Three-dimensional object feature acquisition method, device, computer equipment and storage medium
By extracting the vertices in the target grid information of three-dimensional objects and their adjacent and nearest neighbor vertices for convolution operations, combining the K nearest neighbor algorithm and multi-layer perceptron network, the problems of computing power limitation and resource utilization in the prior art are solved. , improves the accuracy and performance of 3D target classification.
Patent Information
- Application Number
- CN202010674249.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-07-14
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2040-07-14
AI Technical Summary
In the 3D target classification task, the prior art has lost detailed information due to the limitation of hardware computing capabilities. The point cloud-based method has reduced classification performance due to the limitation of sampling method, while the multi-view 2D projection method requires a large amount of calculations. Power resources.
By obtaining the target grid information of a three-dimensional object, extracting all adjacent vertices and nearest neighbor vertices of each vertex, performing convolution operations to obtain update features and importance scores, and deleting them according to the score, updating grid information, and using K Nearest neighbor algorithms and multi-layer perceptron networks calculate convolution kernel parameters to reduce computing power resource usage.
It improves the accuracy and classification performance of three-dimensional object feature acquisition, reduces the occupation of computing resources, and can classify 3D targets based on fine granular structure information, improving the performance of subsequent tasks.
Smart Images

Figure CN111898654B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, computer equipment and storage medium for acquiring features of a three-dimensional object. Background Art
[0002] In some 3D target classification tasks, the input data is the model file of the object, that is, the target is expressed in mesh format, including information such as vertices and triangular faces. For these model files, deep learning-based methods are generally used to extract features for subsequent classification tasks. Specifically, the methods for extracting features can be divided into three categories: voxel-based, point cloud-based, and multi-view 2D projection-based. The voxel-based method first divides the space where the target model is located into uniform grids, and then extracts features through 3D convolution operations. The point cloud-based method first obtains point cloud data by sampling on the model surface, and then uses methods such as Pointnet to extract features. In the method based on multi-view 2D projection, the projection images of the model on 2D planes at different angles are first obtained, and then the convolutional neural network is used to extract features from these 2D images respectively and fuse them to give the final features.
[0003] However, the voxel-based method is limited by the hardware computing power and can only use a coarse-grained voxel division method, so a lot of detailed information will be lost. The point cloud-based method will also lose many key points due to the sampling method, resulting in reduced subsequent classification performance. Although the method based on multi-view 2D projection can achieve the best performance at present, it requires a lot of computing resources because it needs to perform feature extraction and fusion operations on multiple projected 2D images for each target model. Summary of the invention
[0004] Based on this, it is necessary to provide a three-dimensional object feature acquisition method, device, computer equipment and storage medium that can reduce computing resource usage and improve accuracy in response to the above technical problems.
[0005] A method for acquiring features of a three-dimensional object, the method comprising:
[0006] Obtaining target grid information of the three-dimensional object to be classified;
[0007] Extracting all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information;
[0008] Taking each vertex as the convolution center, performing a convolution operation on all adjacent vertices and nearest neighbor vertices corresponding to the vertex as convolution objects, and obtaining an updated feature and importance score corresponding to each vertex;
[0009] Deleting the vertices according to the importance scores;
[0010] The target mesh information is updated by acquiring the remaining vertices after the deletion process and the update features corresponding to the remaining vertices.
[0011] In one embodiment, extracting all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information includes:
[0012] Extracting all adjacent vertices of each vertex in the target mesh information;
[0013] The nearest neighbor vertex corresponding to each vertex in the target mesh information is extracted according to the K nearest neighbor algorithm.
[0014] In one embodiment, extracting the nearest neighbor vertex corresponding to each vertex in the target mesh information according to the K nearest neighbor algorithm includes:
[0015] Get the K value set according to different convolution receptive fields;
[0016] According to the K value, the nearest neighbor vertex corresponding to each vertex in the target mesh information is extracted by a K nearest neighbor algorithm.
[0017] In one embodiment, the convolution kernel parameters corresponding to the convolution operation are calculated by:
[0018] Calculate the spatial coordinate difference between each convolution center and the corresponding convolved object;
[0019] The spatial coordinate difference is input into a multi-layer perceptron network to calculate the convolution kernel parameters.
[0020] In one embodiment, deleting the vertices according to the importance scores includes:
[0021] Get the preset downsampling operation parameters;
[0022] sorting the vertices according to the importance scores;
[0023] Extract vertices whose number corresponds to the downsampling operation parameter from the sorted vertices, and delete the unextracted vertices.
