching degree between the first physical work and N second physical works, at least one similar physical work is determined. If the first matching degree is less than the second threshold, the work judgment information of the first physical work is determined to be passed; the first physical work is scored by a work scoring model to determine the work score information of the first physical work, wherein the work scoring model is a neural network model used to score physical works.

[0389] In some embodiments, the work scoring information includes scores in at least two of the following dimensions: creativity, theme expression, visual expression, craftsmanship, and entertainment value. In some embodiments, the result determination module 1430 is configured to: Input at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork into the artwork scoring model to obtain the scores of the at least two dimensions output by the artwork scoring model; The scores of the at least two dimensions are weighted and summed to determine the overall score of the first physical work, wherein the overall score of the first physical work is used to determine the ranking of the first physical work in the review activity.

[0390] In some embodiments, the work scoring model includes: a multimodal encoder, a text encoder, and a regression decoder; In some embodiments, the result determination module 1430 is configured to: The multimodal encoder encodes at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork to obtain multimodal coding features. The text encoder encodes the descriptive text of the first physical work to obtain text encoding features. The descriptive text of the first physical work is used to describe the judging activities in which the first physical work participated. The regression decoder decodes the multimodal encoded features and the text encoded features to obtain scores for at least two dimensions.

[0391] In some embodiments, the training process of the work scoring model includes: a first-stage training process and a second-stage training process; During the first stage of training, the work scoring model is trained using a first type of training sample, which is a static and unchanging training sample. During the second stage of training, the work scoring model is trained using a second type of training sample, which is a dynamically updated training sample.

[0392] In some embodiments, the training sample set of the work rating model includes: The first training sample set includes physical works that have been identified as Type I physical works by manual review during historical review activities. The second training sample set includes physical works that were shortlisted in the aforementioned historical review activities. The third training sample set includes training samples that, during the review activities within a preset time period, consist of physical works that have passed the review and physical works that have received additional manual scoring. The first type of training samples includes training samples from the first training sample set and training samples from the second training sample set, and the second type of training samples includes training samples from the third training sample set.

[0393] In some embodiments, the result determination module 1430 is configured to: If the first physical work is deemed to have passed the manual review in conjunction with at least one similar physical work, the first physical work is scored using the work scoring model to determine the work score information of the first work.

[0394] In some embodiments, the matching degree determination module 1420 is used for: Based on the voxelization result of the first physical work determined by the three-dimensional geometric information of the first physical work, and the voxelization result of each second physical work in the historical review database, a first index value of the first physical work is determined. The first index value of the first physical work is used to indicate a first physical work set. The first physical work set includes at least one second physical work that is similar to the first physical work in geometric dimensions. Based on the appearance coding features of the first physical work determined from at least two images of the first physical work, and the appearance coding features of each second physical work in the historical review database, a second index value of the first physical work is determined. The second index value of the first physical work is used to indicate a second physical work set, which includes at least one second physical work that is similar to the first physical work in appearance dimension. Based on the style coding features of the first physical work determined from at least two images of the first physical work, and the style coding features of each second physical work in the historical review database, a third index value of the first physical work is determined. The third index of the first physical work is used to indicate a third physical work set, which includes at least one second physical work that is similar to the first physical work in terms of style. For the second physical work in the set of similar works, the step of determining the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work is executed respectively. The set of similar works includes the first set of physical works, the second set of physical works, and the third set of physical works. The set of similar works includes N second physical works, where N is a positive integer.

[0395] In some embodiments, the review results include information on the determination of the work.

[0396] In some embodiments, the device 1400 further includes a map update module (not shown in the figure).

[0397] In some embodiments, the map updating module is used for: If the work is deemed to have passed the judgment, the historical review database is updated based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work to obtain the updated review database. The nodes in the review graph are used to indicate the physical works that have been reviewed, the edge weights between the nodes in the review graph are used to indicate the matching degree between different physical works, and the updated review graph is used to review the physical works to be reviewed.

[0398] In some embodiments, the evaluation chart includes a second node, which is used to indicate the second physical work; In some embodiments, the map updating module is used for: Based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work, a first node is created in the review atlas; If the matching degree between the first physical work and the second physical work is greater than or equal to the third threshold, an edge is added between the first node and the second node to obtain the updated review graph. The edge weight between the first node and the second node is determined based on the matching degree between the first physical work and the second physical work.

