Point cloud classification / segmentation method based on local geometric contour and global topological structure maintenance

Abstract

The invention discloses a point cloud classification / segmentation method based on local geometric contour and global topological structure maintenance, which comprises the following steps of: firstly, constructing a neighborhood relation graph according to a spatial distance through a neighborhood relation construction module, and accurately capturing complex geometric and semantic information in point cloud data through spatial distance calculation and a k-nearest neighbor algorithm; meanwhile, a neighborhood relation graph self-diffusion mechanism further enhances the relevance between local and global features of the point cloud, so that the model can better capture the geometric contour (shape) and global topological structure features of the point cloud, the geometric distribution characteristics in the point cloud data are fully considered, and the accuracy and robustness of feature extraction are improved. The sensitivity of the model to the local geometric details of the point cloud is enhanced, the efficiency of local feature aggregation is effectively enhanced, the influence of redundant information is reduced, feature redundancy caused by a method which only depends on a nearest neighbor method to extract features is avoided, the precision of point cloud classification and segmentation tasks is improved, and the accuracy of point cloud classification and segmentation is improved. And meanwhile, sharing of global information is enhanced, and the accuracy and robustness of point cloud classification and segmentation tasks are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of dialogue recommendation, and more specifically, relates to a point cloud classification / segmentation method based on maintaining local geometric contours and global topological structures. Background Art

[0002] Deep neural networks have been widely used in computer vision and pattern recognition tasks. In this context, point cloud analysis, as a hot research direction in the field of three-dimensional understanding, has received widespread attention from academia and industry in recent years. With the advancement of technology, the means of acquiring point cloud data have become increasingly intelligent and convenient, with a wide variety of methods, such as optical radar (LiDAR) laser detection, three-dimensional model calculation to generate point clouds, and three-dimensional reconstruction through two-dimensional images. In the process of point cloud processing, point cloud data of the same category usually have similar reflection intensity, color and other information, and often present similar macroscopic features and local geometric features. Therefore, these features can be used to form feature vectors to classify and extract laser point clouds. Point cloud classification, as a basic point cloud analysis task, is widely used in many fields such as security detection, target object recognition, medical image analysis, and three-dimensional reconstruction.

[0003] Unlike images represented by regular two-dimensional grids, point clouds are an unordered and irregular set of points. In addition, the natural characteristics of point clouds, such as sparsity and noise, further limit the performance of traditional algorithms. Therefore, it is challenging to directly apply standard deep learning techniques to point cloud data. Nevertheless, thanks to the rapid development of deep neural networks, many point cloud analysis tasks have made significant progress, including three-dimensional shape classification, component segmentation, and semantic segmentation. Limited by the unstructured form of point clouds, existing methods mainly rely on neighborhood point feature aggregation algorithms, which implement message passing between non-continuous points in deep learning. Current feature aggregation algorithms are mainly divided into local feature aggregation and non-local feature aggregation. For example, in the invention patent application published on October 20, 2023 with publication number CN116912561A, a point cloud data classification and segmentation method based on a multi-view adaptive graph convolutional network is disclosed. By constructing a graph network in the offline stage, the collected point cloud data is input into the adaptive rotation matrix generator to obtain the adaptively rotated point cloud data and generate three multi-view projection images; the point cloud data and the three multi-view projection images are respectively constructed to obtain global information graph data and local information graph data, and then respectively input into the global feature extraction network and the local feature extraction network, and after extracting the global features and local features, the two are fused and input into the functional neural network output head, and the loss function is calculated according to the obtained results to realize the training of the graph network; in the online stage, the point cloud data collected in real time is input into the trained graph network to obtain the point cloud classification or segmentation result. The present invention obtains the optimized observation angle by adaptively rotating the point cloud data, flexibly adjusts the position of the angle according to the geometric features and spatial distribution of the point cloud data, and introduces the connection between points of different depths on a specific projection surface, so as to better capture the local and global features of the point cloud data, and make the composition more accurate and robust. The global feature extraction network of the invention patent application is a six-layer tandem neural network with a maximum pooling layer, and the local feature extraction network is a two-layer tandem neural network, which simply uses a convolutional neural network to extract global features and local features. Some other methods extract features through different distance measurement methods or different neighborhood relationship graph construction methods.