[0024] In one embodiment, the obtaining of the remaining vertices after the deletion process and the update features corresponding to the remaining vertices to update the target mesh information includes:
[0025] The remaining vertices after the deletion process and the update features corresponding to the remaining vertices are obtained, and the target mesh information is updated by using the triangle edge collapse method with the importance score as the collapse cost.
[0026] In one embodiment, after obtaining the remaining vertices after the deletion process and updating the target mesh information with the update features corresponding to the remaining vertices, the method further includes:
[0027] Determine whether the preset stop condition is met;
[0028] When the preset stop condition is not met, the updated target mesh information is obtained as new target mesh information, and all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information are continuously extracted;
[0029] When the preset stop condition is met, the updated target grid information is output.
[0030] A three-dimensional object classification method, the classification method comprising:
[0031] Acquire updated target mesh information corresponding to the three-dimensional object to be classified according to the three-dimensional object features described in any of the above embodiments;
[0032] The three-dimensional object to be classified is classified according to the updated target grid information.
[0033] A three-dimensional object feature acquisition device, the device comprising:
[0034] The information to be processed acquisition module is used to acquire the target grid information of the three-dimensional object to be classified;
[0035] A vertex extraction module, used to extract all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information;
[0036] A convolution operation module, used to perform a convolution operation with each vertex as a convolution center and all adjacent vertices and nearest neighbor vertices corresponding to the vertex as convolution objects, to obtain an updated feature and an importance score corresponding to each vertex;
[0037] A deletion module, used for deleting the vertices according to the importance scores;
[0038] The updating module is used to obtain the remaining vertices after the deletion process and the update features corresponding to the remaining vertices to update the target mesh information.
[0039] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.
[0040] A computer-readable storage medium stores a computer program, which implements the steps of the above method when executed by a processor.
[0041] The above-mentioned three-dimensional object feature acquisition method, device, computer equipment and storage medium calculate the updated features based on all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information, that is, the fine-grained structural information brought by the target mesh information is utilized to improve accuracy and lay the foundation for the subsequent performance improvement of 3D object classification. And only each vertex needs to be used as the convolution center, and all adjacent vertices and nearest neighbor vertices of the corresponding vertex are used as the convoluted objects for convolution operations, so that the updated features and importance scores corresponding to each vertex can be obtained, without occupying a large amount of computing resources. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 A schematic diagram of a flow chart of a method for acquiring features of a three-dimensional object in one embodiment;
[0043] Figure 2 is a schematic diagram of the structure of a multilayer perceptron network in one embodiment;
[0044] Figure 3 is a schematic flow chart of a three-dimensional object classification method in one embodiment;
[0045] Figure 4 is a structural block diagram of a three-dimensional object feature acquisition device in one embodiment;
[0046] Figure 5 is a structural block diagram of a three-dimensional object classification device in one embodiment;
[0047] Figure 6 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0048] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0049] In one embodiment, Figure 1 As shown, a method for acquiring three-dimensional object features is provided. This embodiment uses the method applied to a server as an example for illustration. It can be understood that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:
[0050] S102: Obtain target grid information of the three-dimensional object to be classified.
[0051] Specifically, mesh information refers to the model file of an object expressed in mesh format. The mesh is composed of a set of convex polygon vertices and convex polygon surfaces, also known as an unstructured mesh. Polygonal meshes are intended to represent three-dimensional object models in an easy-to-render way, and play a great role in three-dimensional visualization and other aspects. The server can obtain the target mesh information of the three-dimensional object to be classified stored in the database as needed, and use a feature extraction method based on a mixture of mesh and K nearest neighbor vertices to extract the features corresponding to the three-dimensional object, so that the subsequent server can perform 3D classification tasks based on the extracted features.
[0052] S104: Extract all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information.
[0053] The adjacent vertex refers to the vertex that is directly connected to the vertex through an edge. The nearest neighbor vertex refers to the K points closest to the vertex, where the nearest neighbor vertex can obtain the distance between other vertices and vertices, and then sort the distances. Then the server selects the vertex with a higher sorting, that is, the vertex with a closer distance to the vertex as the nearest neighbor vertex according to the preset K value.
