Testing and evaluating method for live-action modeling of nerve radiation field

Abstract

The invention discloses a test and evaluation method for live-action modeling of a nerve radiation field, and the method comprises the following steps: S10, carrying out data set preparation: collecting a public data set, carrying out the live-action shooting of a forward data set and a reverse data set with a designed place, carrying out the targeted enhancement of a data set, and generating a data set through employing AIGC; the data set is used for training a modeling model or a test model of the system; s20, based on the situation that a user may use different functions of applications in parallel in an actual user scene, policy task issuing is carried out in an interface calling mode; s30, carrying out system equipment performance monitoring and modeling duration recording, analyzing state data and a trend thereof, analyzing whether a task is abnormal or not and whether performance consumption and task duration are reasonable or not, and carrying out targeted optimization based on a target; and S40, after task operation is completed and operation data is recorded, the modeling success rate and the modeling effect are evaluated.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional reconstruction, and in particular relates to a test and evaluation method for real-scene modeling of a neural radiation field. Background Art

[0002] 3D reconstruction is a computer technology that converts entities in the physical world into digital models. It is a rapidly developing field that is changing the way we interact with the world. Neural Radiance Fields (NeRF) is a 3D scene reconstruction method based on neural networks. It can restore realistic 3D scenes from 2D image data and achieve new perspective synthesis at the photo level. It has changed the traditional 3D reconstruction method. The generated model is more realistic and detailed, and can capture complex effects such as lighting and shadows in real scenes. NeRF technology has been widely used in various practical applications, such as virtual reality, autonomous driving, game development, smart tourism, program production, etc., so that these applications provide a more immersive and interactive experience.

[0003] NeRF itself currently has some shortcomings. For example, NeRF-based modeling requires a large number of images or videos from different perspectives as input; sufficient computing power and time are required to train for each scene; it mainly considers static scenes; and the generalization ability is somewhat lacking for trained scenes. Currently, the system application forms based on NeRF modeling provided by providers in different application fields are various, such as embedded applications, terminal applications, server applications, etc. From the user's perspective, for different modeling systems, the basic framework process of use is somewhat similar, such as Figure 1 As shown, it generally includes: users shoot modeling materials, users upload the shot materials, model the materials, and then users browse the modeling effects, and users download or share the modeling file materials.

[0004] From the requirements for the use of this type of neural radiation scene modeling, there are several key points that users may be concerned about: (1) What kind of materials can the system support: Does it only support professionally shot materials, or can it be compatible with materials shot anywhere? Does it only support a certain type of material, or can it support various types of materials? We call this the domain applicability of the modeling algorithm. (2) After the material is uploaded, how long does it take to model? Can the modeling be successful? If the modeling is successful, is the modeling effect realistic? This involves requirements such as the performance duration, modeling success rate, and modeling accuracy of the modeling algorithm.

[0005] The following is a further explanation of the pain points from the user perspective, which are also indicators for evaluating the pros and cons of the system: (1) Domain applicability: refers to the degree of support for scenarios in a specific application field when the system is used. Different application fields have different scenarios. For example, in the field of autonomous driving, the modeling scenarios come from all aspects of driving and traffic; in the field of smart cultural tourism, the modeling scenarios come from various natural, cultural and social landscapes. Since NeRF has certain deficiencies in scene generalization capabilities, it is necessary to collect and use corresponding field materials for targeted training for different fields. The greater the coverage of the data set, the greater the space for system optimization, and the higher the adaptability to the field. (2) Modeling performance and time: When the NeRF method generates an image, each pixel requires nearly 200 forward predictions of the MLP (multi-layer perceptron) deep model. Although the single calculation scale is not large, the amount of calculation required to complete the rendering of the entire image pixel by pixel is still considerable. Therefore, for different operating devices, the modeling performance and time consumption will vary greatly. It is necessary to record and analyze the performance indicators and time consumption of the operating devices for different scene materials in order to evaluate the modeling capability level of different devices and modeling algorithms. (3) Modeling success rate: For a single material, determine whether the modeling is successful or not; for a batch of materials, calculate the number and ratio of successful modeling. The modeling success rate is affected by comprehensive factors, such as material characteristics, model generalization ability, device performance, concurrent tasks, etc. It is necessary to record the success or failure of modeling of different materials in various modeling scenarios. The higher the success rate, the better the algorithm model. (4) Modeling accuracy: refers to the degree of realism of the modeling effect. Based on the advantages of NeRF technology itself, through the captured scene pictures under known perspectives, the scene is modeled into a continuous 5D radiation field and implicitly stored in the neural network using a deep learning algorithm. A neural radiation field model is obtained through training. Based on this model, clear photos from any perspective can be rendered. Therefore, in order to ensure the realism of the modeling, it is necessary to ensure the quality and magnitude of the input data: multiple perspectives, sufficient clarity, sufficient quantity, wide scenes, etc.