[0399] In some embodiments, the first node in the review graph is added with at least one of the following information: At least two images of the first physical work; The three-dimensional geometric information of the first physical work includes the three-dimensional point cloud of the first physical work; The timestamp of the first physical work is used to indicate the time when the image of the first physical work was captured; The user identifier of the first physical work is used to indicate the user who submitted the first physical work to participate in the review; The session identifier of the first physical work is used to indicate the session in which the first physical work is judged. The work rating information for the first physical work.

[0400] In some embodiments, the map updating module is used for: Based on the updated review map, ancestral similar physical works and sibling similar physical works of the first physical work are determined, and the review results also include the ancestral similar physical works and the sibling similar physical works; The ancestral similar physical works are used to indicate the second physical works in the historical review database that satisfy the first similarity relationship with the first physical works and have the earliest review time. The sibling similar physical works are used to indicate the second physical works in the historical review database that satisfy the second similarity relationship with the first physical works and have the same review time period as the first physical works.

[0401] In some embodiments, the result determination module 1430 is configured to: From the updated review graph, at least one node with an edge weight greater than or equal to the fourth threshold is identified; the physical work corresponding to the node with the earliest timestamp among the at least one node is identified as the ancestor-similar physical work. The physical works corresponding to the nodes whose timestamps are the same as the first node and whose edge weights with the first node are the largest among the at least one nodes are identified as the sibling similar physical works.

[0402] This application acquires images and three-dimensional geometric information of physical works. Since the acquired data of physical works is multi-dimensional (images taken from at least two shooting angles and three-dimensional geometric information used to indicate the geometric shape of physical works), the rich data source improves the effectiveness of the evaluation of the physical works, rather than limiting the evaluation to a single image, such as in related technologies.

[0403] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules when implementing its functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0404] Please refer to Figure 15 The diagram illustrates a structural block diagram of a computer device 1500 provided in one embodiment of this application. The computer device 1500 can be any electronic device capable of data calculation, processing, and storage. The computer device 1500 can be used to implement the physical work evaluation method provided in the above embodiments.

[0405] Typically, computer device 1500 includes a processor 1501 and a memory 1502.

[0406] Processor 1501 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 1501 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA, and PLA (Programmable Logic Array). Processor 1501 may also include a main processor and a coprocessor. The main processor, also known as the CPU, is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1501 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content required to be displayed on the screen. In some embodiments, processor 1501 may also include an AI processor, which is used to handle computational operations related to machine learning.

[0407] Memory 1502 may include one or more computer-readable storage media, which may be tangible and non-transitory. Memory 1502 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in memory 1502 stores a computer program that is loaded and executed by processor 1501 to implement the physical work evaluation methods provided in the above-described method embodiments.

[0408] Those skilled in the art will understand that Figure 15 The structure shown does not constitute a limitation on the computer device 1500, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0409] In an exemplary embodiment, a computer-readable storage medium is also provided, wherein a computer program is stored therein, which, when executed by a processor, implements the above-described method for evaluating physical works.

[0410] Optionally, the computer-readable storage medium may include: ROM (Read-Only Memory), RAM (Random Access Memory), SSD (Solid State Drives), or optical disc, etc. The random access memory may include ReRAM (Resistance Random Access Memory) and DRAM (Dynamic Random Access Memory).

[0411] In an exemplary embodiment, a computer program product is also provided, the computer program product including a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, causing the computer device to perform the above-described method for evaluating physical works.

[0412] It should be noted that the collection and processing of relevant data (including physical works, historical review databases, etc.) in this application should strictly comply with the requirements of relevant national laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0413] It should be understood that "multiple" as used herein refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. Furthermore, the step numbers described herein are merely illustrative of one possible execution order. In some other embodiments, the steps may not be executed in numerical order, such as two steps with different numbers being executed simultaneously, or two steps with different numbers being executed in the reverse order of the illustration. This application does not limit this.

[0414] The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for evaluating physical works, characterized in that, The method is applied to a physical object evaluation device, the physical object evaluation device comprising: a first camera unit, a second camera unit, a processing unit, and a display unit, and includes: The first camera unit acquires at least two images of the first physical artwork, wherein the different images are obtained from shooting angles of the first physical artwork; The second camera unit acquires the three-dimensional geometric information of the first physical work, and the three-dimensional geometric information of the first physical work is used to indicate the geometric shape of the first physical work in three-dimensional space; The processing unit determines the matching degree between the first physical work and a second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work; and determines the review result of the first physical work based on the matching degree between the first physical work and the second physical work. The display unit displays the review results of the first physical artwork.