[0054] In actual operation, the server can traverse all vertices in the target mesh information, that is, the traversed vertex is used as the current vertex, and then it can be processed by threads, that is, one thread queries the adjacent vertices corresponding to the current vertex, and the other thread queries the nearest neighbor vertex corresponding to the current vertex, thereby improving efficiency.
[0055] Optionally, the server may save the extracted adjacent vertices and nearest neighbor vertices into a set, such as set A, i.e., each vertex corresponds to a set A, in which the adjacent vertices and nearest neighbor vertices corresponding to the vertex are stored. Optionally, the set A may store the adjacent vertices and nearest neighbor vertices in the form of entries in a table, i.e., one entry in a table stores the adjacent vertices and nearest neighbor vertices corresponding to a vertex.
[0056] S106: Taking each vertex as the convolution center, performing a convolution operation on all adjacent vertices and the nearest neighbor vertices of the corresponding vertex as convolution objects, and obtaining an updated feature and importance score corresponding to each vertex.
[0057] Specifically, the convolved object is all adjacent vertices and nearest neighbor vertices, which constitute the convolution receptive field, that is, the local grid information in the grid. The convolution center is relative to the convolution kernel. Each vertex is taken as the convolution center, and multiple convolution kernel areas are delineated according to the step size, so that the server convolves the delineated convolution kernel area with all adjacent vertices and nearest neighbor vertices in turn to obtain the updated features and importance scores corresponding to each vertex. The convolution kernel parameters are calculated by the difference between the spatial coordinates of the convolution center and the corresponding convolved object through a multi-layer perceptron network (MLP).
[0058] S108: Deleting vertices according to importance scores.
[0059] Specifically, the deletion process refers to deleting vertices with low scores according to the importance scores, thereby leaving vertices with high scores, that is, leaving vertices with more detailed information. For example, the server can retain the first N% of the vertices with high importance scores according to pre-set parameters, and delete the rest.
[0060] S110: Acquire the remaining vertices after the deletion process and the update features corresponding to the remaining vertices to update the target mesh information.
[0061] Specifically, the server updates the target mesh information according to the remaining vertices after the deletion process and the updated features corresponding to the remaining vertices, that is, the remaining vertices are used as vertices in the updated target mesh, and the updated features are used as features of each vertex.
[0062] The above three-dimensional object feature acquisition method calculates the updated features based on all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information, that is, it utilizes the fine-grained structural information brought by the target mesh information to improve accuracy and lay the foundation for the subsequent performance improvement of 3D object classification. And it only needs to take each vertex as the convolution center and all adjacent vertices and nearest neighbor vertices of the corresponding vertex as the convoluted objects for convolution operations, that is, the updated features and importance scores corresponding to each vertex can be obtained, without occupying a large amount of computing resources.
[0063] In one of the embodiments, all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information are extracted, including: extracting all adjacent vertices of each vertex in the target mesh information; extracting the corresponding nearest neighbor vertices of each vertex in the target mesh information according to the K nearest neighbor algorithm.
[0064] Specifically, for a given target mesh information M=(V, T) of a target model to be classified, V is a set of all vertices, and T is a set of all triangular faces, and each triangular face ti∈T contains three vertices.
[0065] For each vertex vi∈V, the server finds all other adjacent vertex sets Vmesh by traversing all triangles T according to the mesh structure. In addition, the server finds the K nearest vertex sets Vdis of the vertex vi in 3D space according to the spatial Euclidean distance. Finally, the server merges these two vertex sets to obtain Vjoint=Vmesh∪Vdis, and saves each pair (vi, Vjoint) as an entry in table A.
[0066] Optionally, the search for adjacent vertices and nearest neighbor vertices may be performed in parallel to improve processing efficiency.
[0067] In the above embodiment, the nearest neighbor vertices in the mesh information are extracted by the K-nearest neighbor algorithm, and the adjacent vertices are extracted, which can make full use of the fine-grained structural information brought by the model mesh, help improve the performance of subsequent classification tasks, and require relatively few computing resources.
[0068] In one embodiment, the corresponding nearest neighbor vertex of each vertex in the target mesh information is extracted according to the K nearest neighbor algorithm, including: obtaining a K value set according to different convolution receptive fields; and extracting the corresponding nearest neighbor vertex of each vertex in the target mesh information according to the K value and through the K nearest neighbor algorithm.