[0006] In summary, in order to better solve the pain points of users using the neural radiation field real-scene modeling system and continuously optimize and evaluate the system, a set of systematic and professional processes and methods are needed. Summary of the invention

[0007] In view of the above problems, the present invention provides a test and evaluation method for real-scene modeling of neural radiation fields, which is used to solve the optimization problem of real-scene modeling evaluation of neural radiation fields and improve efficiency.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A test and evaluation method for real-scene modeling of a neural radiation field comprises the following steps:

[0010] S10, prepare data sets, including collecting public data sets, designed real-life shooting forward data sets and reverse data sets, targeted enhancement data sets, and data sets generated by AIGC; data sets are used to train the system's modeling model or test model;

[0011] S20, based on the fact that users may use different functions of the application in parallel in actual user scenarios, strategically dispatch tasks using an interface call method;

[0012] S30, monitor the performance of system equipment and record the modeling duration, analyze the status data and its trend, analyze whether there are any abnormalities in the task, whether the performance consumption and task duration are reasonable, and perform targeted optimization based on the goal;

[0013] S40, after the task is completed and the operation data is recorded, the modeling success rate and modeling effect are evaluated.

[0014] In a possible implementation, in S10, the designed real-scene shooting forward data set and reverse data set specifically include: purposeful design shooting of information including the photographed objects, environment, and trajectory according to training needs; the reverse data set includes a reflective data set that is easy to model holes and concave-convex effects; an inferior data set with moving targets, sudden changes in light, and perspective that is easy to fail in modeling; a data set with incomplete perspective, too fast movement speed, or too large intervals between adjacent images; the data set source is an image video shot by a pure camera, or an image video with a laser radar, structured light, and time-of-flight sensor to actively detect and obtain depth information.

[0015] In one possible implementation, in S10, the targeted enhanced data set is image processing based on real-scene data to simulate materials with different resolutions, clarity, lighting conditions, shooting speeds, and angles, wherein the image processing includes up and down changes in resolution, changes in clarity, changes in lighting, shooting speed control, and changes in angles.

[0016] In a possible implementation, in S20, the actual application scenario includes at least one of material uploading, material modeling, model rendering and display, or model downloading.

[0017] In one possible implementation, in S30, system equipment performance monitoring includes: task logs designed and output by the application system itself, task alarm information, single task running time, system logs, task-level and system-level CPU, GPU, memory, video memory, network bandwidth, and storage consumption.

[0018] In a possible implementation, in S30, the modeling effect evaluation includes manual evaluation, semi-automatic evaluation and automatic evaluation.

[0019] In one possible implementation, the semi-automatic evaluation includes comparing the generated model with the non-standard model, intercepting each image of the generated model on the shooting trajectory according to certain rules or intervals, and performing image-by-image quality comparison with images at corresponding positions between the real video or image.

[0020] In a possible implementation, the fully automated evaluation includes testing a data set. After applying an algorithm to generate a model, a point cloud of the generated model is compared with a point cloud acquired by a structured light scanner to obtain an evaluation of the model effect.