2. The method according to claim 1, characterized in that, The processing unit determines the matching degree between the first physical artwork and a second physical artwork in the historical review database based on at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork, including: The processing unit determines the geometric similarity, appearance similarity, and style similarity between the first physical work and the second physical work based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work. The processing unit determines the matching degree between the first physical artwork and the second physical artwork based on the geometric similarity, the appearance similarity, and the style similarity; Wherein, the geometric similarity is used to indicate the similarity between the first physical work and the second physical work in the geometric dimension, the appearance similarity is used to indicate the similarity between the first physical work and the second physical work in the appearance dimension, and the style similarity is used to indicate the similarity between the first physical work and the second physical work in the style dimension.

3. The method according to claim 2, characterized in that, The processing unit determines the geometric similarity, appearance similarity, and style similarity between the first physical artwork and the second physical artwork based on at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork, including: The processing unit performs voxelization processing on the first physical work based on the three-dimensional geometric information of the first physical work to obtain the voxelization processing result of the first physical work; and determines the geometric similarity between the first physical work and the second physical work based on the voxelization processing result of the first physical work and the voxelization processing result of the second physical work. The processing unit employs an appearance encoder to perform feature encoding on at least two images of the first physical work to obtain the appearance encoding features of the first physical work; based on the appearance encoding features of the first physical work and the appearance encoding features of the second physical work, it determines the appearance similarity between the first physical work and the second physical work; wherein, the appearance encoder is a neural network model used to encode appearance features of images. The processing unit employs a style encoder to perform feature encoding on at least two images of the first physical work to obtain style encoding features of the first physical work; based on the style encoding features of the first physical work and the style encoding features of the second physical work, it determines the style similarity between the first physical work and the second physical work; wherein, the style encoder is a neural network model used for style feature encoding of images.

4. The method according to claim 3, characterized in that, The training samples for the appearance encoder and the style encoder include: In the same historical review activity, the results of manual review consist of a sample pair of physical works that passed the review and those that failed the review.

5. The method according to any one of claims 2 to 4, characterized in that, The processing unit determines the matching degree between the first physical artwork and the second physical artwork based on the geometric similarity, the appearance similarity, and the style similarity, including: The processing unit normalizes the geometric similarity, appearance similarity, and style similarity respectively to obtain the normalization results corresponding to each of the geometric similarity, appearance similarity, and style similarity. The processing unit performs a weighted summation of the normalized processing results corresponding to the geometric similarity, appearance similarity, and style similarity to obtain the matching degree between the first physical work and the second physical work.

6. The method according to any one of claims 1 to 5, characterized in that, The review results include work judgment information, which is one of the following: passed, subject to review, or failed. If the work is deemed to have passed the judgment, the review result also includes work scoring information. The work scoring information is determined by scoring the first physical work using a work scoring model, which is a neural network model used to score physical works. When the work judgment information indicates a review, the review result also includes at least one similar physical work. The at least one similar physical work is a second physical work that is similar to the first physical work and is determined from the historical review database. The first physical work and the at least one similar physical work are used to review the first physical work manually.

7. The method according to claim 6, characterized in that, The display unit displays the review results of the first physical artwork, including: The display unit displays the artwork judgment information of the first physical artwork, and if the artwork judgment information indicates a pass, it displays the artwork rating information of the first physical artwork in a graphical form; wherein, displaying the artwork rating information in a graphical form includes: displaying the scores of multiple dimensions included in the artwork rating information in the form of a radar chart; or, The display unit displays the work determination information of the first physical work, and if the work determination information indicates a review, it displays at least one similar physical work of the first physical work; or... The display unit displays the work determination information of the first physical work, and the work determination information indicates that the work has failed the determination.

8. The method according to claim 6 or 7, characterized in that, If the work is deemed to have passed the evaluation, the review result also includes: Ancestor-similar physical works are used to indicate the second physical work in the historical review database that satisfies the first similarity relationship with the first physical work and has the earliest review time; The term "sibling similar physical works" is used to indicate a second physical work in the historical review database that satisfies a second similarity relationship with the first physical work and is located in the same review period as the first physical work.