[0069] Specifically, the K value here can be set in advance, which is related to the difficulty of the project, the field to which it belongs, and the experience of the personnel. The larger the K, the more nearest vertices are selected, that is, the larger the convolution receptive field. That is to say, before the project starts, the server can output prompt information to facilitate the user to input the set K value, so that the server extracts the corresponding nearest neighbor vertices of each vertex in the target grid information according to the K value and the K nearest neighbor algorithm, that is, selects K vertices that are closest to the current vertex. Thus, the server can calculate the final output features according to the selected vertices according to the above embodiment, and debug according to the calculation results, that is, the accuracy of the final output features obtained and the resource occupancy, until the appropriate K value is selected.
[0070] In the above embodiment, the K value can be set as needed, thereby increasing the application scope of the three-dimensional object feature acquisition method.
[0071] In one embodiment, specifically, see Figure 2 As shown, Figure 2Schematic diagram of the structure of a multilayer perceptron network in an embodiment. In this embodiment, the convolution kernel parameters are calculated by the multilayer perceptron network. Specifically, the calculation method of the convolution kernel parameters corresponding to the convolution operation includes: calculating the spatial coordinate difference between each convolution center and the corresponding convolved object; and inputting the spatial coordinate difference into the multilayer perceptron network to calculate the convolution kernel parameters.
[0072] Specifically, the convolution kernel parameter is a convolution kernel parameter from the convolution center to each convolved vertex, which is calculated by the difference between the spatial coordinates of the convolution center and the convolved object through a multi-layer perceptron network (MLP). For example, assuming that pc and pv are the spatial coordinates of a convolution center and a convolution object, respectively, the convolution kernel parameter f between the convolution center and the convolution object is f = MLP (pv – pc). Figure 2 The input is the relative coordinate, and the output is the convolution kernel parameter of the corresponding channel.
[0073] In this way, the server uses the vertex vi in each entry in table A as the convolution center, and the vertices in Vjoint corresponding to vi as the convolution objects to perform convolution operations to obtain new features of these vertices and the importance score of the convolution center.
[0074] In the above embodiment, the convolution kernel parameters are calculated by the difference between the convolution center and the spatial coordinates of the convolved object through a multi-layer perceptron network (MLP), which is relatively simple.
[0075] In one of the embodiments, deleting vertices according to importance scores includes: obtaining preset downsampling operation parameters; sorting the vertices according to the importance scores; extracting a number of vertices corresponding to the downsampling operation parameters from the sorted vertices, and deleting the unextracted vertices.
[0076] Specifically, the downsampling operation parameter is used to characterize the number of vertices to be deleted during the deletion process. For example, if there are B vertices originally and B / 4 vertices are retained, it is necessary to downsample by 4 times, that is, to retain the first 25% of the vertices. Therefore, the server can set the downsampling operation parameter by setting the downsampling by n times, or retaining the first N% of the vertices (rounded to an integer).
[0077] The server sorts the vertices according to the importance score, and then extracts the number of vertices corresponding to the downsampling operation parameter according to the downsampling operation parameter, and deletes the unextracted vertices. For example, if the downsampling operation parameter is to retain the first N% vertices, then the first N% vertices after sorting are extracted, and the other points are deleted. If the downsampling operation parameter is to downsample n times, then the first 1 / n% points after sorting are extracted, and the other points are deleted.
[0078] In addition, it should be noted that the downsampling operation parameters can be set in advance, which is related to the difficulty of the project, the field to which it belongs, and the experience of the personnel. Different downsampling operation parameters will result in different downsampling operations and thus different output sizes.
[0079] In one embodiment, obtaining the remaining vertices after the deletion process and the update features corresponding to the remaining vertices to update the target mesh information includes: obtaining the remaining vertices after the deletion process and the update features corresponding to the remaining vertices, and updating the target mesh information by using a triangle edge collapse method with an importance score as a collapse cost. Specifically, the update of the target mesh information can be performed by using a triangle edge collapse method, wherein the collapse cost is replaced by the importance score.
[0080] In one of the embodiments, after obtaining the remaining vertices after the deletion process and the update features corresponding to the remaining vertices to update the target mesh information, it also includes: determining whether a preset stop condition is met; when the preset stop condition is not met, obtaining the updated target mesh information as new target mesh information, and continuing to extract all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information; when the preset stop condition is met, outputting the updated target mesh information.
[0081] Specifically, the preset stop condition may be manually set, which is related to the difficulty of the project, the field to which it belongs, and the experience of the personnel. For example, the preset stop condition may be to stop after a fixed number of repetitions.