[0021] In one possible implementation, the evaluation index includes calculating the Qinfo distance, which is used to specify the original model and generate the model point cloud; for each reference point, calculating the distance to the nearest point in the target point cloud; for each target point, calculating the distance to the nearest point in the reference point cloud; accumulating or averaging the distance values ​​to form a comprehensive chamfer distance, a large distance indicates that the two groups of point clouds are more different, and a small distance indicates that the reconstruction effect is better.

[0022] In one possible implementation, the evaluation index includes a mean square error, which is used to specify the original model and the generated model point cloud; the distance between corresponding points is calculated; the error of each point is squared to obtain the square error of each point; the square errors of all points are added to obtain the total square error; the total square error is divided by the number of points to obtain the average mean square error MSE. The smaller the value of MSE, the smaller the difference between the generated 3D model and the true model, and the better the model performance.

[0023] The present invention has the following beneficial effects:

[0024] (1) By collecting public data sets, designing real-life forward and reverse data sets, using different methods and tools to enhance the data in terms of resolution, clarity, lighting conditions, shooting speed control, angle changes, etc., using AIGC technology to generate modeling materials, etc., we can obtain comprehensive and difficult-to-obtain materials, greatly reducing the difficulty and cost of manual entity shooting.

[0025] (2) After obtaining the data set, analyze the various usage possibilities in actual application scenarios, consider the normal testing requirements of algorithm optimization regression, use interface calls to strategically issue tasks, and collect and analyze application and system logs, alarms, task running time, performance indicators and other information in a timely manner during the test process to evaluate the algorithm and device performance, thereby improving the pertinence, effectiveness, comprehensiveness and efficiency of the test.

[0026] (3) By using different indicators and tools to compare the gap between the generated model and the real model, and by using full-parameter and no-parameter image quality evaluation methods to evaluate the quality difference between the model image and the captured video image, and by calculating the ratio of successful modeling data sets, the accuracy, compatibility and stability of the model modeling can be obtained, which is convenient for equipment selection and operation and maintenance expansion and reduction, thereby reducing production costs.

[0027] (4) The entire solution is easy to repeat and quickly execute, and it conducts a comprehensive and systematic evaluation of the algorithm model and device performance, greatly improving the efficiency of the test and its actual application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A schematic diagram of a general process of modeling the reality of neural radiation fields in the prior art;

[0029] Figure 2 A flowchart of a test and evaluation method for real-scene modeling of a neural radiation field according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] See also Figure 2 , which is a flow chart of a method for testing and evaluating a neural radiation field real scene modeling according to an embodiment of the present invention, comprising the following steps:

[0032] S10, prepare data sets, including collecting public data sets, designed real-life shooting forward data sets and reverse data sets, targeted enhancement data sets, and data sets generated by AIGC; data sets are used to train the system's modeling model or test model;

[0033] S20, based on the fact that users may use different functions of the application in parallel in actual user scenarios, strategically dispatch tasks using an interface call method;

[0034] S30, monitor the performance of system equipment and record the modeling duration, analyze the status data and its trend, analyze whether there are any abnormalities in the task, whether the performance consumption and task duration are reasonable, and perform targeted optimization based on the goal;

[0035] S40, after the task is completed and the operation data is recorded, the modeling success rate and modeling effect are evaluated.

[0036] Specifically, in S10, the data required by NeRF usually comes from images or videos taken from multiple angles of the same scene. When shooting, it is necessary to consider different camera shooting parameter settings (focal length, aperture, exposure time, ISO), different lighting conditions, different shooting routes and angles (front, side, top view and top view), different shooting speeds, avoid dynamic targets, etc.; according to the characteristics of the shooting object itself, whether there is clear texture, whether there is occlusion, reflection, perspective, etc.; according to the scene, it can be divided into indoor objects, indoor scenes, outdoor objects, outdoor scenes, etc. In summary, various factors when shooting different scenes will ultimately affect the effect and success rate of modeling.