[0004] However, existing feature aggregation methods have the following problems: 1. Local neighborhood construction methods based on spatial distance usually rely on the geometric distance between points to select neighbors. This method is easily affected by changes in the density of point cloud distribution in point clouds. In sparse areas, neighbor points may be insufficient, resulting in incomplete local feature information. 2. Methods based on feature distance can transcend spatial limitations when capturing global structures and capture points that are far away but have similar features. However, this method is prone to feature redundancy problems because its reliance on features causes similar points to be frequently included in the neighborhood, especially when feature distributions are highly similar or repeated. This redundancy can lead to excessive concentration of information, resulting in a decrease in information differences within the neighborhood, thereby weakening the model's sensitivity to local details. 3. Existing methods use simple splicing methods to fuse local and global information. Although it can retain information at both scales to a certain extent, this method often leads to insufficient use of information. For point cloud scenes with complex spatial structures or rich semantic levels, splicing alone cannot allow local features to fully play a role in the global context. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a point cloud classification / segmentation method based on local geometric contours and global topological structure preservation, so as to fully consider the geometric contour information, i.e. the geometric features of the point location, and the topological structure information for feature aggregation, thereby improving the accuracy and robustness of point cloud classification / segmentation.

[0006] To achieve the above-mentioned object of the invention, the present invention is based on a point cloud classification / segmentation method that maintains local geometric contours and global topological structures, characterized in that it includes the following steps:

[0007] (1) Use a multi-stage cascade feature extractor to extract features from point cloud data

[0008] The first-level feature extractor takes the three-dimensional coordinates of the midpoints in the point cloud data as input features, and the subsequent feature extractors take the features output by the previous-level feature extractor as input features. Each level of feature extractors includes a neighborhood relationship construction module, a neighborhood relationship diffusion module, a feature extraction network, and a farthest point sampling module. In each level of feature extractors, the input features are input into the neighborhood relationship construction module respectively;

[0009] 1.1) Get the neighborhood relationship graph

[0010] In the neighborhood relationship construction module, the spatial distance between each point and other points is first calculated based on the input features, and then the neighborhood point set of each point is determined from the point cloud data according to the spatial distance, that is, the neighborhood relationship graph A is constructed by selecting the first k minimum spatial distances, and the asymmetric neighborhood relationship graph A is converted into a symmetric matrix and normalized. The normalization of the neighborhood relationship graph is:

[0011]

[0012] Among them, A represents the neighborhood relationship graph, and D represents the degree matrix of the neighborhood relationship graph;

[0013] 1.2) Using self-diffusion mechanism to propagate and strengthen features

[0014] In the neighborhood relationship diffusion module, the acquired neighborhood relationship graph is used to propagate and enhance features using the self-diffusion mechanism, so as to quickly expand to the surface contour points of the point cloud. Finally, the top k maximum values ​​and indexes on the specified dimension in the diffused neighborhood relationship graph are determined as the input features of the feature extraction network. The calculation formula of the self-diffusion mechanism is:

[0015]

[0016] in, The subscript n indicates the number of diffusions, the superscript n+1 and ni indicate the power, and α is a custom parameter used to adjust the local neighborhood field of view;

[0017] 1.3) Extract features

[0018] The feature extraction network is a multi-layer cascade neural network, each layer of which consists of two two-dimensional convolutional layers and a maximum pooling layer, and its update function is a neighborhood feature aggregation function with residual connections, specifically:

[0019]

[0020] Among them, h i (n) is the input feature of the ith point of the nth layer neural network, h i (n+1) is the output feature of the i-th point of the n-th layer neural network, p i is the point cloud coordinate of the i-th point, N(i) represents the neighborhood point set of the i-th point, j represents the j-th point in the neighborhood point set N(i), p j is the point cloud coordinate of the jth point in the neighborhood point set N(i), Θ and Φ are the parameters of the two 2D convolutional layers, max() is the maximum pooling function, and the comma in the brackets indicates concatenation;

[0021] The input features of the feature extraction network are input into the first layer of the neural network, and the output features of the last layer of the feature extraction network are the extracted features;

[0022] 1.4) Sampling at the farthest point

[0023] The farthest point sampling method is used in the farthest point sampling module. According to the extracted features, the farthest point is selected iteratively to extract a representative point set from the point cloud data. The features corresponding to each point in the representative point set constitute the features of the point cloud data.

[0024] (2) Point cloud classification / segmentation

[0025] The classification / segmentation output head is used to classify / segment the point cloud data according to its features to obtain the point cloud classification / segmentation results.