[0082] That is, each time the server updates the target mesh information, it determines whether the preset stop condition is met, such as whether the preset number of times is reached. If so, the updated target mesh information is output, including the remaining vertices after the deletion process and the update features corresponding to the remaining vertices. If not, the updated target mesh information is obtained as the new target mesh information, and all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information are continuously extracted, that is, the next mesh update is performed.
[0083] In one embodiment, a complete embodiment of the above-mentioned three-dimensional object feature acquisition method is given, and the three-dimensional object feature acquisition method may include:
[0084] For a given target model to be classified M = (V, T), where V is the set of all vertices and T is the set of all triangles, each triangle ti∈T contains three vertices. For each vertex vi∈V, the server first finds all other adjacent vertex sets Vmesh according to the mesh structure (by traversing all triangles T), and then finds its K nearest vertex sets Vdis according to the spatial Euclidean distance. Finally, the server merges the two vertex sets to obtain Vjoint = Vmesh∪Vdis, and saves each pair (vi, Vjoint) as an entry in table A.
[0085] Secondly, the server uses the vertex vi in each entry in Table A as the convolution center, and uses the vertex in Vjoint corresponding to vi as the convolution object to perform convolution operations to obtain the new features of these vertices and the importance score confidence of the convolution center. The convolution kernel parameters (including one convolution kernel parameter from the convolution center to each convolution vertex) are calculated by the spatial coordinate difference between the convolution center and the convolution object through a multi-layer perceptron network (MLP). For example, assuming that pc and pv are the spatial coordinates of a convolution center and a convolution object, respectively, the convolution kernel parameter f between the convolution center and the convolution object is MLP (pv–pc).
[0086] The server sorts all convolution centers (vertices) according to the importance score confidence calculated above, retains the first N% (rounded to integers), and reconstructs the mesh structure using the triangle edge collapse method (where the collapse cost is replaced by the importance score confidence).
[0087] Finally, the server repeats the above steps as required until the stop condition is met to obtain the final output features. The stop condition is set manually, such as stopping after a fixed number of repetitions.
[0088] The above three-dimensional object feature acquisition method calculates the updated features based on all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information, that is, it utilizes the fine-grained structural information brought by the target mesh information to improve accuracy and lay the foundation for the subsequent performance improvement of 3D object classification. And it only needs to take each vertex as the convolution center and all adjacent vertices and nearest neighbor vertices of the corresponding vertex as the convoluted objects for convolution operations, that is, the updated features and importance scores corresponding to each vertex can be obtained, without occupying a large amount of computing resources.
[0089] In one embodiment, see Figure 3, provides a three-dimensional object classification method. This embodiment uses the method applied to a server as an example for illustration. It can be understood that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:
[0090] S302: Acquire updated target mesh information corresponding to the three-dimensional object to be classified according to the three-dimensional object features described in any of the above embodiments.
[0091] Specifically, the specific method for obtaining the updated target grid information can be found above and will not be repeated here.
[0092] S304: Classify the three-dimensional object to be classified according to the updated target grid information.
[0093] Specifically, after acquiring the updated target grid information, the server may input the updated features in the updated target grid information into a preset classification model to classify the three-dimensional object.
[0094] In the above embodiment, the fine-grained structural information brought by the model mesh is fully utilized, which helps to improve the performance of the classification task while requiring relatively less computing resources.
[0095] It should be understood that although Figure 1 and Figure 3 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 Figure 3 At least part of the steps may include multiple steps or multiple stages. These steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed in turn or alternately with other steps or at least part of the steps or stages in other steps.
[0096] In one embodiment, if Figure 4 As shown, a three-dimensional object feature acquisition device is provided, comprising: a to-be-processed information acquisition module 100, a vertex extraction module 200, a convolution operation module 300, a deletion module 400 and an update module 500, wherein:
[0097] The to-be-processed information acquisition module 100 is used to acquire target grid information of the three-dimensional object to be classified;
[0098] A vertex extraction module 200 is used to extract all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information;
[0099] A convolution operation module 300 is used to perform a convolution operation with each vertex as a convolution center and all adjacent vertices and nearest neighbor vertices of the corresponding vertex as convolution objects to obtain an updated feature and importance score corresponding to each vertex;
[0100] A deletion module 400 is used to delete vertices according to importance scores;
[0101] The updating module 500 is used to obtain the remaining vertices after the deletion process and the update features corresponding to the remaining vertices to update the target mesh information.