[0037] Generally speaking, test data is best obtained from real shooting, but it is easy to imagine the difficulty of shooting. Scenes cannot be controlled, such as weather and lighting; scenes cannot be shot, such as dangerous environments; scenes are difficult to obtain, such as treasures; scenes are difficult to shoot, such as grand landscapes or buildings. Even if there is a good shooting environment and shooting equipment, the investment cost of manpower and time is very considerable. Therefore, data sets will be an obvious bottleneck for general application providers.

[0038] In order to optimize the dataset acquisition path at the lowest cost, this method designs the following methods to prepare the required datasets: (1) Public datasets: such as DTU, LLFF, ScanNet, Tanks and Temples, ETH3D, etc. Each type of dataset has its own scene characteristics, shooting techniques, different annotations / auxiliary content, etc., which can be selected and used according to different training needs. (2) Real scene shooting: The objects, environment, trajectory, etc. are shot according to the training needs. The captured dataset can be a standard dataset for training or a reverse dataset for verifying the algorithm compatibility or optimization effect. The reverse dataset includes but is not limited to: reflective datasets that are easy to model holes and concave and convex effects; low-quality datasets with moving targets, sudden changes in light, and perspective that are easy to fail to model; datasets with incomplete perspective, too fast movement speed, or too large intervals between adjacent images, etc. It can be an image video shot by a pure camera, or an image video with sensors such as lidar, structured light, and ToF (time of flight) to actively detect and obtain depth information. (3) Data enhancement: Perform certain image processing based on real-scene data to simulate materials with different resolutions, clarity, lighting conditions, shooting speeds, and angles. Including but not limited to: a. Resolution changes: The resolution of the material can be enlarged or reduced using transcoding software to change the size of the material. This can solve the problem of repeated shooting of the material when the algorithm adapts to different resolutions; b. Clarity changes: The clarity can be changed through transcoding methods such as bit rate change, sharpening, denoising, and brightness change; c. Lighting changes: Using the DaVinci dimming and color adjustment function, the light path of the picture, the light and dark contrast of the area, the hue and color system, etc. can be directly changed; d. Shooting speed control: Different shooting speeds determine the overlap between adjacent images, thereby affecting the modeling effect. When the video is fixed, the frame rate can be changed through different interpolation strategies of the transcoding software. The higher the frame rate, the higher the overlap between adjacent images; e. Angle changes: The material can be rotated, cropped, filled, etc. to change the perspective of the material, etc. Tools such as Monai are used. (4) AIGC: Currently, there are many models and applications for generating videos from text and images, such as Sora, Stable Video Diffusion, Gemini, Pika, Emu Video, etc. With the development of technology, the effect of generating videos is getting better and better, the time is longer, the logic is stronger, and the stability is higher. For the training materials required for 3D modeling, it can provide greater possibilities, convenience and imagination space. Based on the technical characteristics of 3D modeling, special prompt words can be designed, such as object features, lighting conditions, observation angles, and surround speeds, to generate text and images, and the prompt words can be optimized based on the generated effects for generation; when the text description is limited, a small number of pictures can be combined to generate text and images.The generated video can be processed using video editing software, such as adding texture noise to improve realism.

[0039] Specifically, in S20, after the data set is prepared, it cannot be issued directly, because in the specific usage scenarios of the application, different tasks will be involved in parallel, such as: material uploading, material modeling, model rendering and display, model downloading, etc. Material uploading is to upload the data set, which may be a single data set or multiple data sets, depending on the specific support of the application. Material modeling is to convert the uploaded material from 2D images to 3D models, which is also the core function of the application. Based on different applications or device performance differences, it may only support single material modeling, or multi-material parallel modeling. Model rendering and display is for the modeled 3D model, which can be observed and explored from any perspective relatively casually. Model downloading is to download the modeled 3D model for more post-processing.

[0040] Application software needs to be designed with reasonable task allocation before verification based on the actual user scenario in which users may use different functions of the application in parallel: how many levels of user operations may be possible or need to be supported at the same time, what is the proportion of various operations, how long each operation may last, etc. After designing the task classification, level, and duration, you can use tools such as Postman, Python, Jmeter, etc. to issue tasks of corresponding levels and execution cycles to each task through interface calls.