9. The method according to any one of claims 1 to 8, characterized in that, The review results include work judgment information, which is one of the following: passed, subject to review, or failed. The processing unit determines the review result of the first physical work based on the matching degree between the first physical work and the second physical work, including: The processing unit determines a first matching degree based on the matching degree between the first physical artwork and each of the N second physical artworks; wherein, the first matching degree is the maximum value of the matching degree between the first physical artwork and each of the N second physical artworks, and N is a positive integer; If the first matching degree is greater than or equal to the first threshold, the processing unit determines that the work judgment information of the first physical work is a judgment failure; If the first matching degree is less than the first threshold and greater than or equal to the second threshold, the processing unit determines the work judgment information of the first physical work as a judgment review; and determines at least one similar physical work based on the matching degree between the first physical work and N second physical works respectively. If the first matching degree is less than the second threshold, the processing unit determines that the work judgment information of the first physical work is judged as passed; the first physical work is scored by a work scoring model to determine the work score information of the first physical work, wherein the work scoring model is a neural network model used to score physical works.

10. The method according to claim 9, characterized in that, The scoring information for the works includes scores in at least two of the following dimensions: creativity, theme expression, visual expression, craftsmanship, and entertainment value. The step of scoring the first physical work using a work scoring model to determine the work scoring information of the first work includes: Input at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork into the artwork scoring model to obtain the scores of the at least two dimensions output by the artwork scoring model; The method further includes: The scores of the at least two dimensions are weighted and summed to determine the overall score of the first physical work, wherein the overall score of the first physical work is used to determine the ranking of the first physical work in the review activity.

11. The method according to claim 10, characterized in that, The work scoring model includes: a multimodal encoder, a text encoder, and a regression decoder; The step of inputting at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork into the artwork scoring model to obtain the scores of the at least two dimensions output by the artwork scoring model includes: The multimodal encoder encodes at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork to obtain multimodal coding features. The text encoder encodes the descriptive text of the first physical work to obtain text encoding features. The descriptive text of the first physical work is used to describe the judging activities in which the first physical work participated. The regression decoder decodes the multimodal encoded features and the text encoded features to obtain scores for at least two dimensions.

12. The method according to any one of claims 9 to 11, characterized in that, The training process of the work scoring model includes: a first-stage training process and a second-stage training process; During the first stage of training, the work scoring model is trained using a first type of training sample, which is a static and unchanging training sample. During the second stage of training, the work scoring model is trained using a second type of training sample, which is a dynamically updated training sample.

13. The method according to claim 12, characterized in that, The training sample set for the work rating model includes: The first training sample set includes physical works that have been identified as Type I physical works by manual review during historical review activities. The second training sample set includes physical works that were shortlisted in the aforementioned historical review activities. The third training sample set includes training samples that, during the review activities within a preset time period, consist of physical works that have passed the review and physical works that have received additional manual scoring. The first type of training samples includes training samples from the first training sample set and training samples from the second training sample set, and the second type of training samples includes training samples from the third training sample set.

14. The method according to any one of claims 9 to 13, characterized in that, The method further includes: If the first physical work is deemed to have passed the manual review in conjunction with at least one similar physical work, the processing unit scores the first physical work using the work scoring model to determine the work scoring information of the first work.

15. The method according to any one of claims 1 to 14, characterized in that, The method further includes: The processing unit determines the first index value of the first physical work based on the voxelization processing result of the first physical work determined by the three-dimensional geometric information of the first physical work, and the voxelization processing result of each second physical work in the historical review database. The first index value of the first physical work is used to indicate the first physical work set, which includes at least one second physical work that is similar to the first physical work in geometric dimensions. The processing unit determines a second index value for the first physical work based on the appearance coding features of the first physical work determined from at least two images of the first physical work, and the appearance coding features of each second physical work in the historical review database. The second index value of the first physical work is used to indicate a second physical work set, which includes at least one second physical work that is similar to the first physical work in appearance dimension. The processing unit determines a third index value for the first physical work based on the style coding features of the first physical work determined from at least two images of the first physical work, and the style coding features of each second physical work in the historical review database. The third index of the first physical work is used to indicate a third physical work set, which includes at least one second physical work that is similar to the first physical work in terms of style. The processing unit performs the step of determining the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work for the second physical work in the similar works set. The set of similar works includes the first set of physical works, the second set of physical works, and the third set of physical works. The set of similar works includes N second physical works, where N is a positive integer.