[0102] In one embodiment, the vertex extraction module 200 includes:
[0103] An adjacent vertex extraction unit, used for extracting all adjacent vertices of each vertex in the target mesh information;
[0104] The nearest neighbor vertex extraction unit is used to extract the nearest neighbor vertex corresponding to each vertex in the target mesh information according to the K nearest neighbor algorithm.
[0105] In one embodiment, the nearest neighbor vertex extraction unit includes:
[0106] A K value acquisition subunit is used to obtain the K value set according to different convolution receptive fields;
[0107] The extraction subunit is used to extract the nearest neighbor vertex corresponding to each vertex in the target mesh information according to the K value and through the K nearest neighbor algorithm.
[0108] In one embodiment, the three-dimensional object feature acquisition device may further include:
[0109] A spatial coordinate difference calculation module, used to calculate the spatial coordinate difference between each convolution center and the corresponding convolved object;
[0110] The convolution kernel parameter calculation module is used to input the spatial coordinate difference into the multi-layer perceptron network to calculate the convolution kernel parameters.
[0111] In one embodiment, the deletion module 400 includes:
[0112] A parameter acquisition unit, used to acquire preset downsampling operation parameters;
[0113] A sorting unit, used to sort vertices according to their importance scores;
[0114] The deleting unit is used to extract vertices corresponding to the downsampling operation parameter from the sorted vertices and delete the unextracted vertices.
[0115] In one of the embodiments, the update module 500 is further used to obtain the remaining vertices after the deletion process and the update features corresponding to the remaining vertices, and to update the target mesh information by using the triangle edge collapse method with the importance score as the collapse cost.
[0116] In one embodiment, the three-dimensional object feature acquisition device may further include:
[0117] A judgment module, used to judge whether a preset stop condition is met;
[0118] The processing module is used to obtain the updated target mesh information as the new target mesh information when the preset stop condition is not met, and continue to extract all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information; when the preset stop condition is met, the updated target mesh information is output.
[0119] For the specific definition of the three-dimensional object feature acquisition device, please refer to the definition of the three-dimensional object feature acquisition method above, which will not be repeated here. Each module in the above-mentioned three-dimensional object feature acquisition device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0120] In one embodiment, if Figure 5 As shown, a three-dimensional object classification device is provided, including: a grid information extraction module 600 and a classification module 700, wherein:
[0121] A mesh information extraction module 600, configured to obtain updated target mesh information corresponding to the three-dimensional object to be classified according to the three-dimensional object features described in any of the above embodiments;
[0122] The classification module 700 is used to classify the three-dimensional object to be classified according to the updated target grid information.
[0123] For the specific definition of the three-dimensional object classification device, please refer to the definition of the three-dimensional object classification method above, which will not be repeated here. Each module in the above-mentioned three-dimensional object classification device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0124] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store target grid information, update features and updated target grid information. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a three-dimensional object feature acquisition method is implemented.
[0125] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0126] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the following steps when executing the computer program: obtaining target mesh information of a three-dimensional object to be classified; extracting all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information; performing a convolution operation with each vertex as a convolution center and all adjacent vertices and nearest neighbor vertices of the corresponding vertex as convolved objects to obtain an updated feature and an importance score corresponding to each vertex; deleting the vertices according to the importance scores; and obtaining the vertices remaining after the deletion and the updated features corresponding to the remaining vertices to update the target mesh information.
[0127] In one embodiment, the extraction of all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information involved in the execution of a computer program by a processor includes: extracting all adjacent vertices of each vertex in the target mesh information; extracting the corresponding nearest neighbor vertices of each vertex in the target mesh information according to a K-nearest neighbor algorithm.
[0128] In one embodiment, when a processor executes a computer program, extracting the corresponding nearest neighbor vertex of each vertex in the target mesh information according to the K nearest neighbor algorithm includes: obtaining a K value set according to different convolution receptive fields; and extracting the corresponding nearest neighbor vertex of each vertex in the target mesh information according to the K value and through the K nearest neighbor algorithm.
[0129] In one embodiment, a method for calculating convolution kernel parameters corresponding to a convolution operation involved in executing a computer program by a processor includes: calculating the spatial coordinate difference between each convolution center and the corresponding convolved object; and inputting the spatial coordinate difference into a multilayer perceptron network to calculate the convolution kernel parameters.