[0041] For example, a server-side application system needs to support parallel access from 1,000 users, 20 parallel material upload tasks, 10 modeling tasks, 50 concurrent rendering tasks, and 50 parallel model download tasks. The system needs to run 24 / 7. At the same time, for each task, set the corresponding number of concurrent tasks, call different task interfaces concurrently, and after the task is completed, continue to initiate the next interface call task.

[0042] In addition to designing task strategies based on application scenarios, the algorithm itself also needs to be tested after algorithm regression. At this time, you only need to call the corresponding interface at the same time. For example, at the same time, send material modeling and perform modeling rendering, etc. Wait until the task is completed before sending the next task.

[0043] Through strategic task distribution design, the pertinence, effectiveness and efficiency of testing can be greatly improved.

[0044] Specifically, in S30, whether it is algorithm testing or scenario testing, its main purpose is to evaluate the performance and long-term stability of algorithms and devices, so as to facilitate subsequent algorithm optimization or equipment selection, and to evaluate the business carrying capacity of actual application scenarios for flow control or capacity expansion. Therefore, during the test, it is necessary to record various status data of tasks and devices, including but not limited to: a. Task logs, task alarms and other information designed and output by the application system itself. b. Single task running time. c. System log. d. Task-level and system-level CPU, GPU, memory, video memory, network bandwidth, storage and other consumption.

[0045] By analyzing the above status data and its trends, we can analyze whether there are any abnormalities in the task, whether the performance consumption and task duration are reasonable, and perform targeted optimization based on the goals.

[0046] Specifically, in S40, after the task is completed and the running data is recorded, the modeling success rate can be easily calculated. The modeling success rate can be defined based on the actual situation. In a narrow sense, it can only refer to the ratio of the number of successfully run data sets to the number of issued data sets; in a broad sense, it can also include the situation that although the task status is successful, the actual rendering or downloading of the model is abnormal, or the rendering effect is obviously poor. The modeling success rate can reflect indicators including but not limited to the compatibility and capability of the algorithm with the data set, the stability of the application system, and other indicators.

[0047] Regarding the evaluation of the modeling effect, the evaluation method in the embodiment of the present invention is divided into three modes: manual, semi-automatic and automatic.

[0048] For the manual method, it is mainly to directly observe the generated model with the naked eye to judge the difference and distance between it and the actual scene seen by the naked eye. Due to the limitations of the scenes that the naked eye can actually see, such as a building, it is not easy for the naked eye to observe from any angle. Therefore, in addition to subjective feelings and experience, the judgment of the difference between the model and the real object can also be carried out by borrowing some ideas of video quality evaluation, such as full parameter and no parameter evaluation.

[0049] Semi-automation and automation are mainly about the generation method of the original model. Taking the public dataset DTU as an example, the DTU dataset is an MVS dataset shot by a 6-axis industrial robot equipped with a camera and a structured light scanner. The robot provides precise camera positioning, and the MATLAB calibration toolbox is used to carefully calibrate the camera pose and internal parameters. The structured light scanner provides a dense point cloud. For the test of this dataset, after applying the algorithm to generate the model, the point cloud of the generated model is compared with the point cloud obtained by the structured light scanner to obtain an evaluation of the model effect. No human intervention is required in the middle, which is called an automated method. For semi-automation, because the dataset does not have special depth information when the video is shot, it is only pure two-dimensional data. At this time, there is no standard model to compare the generated model. Then some reverse operations can be performed, that is, based on the model generated by its own algorithm or excellent third-party tools, use tools such as Blender, UnrealEngine, etc. to perform certain optimization and modification, try to repair an artificial standard model, and then use it as a standard model to compare with the generated model. We call this method a semi-automated method. Whether it is automated or semi-automated, it is assumed that there is a standard model in the dataset. That is, the semi-automatic evaluation includes comparing the generated model with the non-standard model, intercepting each image of the generated model on the shooting trajectory according to certain rules or intervals, and comparing the image quality with the corresponding position of the real video or image. The fully automated evaluation includes testing the data set. After applying the algorithm to generate the model, the point cloud of the generated model is compared with the point cloud obtained by the structured light scanner to obtain an evaluation of the model effect.