[0026] The object of the present invention is achieved in this way.

[0027] The point cloud classification / segmentation method based on local geometric contours and global topological structure preservation of the present invention first constructs a neighborhood relationship graph according to spatial distance through a neighborhood relationship construction module, wherein spatial distance calculation and the k-nearest neighbor algorithm accurately capture the complex geometric and semantic information in the point cloud data, and enhance the accuracy and robustness of feature extraction. At the same time, the self-diffusion mechanism of the neighborhood relationship graph further enhances the correlation between local and global features of the point cloud, so that the model can better capture the geometric contour (shape) and global topological structure features of the point cloud. In this way, the geometric distribution characteristics in the point cloud data are fully considered, the sensitivity of the model to the local geometric details of the point cloud is enhanced, the efficiency of local feature aggregation is effectively enhanced, the influence of redundant information is reduced, and the feature redundancy caused by the method of extracting features only by relying on the nearest neighbor method is avoided, the accuracy of point cloud classification and segmentation tasks is improved, and at the same time, the sharing of global information is strengthened to achieve a significant improvement in the accuracy and robustness of point cloud classification and segmentation tasks.

[0028] Through the organic combination of modular design and innovative technology, the present invention shows significant advantages in task adaptability, scalability and multi-view composition. When the point cloud is sparse or the density is uneven, the adaptive optimization of neighborhood selection can be achieved by replacing the k-nearest neighbor algorithm with the ball query algorithm and adjusting the ball query radius or the number of self-diffusion of the neighborhood relationship graph. At the same time, the neighborhood relationship graph self-diffusion algorithm is essentially a form of feature propagation, which can be naturally and seamlessly combined with other graph neural network technologies (such as graph convolution, graph attention, etc.) to further enhance the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flowchart of a specific implementation of the point cloud classification / segmentation method based on local geometric contours and global topological structure preservation of the present invention;

[0030] Figure 2 It is the overall network structure diagram of the point cloud classification / segmentation method based on the local geometric contour and global topological structure preservation of the present invention;

[0031] Figure 3It is a self-diffusion effect diagram of the neighborhood relationship graph of any point in the point cloud data, where (a) is before self-diffusion and (b) is after self-diffusion;

[0032] Figure 4 This is the point cloud segmentation effect diagram of 'airplane', 'chair', 'table' and 'lamp' in the ShapNetPart dataset. DETAILED DESCRIPTION

[0033] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0034] Example 1

[0035] Figure 1 It is a flow chart of a specific implementation of the point cloud classification / segmentation method based on local geometric contours and global topological structure preservation of the present invention.

[0036] In this embodiment, if Figure 1 As shown, the point cloud classification / segmentation method based on local geometric contours and global topological structure preservation of the present invention includes the following steps:

[0037] Step S1: Use a multi-stage cascade feature extractor to extract features from point cloud data

[0038] In this embodiment, if Figure 2 As shown, four-stage series feature extractors 1 to 4 are used to extract fusion features from point cloud data. The first-stage feature extractor uses the three-dimensional coordinates of the midpoints in the point cloud data as input features, and the subsequent feature extractors use the features output by the previous-stage feature extractor as input features. Each stage of feature extractors includes a neighborhood relationship construction module, a neighborhood relationship diffusion module, a feature extraction network, and a farthest point sampling module. In each stage of feature extractors, the input features are respectively input into the neighborhood relationship construction module.

[0039] In this embodiment, point cloud data with three-dimensional coordinates P=[x, y, z] is obtained through laser radar (LiDAR) scanning or a three-dimensional model generation tool, and the data size is N×3, where N is the number of point cloud points.

[0040] Step S1.1: Obtain neighborhood relationship graph

[0041] In the neighborhood relationship construction module, the spatial distance between each point and other points is first calculated based on the input features, and then the neighborhood point set of each point is determined from the point cloud data according to the spatial distance, that is, the neighborhood relationship graph A is constructed by selecting the first k minimum spatial distances, and the asymmetric neighborhood relationship graph A is converted into a symmetric matrix and normalized. The normalization of the neighborhood relationship graph is:

[0042]

[0043] Among them, A represents the neighborhood relationship graph, and D represents the degree matrix of the neighborhood relationship graph.