[0130] In one embodiment, the deleting process of vertices according to importance scores involved in executing a computer program by a processor includes: obtaining preset downsampling operation parameters; sorting the vertices according to the importance scores; extracting a number of vertices corresponding to the downsampling operation parameters from the sorted vertices, and deleting the unextracted vertices.
[0131] In one embodiment, when a processor executes a computer program, the process of obtaining the remaining vertices after the deletion process and the update features corresponding to the remaining vertices is performed to update the target mesh information, including: obtaining the remaining vertices after the deletion process and the update features corresponding to the remaining vertices, and updating the target mesh information through the triangle edge collapse method with the importance score as the collapse cost.
[0132] In one embodiment, after the processor executes a computer program involving obtaining the remaining vertices after the deletion process and updating the corresponding features of the remaining vertices to update the target mesh information, it also includes: determining whether a preset stop condition is met; when the preset stop condition is not met, obtaining the updated target mesh information as new target mesh information, and continuing to extract all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information; when the preset stop condition is met, outputting the updated target mesh information.
[0133] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the following steps when executing the computer program: obtaining updated target grid information corresponding to the three-dimensional object to be classified according to the three-dimensional object features described in any of the above embodiments; and classifying the three-dimensional object to be classified according to the updated target grid information.
[0134] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented: obtaining target mesh information of a three-dimensional object to be classified; extracting all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information; performing a convolution operation with each vertex as a convolution center and all adjacent vertices and nearest neighbor vertices of the corresponding vertex as convolved objects to obtain an updated feature and an importance score corresponding to each vertex; deleting the vertices according to the importance scores; and obtaining the vertices remaining after the deletion and the update features corresponding to the remaining vertices to update the target mesh information.
[0135] In one embodiment, the computer program executed by a processor involves extracting all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information, including: extracting all adjacent vertices of each vertex in the target mesh information; extracting the corresponding nearest neighbor vertices of each vertex in the target mesh information according to the K nearest neighbor algorithm.
[0136] In one embodiment, when a computer program is executed by a processor, the steps involved in extracting the corresponding nearest neighbor vertices of each vertex in the target mesh information according to the K nearest neighbor algorithm include: obtaining a K value set according to different convolution receptive fields; and extracting the corresponding nearest neighbor vertices of each vertex in the target mesh information according to the K value and through the K nearest neighbor algorithm.
[0137] In one embodiment, a method for calculating convolution kernel parameters corresponding to the convolution operation involved in executing a computer program by a processor includes: calculating the spatial coordinate difference between each convolution center and the corresponding convolved object; and inputting the spatial coordinate difference into a multilayer perceptron network to calculate the convolution kernel parameters.
[0138] In one embodiment, the deleting process of vertices according to importance scores involved in the execution of a computer program by a processor includes: obtaining preset downsampling operation parameters; sorting the vertices according to the importance scores; extracting a number of vertices corresponding to the downsampling operation parameters from the sorted vertices, and deleting the unextracted vertices.
[0139] In one embodiment, when a computer program is executed by a processor, the steps involved in obtaining the remaining vertices after deletion processing and the update features corresponding to the remaining vertices to update the target mesh information include: obtaining the remaining vertices after deletion processing and the update features corresponding to the remaining vertices, and updating the target mesh information through the triangle edge collapse method with the importance score as the collapse cost.
[0140] In one embodiment, after the computer program is executed by a processor and obtains the remaining vertices after the deletion processing and the update features corresponding to the remaining vertices to update the target mesh information, it also includes: determining whether a preset stop condition is met; when the preset stop condition is not met, obtaining the updated target mesh information as new target mesh information, and continuing to extract all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information; when the preset stop condition is met, outputting the updated target mesh information.
[0141] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented: obtaining updated target grid information corresponding to the three-dimensional object to be classified according to the three-dimensional object features described in any of the above embodiments; and classifying the three-dimensional object to be classified according to the updated target grid information.
[0142] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).