[0050] The following are some comparison indicators for reference. More indicators can be expanded based on similar ideas.

[0051] There are comparison indicators of the original model: (1) Qinfo distance: specify the original model and the generated model point cloud; for each reference point, calculate the distance to the nearest point in the target point cloud; for each target point, calculate the distance to the nearest point in the reference point cloud; accumulate or average the distance values ​​to form a comprehensive chamfer distance. A large distance means that the two sets of point clouds are quite different, and a small distance means that the reconstruction effect is better. (2) Mean square error: specify the original model and the generated model point cloud; calculate the distance between corresponding points, such as Euclidean distance; square the error of each point to obtain the square error of each point; add the square errors of all points to obtain the total square error; divide the total square error by the number of points (number of samples) to obtain the mean square error (MSE). The smaller the MSE value, the smaller the difference between the generated 3D model and the real model, and the better the model performance. (3) Completeness: by calculating the overlap between the point cloud or model in the reconstructed model and the real model, determine how much of the real model content is contained in the reconstructed model. Tools include CloudCompare, meshlab, etc.

[0052] There is no comparison index for the original model. In the absence of the original model, the main comparison idea is to intercept the images of the generated model on the shooting trajectory according to certain rules or intervals, and compare the image quality with the images at the corresponding positions between the real video or image. This is equivalent to the full parameter evaluation of video quality. The corresponding evaluation indicators include PSNR, SSIM, VMAF, etc. For modeling images on non-shooting trajectories, since there is no source video material directly corresponding to it, non-parametric evaluation indicators can be used, such as blur, block effect, contrast, noise, etc. Regardless of whether there are parameters or not, the better the value, the higher the quality of the modeling picture and the better the modeling effect.

[0053] The three-dimensional real-scene modeling based on NeRF technology has an increasingly wide range of application fields and various application forms. Regardless of the application, the field applicability, modeling accuracy, modeling performance duration and modeling success rate of the modeling algorithm will be tested. The test and evaluation method of the neural radiation field real-scene modeling set above collects public data sets, designs real-scene shooting of positive and negative data sets, uses different methods and tools to enhance the data from the aspects of resolution, clarity, lighting conditions, shooting speed control, angle changes, etc., uses popular AIGC technology to generate modeling materials, and so on to obtain comprehensive and difficult-to-obtain materials, which greatly reduces the difficulty and cost of artificial entity shooting. After obtaining the data set, analyze various usage possibilities in actual application scenarios, consider the normal test requirements of algorithm optimization regression, use interface calling methods to strategically issue tasks, and collect and analyze application and system logs, alarms, task running time, performance indicators and other information in a timely manner during the test process. Evaluate the performance of algorithms and equipment, and improve the pertinence, effectiveness, comprehensiveness and efficiency of the test. Finally, by using different indicators and tools to compare the gap between the generated model and the real model, the quality difference between the model image and the captured video image is evaluated by full-parameter and no-parameter image quality evaluation, and by calculating the ratio of successful modeling data sets, the accuracy, compatibility and stability of the model modeling are obtained, which is convenient for equipment selection and operation and maintenance expansion and reduction, etc., to reduce production costs. The whole solution is easy to repeat and quickly execute, and conducts a comprehensive and systematic evaluation of the algorithm model and equipment performance, greatly improving the efficiency of the test and the actual application value.

[0054] It should be understood that the exemplary embodiments described herein are illustrative rather than restrictive. Although one or more embodiments of the present invention have been described in conjunction with the accompanying drawings, it should be understood by those skilled in the art that various changes in form and detail may be made without departing from the spirit and scope of the present invention as defined by the appended claims.