16. The method according to any one of claims 1 to 15, characterized in that, The review results include information on the judgment of the works, and the method further includes: If the work is deemed to have passed the judgment, the processing unit updates the review map corresponding to the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work, and obtains the updated review map. The nodes in the review graph are used to indicate the physical works that have been reviewed, the edge weights between the nodes in the review graph are used to indicate the matching degree between different physical works, and the updated review graph is used to review the physical works to be reviewed.

17. The method according to claim 16, characterized in that, The evaluation chart includes a second node, which is used to indicate the second physical work; The processing unit updates the review map corresponding to the historical review database based on the first physical work, obtaining the updated review map, including: The processing unit creates a first node in the review map based on at least two images of the first physical artwork and the three-dimensional geometric information of the first physical artwork; If the matching degree between the first physical work and the second physical work is greater than or equal to the third threshold, the processing unit adds an edge between the first node and the second node to obtain the updated review graph. The edge weight between the first node and the second node is determined based on the matching degree between the first physical work and the second physical work.

18. The method according to claim 16 or 17, characterized in that, The first node in the review graph has at least one of the following pieces of information added: At least two images of the first physical work; The three-dimensional geometric information of the first physical work includes the three-dimensional point cloud of the first physical work; The timestamp of the first physical work is used to indicate the time when the image of the first physical work was captured; The user identifier of the first physical work is used to indicate the user who submitted the first physical work to participate in the review; The session identifier of the first physical work is used to indicate the session in which the first physical work is judged. The work rating information for the first physical work.

19. The method according to any one of claims 16 to 18, characterized in that, The method further includes: The processing unit determines the ancestral similar physical works and the sibling similar physical works of the first physical work based on the updated review map. The review result also includes the ancestral similar physical works and the sibling similar physical works. The ancestral similar physical works are used to indicate the second physical works in the historical review database that satisfy the first similarity relationship with the first physical works and have the earliest review time. The sibling similar physical works are used to indicate the second physical works in the historical review database that satisfy the second similarity relationship with the first physical works and have the same review time period as the first physical works.

20. The method according to claim 19, characterized in that, The processing unit determines, based on the updated review map, ancestral similar physical works and sibling similar physical works of the first physical work, including: The processing unit determines, from the updated review graph, at least one node whose edge weight with the first node is greater than or equal to a fourth threshold; and determines the physical work corresponding to the node with the earliest timestamp among the at least one node as the ancestor-similar physical work. The processing unit identifies the physical works corresponding to the nodes in the at least one node whose timestamps are the same as the first node and whose edge weights with the first node are the largest, as the sibling similar physical works.

21. A physical evaluation device, characterized in that, The physical evaluation device includes: a first camera unit, a second camera unit, a processing unit, and a display unit; The first camera unit is used to capture at least two images of the first physical work, wherein the different images are obtained from shooting angles of the first physical work; The second camera unit is used to acquire the three-dimensional geometric information of the first physical work, and the three-dimensional geometric information of the first physical work is used to indicate the geometric shape of the first physical work in three-dimensional space. The processing unit is configured to determine the matching degree between the first physical work and a second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work; and to determine the review result of the first physical work based on the matching degree between the first physical work and the second physical work. The display unit is used to display the review results of the first physical work.

22. A judging device for physical works, characterized in that, The device includes: The data acquisition module is used to acquire at least two images of the first physical work and the three-dimensional geometric information of the first physical work. The different images are obtained from shooting angles of the first physical work. The three-dimensional geometric information of the first physical work is used to indicate the geometric shape of the first physical work in three-dimensional space. The matching degree determination module is used to determine the matching degree between the first physical work and the second physical work in the historical review database based on at least two images of the first physical work and the three-dimensional geometric information of the first physical work. The result determination module is used to determine the review result of the first physical work based on the matching degree between the first physical work and the second physical work.

23. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the steps performed by the processing unit in the method for evaluating physical works as described in any one of claims 1 to 20.

24. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the steps performed by the processing unit in the method for evaluating physical works as described in any one of claims 1 to 20.

25. A computer program product, characterized in that, The computer program product includes a computer program that is loaded and executed by a processor to implement the steps performed by the processing unit in the method for evaluating physical works as described in any one of claims 1 to 20.

Images