[0143] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0144] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A method for acquiring features of a three-dimensional object, characterized in that: The method comprises: Obtaining target grid information of the three-dimensional object to be classified; Extracting all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information; Taking each vertex as the convolution center, performing a convolution operation on all adjacent vertices and nearest neighbor vertices corresponding to the vertex as convolution objects, and obtaining an updated feature and importance score corresponding to each vertex; Deleting the vertices according to the importance scores; Acquire the remaining vertices after the deletion process and the update features corresponding to the remaining vertices to update the target mesh information; The step of extracting all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information includes: Extracting all adjacent vertices of each vertex in the target mesh information; Extracting the nearest neighbor vertex corresponding to each vertex in the target mesh information according to the K nearest neighbor algorithm; The step of extracting the nearest neighbor vertex corresponding to each vertex in the target mesh information according to the K nearest neighbor algorithm includes: Get the K value set according to different convolution receptive fields, where the K value is obtained before the project starts. The K value is input, and the corresponding nearest neighbor vertex of each vertex in the target mesh information is extracted by the K nearest neighbor algorithm. The final output feature is obtained according to the selected vertex. The final output feature is debugged based on the accuracy and resource occupancy of the final output feature. The larger the K value, the larger the convolution receptive field; According to the K value, the nearest neighbor vertex corresponding to each vertex in the target mesh information is extracted by a K nearest neighbor algorithm.
2. The method according to claim 1, characterized in that: The calculation method of the convolution kernel parameters corresponding to the convolution operation includes: Calculate the spatial coordinate difference between each convolution center and the corresponding convolved object; The spatial coordinate difference is input into a multi-layer perceptron network to calculate the convolution kernel parameters.
3. The method according to claim 1 or 2, characterized in that: The deleting process of the vertices according to the importance scores includes: Get the preset downsampling operation parameters; sorting the vertices according to the importance scores; Extract vertices whose number corresponds to the downsampling operation parameter from the sorted vertices, and delete the unextracted vertices.
4. The method according to claim 3, characterized in that The obtaining of the remaining vertices after the deletion process and the update features corresponding to the remaining vertices to update the target mesh information includes: The remaining vertices after the deletion process and the update features corresponding to the remaining vertices are obtained, and the target mesh information is updated by using the triangle edge collapse method with the importance score as the collapse cost.
5. The method according to claim 1 or 2, characterized in that: After obtaining the remaining vertices after the deletion process and updating the target mesh information with the update features corresponding to the remaining vertices, the method further includes: Determine whether the preset stop condition is met; When the preset stop condition is not met, the updated target mesh information is obtained as new target mesh information, and all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information are continuously extracted; When the preset stop condition is met, the updated target grid information is output.
6. A three-dimensional object classification method, characterized in that: The classification method includes: Acquire updated target mesh information corresponding to the three-dimensional object to be classified according to the three-dimensional object feature described in any one of claims 1 to 5 above; The three-dimensional object to be classified is classified according to the updated target grid information.
7. A three-dimensional object feature acquisition device, characterized in that: The device comprises: The information to be processed acquisition module is used to acquire the target grid information of the three-dimensional object to be classified; A vertex extraction module, used to extract all adjacent vertices and nearest neighbor vertices of each vertex in the target mesh information; A convolution operation module, used to perform a convolution operation with each vertex as a convolution center and all adjacent vertices and nearest neighbor vertices corresponding to the vertex as convolution objects, to obtain an updated feature and an importance score corresponding to each vertex; A deletion module, used for deleting the vertices according to the importance scores; An updating module, used for acquiring the remaining vertices after the deletion process and the update features corresponding to the remaining vertices to update the target mesh information; The vertex extraction module comprises: An adjacent vertex extraction unit, used for extracting all adjacent vertices of each vertex in the target mesh information; A nearest neighbor vertex extraction unit, used for extracting the nearest neighbor vertex corresponding to each vertex in the target mesh information according to a K nearest neighbor algorithm; The nearest neighbor vertex extraction unit comprises: The K value acquisition subunit is used to obtain the K value set according to different convolution receptive fields, wherein the K value is obtained before the project starts, the input K value is obtained, the corresponding nearest neighbor vertex of each vertex in the target mesh information is extracted by the K nearest neighbor algorithm, and the final output feature is obtained according to the selected vertex. The final output feature is debugged based on the accuracy and resource occupancy of the final output feature. The larger the K value, the larger the convolution receptive field; The extraction subunit is used to extract the nearest neighbor vertex corresponding to each vertex in the target mesh information according to the K value and through the K nearest neighbor algorithm.
8. The device according to claim 7, characterized in that The device also includes: A spatial coordinate difference calculation module, used to calculate the spatial coordinate difference between each convolution center and the corresponding convolved object; The convolution kernel parameter calculation module is used to input the spatial coordinate difference into the multi-layer perceptron network to calculate the convolution kernel parameters.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 or 6 are implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 or 6 are implemented.