Claims

1. A test and evaluation method for real-scene modeling of neural radiation fields, characterized in that: The following steps are involved: S10, prepare the dataset, including collecting public datasets, designed real-life shooting forward and reverse datasets, targeted enhancement datasets, and using AIGC to generate datasets; The data set is used to train the system's modeling model or test the model; S20, based on the fact that users may use different functions of the application in parallel in actual user scenarios, strategically dispatch tasks using an interface call method; S30, monitor the performance of system equipment and record the modeling duration, analyze the status data and its trend, analyze whether there are any abnormalities in the task, whether the performance consumption and task duration are reasonable, and perform targeted optimization based on the goal; S40, after the task is completed and the operation data is recorded, the modeling success rate and modeling effect are evaluated.

2. The test and evaluation method for real-scene modeling of neural radiation field according to claim 1, characterized in that: In S10, the designed real-scene shooting forward data set and reverse data set specifically include: purposeful design shooting of information including the photographed objects, environment, and trajectory according to training needs; the reverse data set includes a reflective data set that is easy to model holes and concave-convex effects; an inferior data set with moving targets, sudden changes in light, and perspective that is easy to fail in modeling; a data set with incomplete perspective, too fast movement speed, or too large intervals between adjacent images; the data set source is an image video shot by a pure camera, or an image video with a laser radar, structured light, and time-of-flight sensor to actively detect and obtain depth information.

3. The test and evaluation method for real-scene modeling of neural radiation field according to claim 1, characterized in that: In S10, the targeted enhanced data set is image processing based on real scene data to simulate materials with different resolutions, clarity, lighting conditions, shooting speeds, and angles, wherein the image processing includes up and down changes in resolution, changes in clarity, changes in lighting, shooting speed control, and changes in angles.

4. The test and evaluation method for real-scene modeling of neural radiation field according to claim 1, characterized in that: In S20, the actual application scenario includes at least one of material uploading, material modeling, model rendering and display, or model downloading.

5. The test and evaluation method for real-scene modeling of neural radiation field according to claim 1, characterized in that: In S30, the system equipment performance monitoring includes: task logs designed and output by the application system itself, task alarm information, single task running time, system logs, task-level and system-level CPU, GPU, memory, video memory, network bandwidth, and storage consumption.

6. The test and evaluation method for real-scene modeling of neural radiation field according to claim 1, characterized in that: In S30, the modeling effect evaluation includes manual evaluation, semi-automatic evaluation and automatic evaluation.

7. The test and evaluation method for real-scene modeling of neural radiation field according to claim 6, characterized in that: The semi-automatic evaluation includes comparing the generated model with the non-standard model, intercepting each image of the generated model on the shooting trajectory according to certain rules or intervals, and performing image-by-image quality comparison with the images at corresponding positions between the real video or image.

8. The test and evaluation method for real-scene modeling of neural radiation field according to claim 6, characterized in that: The fully automated evaluation includes testing of the data set. After applying the algorithm to generate the model, the point cloud of the generated model is compared with the point cloud obtained by the structured light scanner to obtain an evaluation of the model effect.

9. The test and evaluation method for real-scene modeling of neural radiation field according to claim 8, characterized in that: The evaluation indicators include calculating the Qinfo distance, which is used to specify the original model and generate the model point cloud; for each reference point, calculating the distance to the nearest point in the target point cloud; for each target point, calculating the distance to the nearest point in the reference point cloud; accumulating or averaging the distance values ​​to form a comprehensive chamfer distance. A large distance indicates that the two groups of point clouds are quite different, and a small distance indicates that the reconstruction effect is better.

10. The test and evaluation method for real-scene modeling of neural radiation field according to claim 8, characterized in that: The evaluation indicators include mean square error, which is used to specify the original model and the generated model point cloud; calculate the distance between corresponding points; square the error of each point to obtain the square error of each point; add the square errors of all points to obtain the total square error; divide the total square error by the number of points to obtain the average mean square error MSE. The smaller the value of MSE, the smaller the difference between the generated 3D model and the real model, and the better the model performance.

Images