[0044] In this embodiment, the neighborhood relationship construction module is composed of a spatial distance calculation algorithm, a k-nearest neighbor algorithm, a neighborhood relationship graph symmetry algorithm, and a neighborhood relationship graph normalization algorithm, wherein: the spatial distance calculation algorithm is used to calculate the spatial distance between each point and other points, the k-nearest neighbor algorithm is used to determine the neighborhood point set of each point from the point cloud data according to the spatial distance, the algorithm constructs a neighborhood relationship graph by selecting the first k minimum values, the neighborhood relationship graph symmetry algorithm is used to convert an asymmetric neighborhood relationship graph into a symmetric matrix to meet the positive definiteness of the matrix diffusion kernel, and the neighborhood relationship graph normalization algorithm is used to normalize the graph structure data, thereby alleviating the numerical instability problem that may arise during the diffusion of neighborhood relationship graph nodes.

[0045] In this embodiment, the spatial distance calculation algorithm is used to calculate the Euclidean distance between two points, and the calculation formula is:

[0046]

[0047] Among them, p1(x1,y1,z1) and p2(x2,y2,z2) are any two points in the point cloud.

[0048] Step S1.2: Use self-diffusion mechanism to propagate and strengthen features

[0049] In the neighborhood relationship diffusion module, the acquired neighborhood relationship graph is used to propagate and enhance features using the self-diffusion mechanism, so as to quickly expand to the surface contour points of the point cloud. Finally, the top k maximum values ​​and indexes on the specified dimension in the diffused neighborhood relationship graph are determined as the input features of the feature extraction network. The calculation formula of the self-diffusion mechanism is:

[0050]

[0051] in, The subscript n indicates the number of diffusions, the superscript n+1 and ni indicate the power, and α is a custom parameter. The smaller α is, the larger the field of view is. In this embodiment, it is set to 0.01 based on experience.

[0052] This self-diffusion mechanism can enhance the correlation between local and global features of the point cloud, enabling the model to better capture the geometric contour (shape) and global topological structure characteristics of the point cloud.

[0053] In this embodiment, n=5, the neighborhood relationship graph self-diffusion effect diagram is as follows Figure 3 As shown in the figure, the neighborhood before diffusion is the nearest neighborhood based on the center point. Since the geometric structure of the point cloud is ignored, some points belonging to different parts are included in the neighborhood of the center point, resulting in noise in feature extraction. After using the self-diffusion method, the neighborhood of the center point is gradually expanded from local to global, while avoiding the introduction of noise points. This mechanism allows the global neighborhood to be locally constrained, while local information is also refined and enhanced globally, thereby achieving global and local consistency in the entire point cloud structure, thereby more efficiently capturing multi-scale relationships in the point cloud and significantly improving information utilization and expression capabilities.

[0054] Step S1.3: Extract features

[0055] The feature extraction network described in this paper is a multi-layer cascade neural network. Each layer of the neural network consists of two two-dimensional convolutional layers and a maximum pooling layer. Its update function is a neighborhood feature aggregation function with residual connections, specifically:

[0056]

[0057] Among them, h i (n) is the input feature of the ith point of the nth layer neural network, h i (n+1) is the output feature of the i-th point of the n-th layer neural network, p i is the point cloud coordinate of the i-th point, N(i) represents the neighborhood point set of the i-th point, j represents the j-th point in the neighborhood point set N(i), p j is the point cloud coordinate of the jth point in the neighborhood point set N(i), Θ and Φ are the parameters of the two two-dimensional convolutional layers, max() is the maximum pooling function, and the comma in the brackets indicates concatenation.

[0058] The input features of the feature extraction network are input into the first layer of the neural network, and the output features of the last layer of the feature extraction network are the extracted features.

[0059] In this embodiment, the feature extraction network is a five-layer serial neural network, wherein: the first layer is a graph convolution layer with an input dimension of 3 and an output dimension of 32, the second layer is a graph convolution layer with an input dimension of 32 and an output dimension of 64, the third layer is a graph convolution layer with an input dimension of 64 and an output dimension of 128, the fourth layer is a graph convolution layer with an input dimension of 128 and an output dimension of 256, and the fifth layer is a graph convolution layer with an input dimension of 256 and an output dimension of 512.

[0060] Step S1.4: Farthest point sampling

[0061] The farthest point sampling method is used in the farthest point sampling module. Based on the extracted features, the farthest point is iteratively selected to extract a representative point set from the point cloud data while maintaining the integrity of the geometric structure information. The features corresponding to each point in the representative point set constitute the features of the point cloud data.

[0062] Step S2: Point cloud classification / segmentation

[0063] The classification / segmentation output head is used to classify / segment the point cloud data according to its features to obtain the point cloud classification / segmentation results.

[0064] In this embodiment, the classification / segmentation output head is a functional neural network output head including several layers of multi-layer perceptrons, and the output dimension is the number of categories in the classification task or the number of components to be segmented in the segmentation task.

[0065] The point cloud data features obtained in step S1 are input into the functional neural network output head, and the network weights are updated after the loss function is calculated based on the obtained results. The functional neural network output head is a multilayer perceptron with two linear layers and one maximum pooling layer in series, and its output result is a vector of the number of categories in the classification task, where the index corresponding to the maximum value is the final result of the network prediction.

[0066] The step of updating the network weights is: calculating the cross entropy error between the obtained output result and the data set label and passing it backward to update the network weights. The updating network weights use the public data set ModelNet40, in which the training set has 9843 point clouds, the test set has 2468 point clouds, and the validation set is not included. 1024 points are randomly sampled from each point cloud. The optimizer used in the training process is the Adam optimizer with a learning rate of 0.001. The training generation is 300 generations, and the final trained graph network is obtained, which includes a neighborhood relationship graph construction module, a neighborhood relationship graph diffusion module, a feature extraction network, a farthest point sampling module and a functional neural network output head.

[0067] After specific practical experiments, using PyTorch as the underlying deep learning framework and adopting the hardware environment of NVIDIA RTX4090 GPU (24GB), the above method was run. On the official test set, the average instance accuracy reached 94.1%, which is 1.2% higher than the existing baseline technology, such as Wang et al. in "Dynamic graph cnn for learning on point clouds" ([J]. Acm Transactions On Graphics (tog), 2019, 38 (5): 1-12).

[0068] Example 2

[0069] In this embodiment, the ShapeNetPart dataset is selected for training and testing, which includes 12137 training set samples, 1870 validation set samples, and 2874 test set samples. The number of points in each point cloud is about 2000, and the specific number of point clouds may vary. During the training process, the Adam optimizer with a learning rate of 0.001 is used, and the training generation number is 200 generations.

[0070] In the point cloud segmentation task, this embodiment first uses the neighborhood relationship graph construction module to construct the corresponding point cloud neighborhood relationship graph. In order to ensure the numerical stability of the graph diffusion process, the neighborhood relationship graph symmetry method and the neighborhood relationship graph normalization algorithm are further applied to symmetrize and normalize the neighborhood relationship graph. Subsequently, the feature information of the point cloud surface contour is further captured through the neighborhood relationship graph diffusion module and the feature extraction network to ensure the effective focusing and enhancement of the feature information. On this basis, the point cloud scale is reduced by the farthest point sampling method, thereby reducing the computational complexity without losing important information.

[0071] According to the requirements of a specific segmentation task, the features of the point cloud data are input into the functional neural network output head. The output head uses a multi-layer perceptron structure to generate classification or segmentation results based on the input fusion features, where the output dimension is the number of categories or components of the point cloud. In the point cloud segmentation task, the prediction result of each point is the component category to which it belongs.

[0072] In the experiment, PyTorch was used as the underlying deep learning framework, and the hardware environment of NVIDIA RTX 4090 GPU (24GB) was used for training and testing. After verification on the training set and the test set, the average instance accuracy on the test set reached 86.4%, an increase of 1.3% compared to the existing technology; the average category accuracy was 83.3%, an increase of 1.0% compared to the existing technology.

[0073] Figure 4 This is the point cloud segmentation effect diagram of 'airplane', 'chair', 'table' and 'table lamp' in the ShapNetPart dataset, where different colors represent different parts. Figure 4 It can be seen that in categories such as 'airplane', 'chair' and 'lamp', the various parts are accurately segmented, and the point cloud segmentation performance is better than the existing methods.

[0074] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A point cloud classification / segmentation method based on local geometric contours and global topological structure preservation, characterized in that: The following steps are involved: (1) Use a multi-stage cascade feature extractor to extract features from point cloud data The first-level feature extractor takes the three-dimensional coordinates of the midpoints in the point cloud data as input features, and the subsequent feature extractors take the features output by the previous-level feature extractor as input features. Each level of feature extractors includes a neighborhood relationship construction module, a neighborhood relationship diffusion module, a feature extraction network, and a farthest point sampling module. In each level of feature extractors, the input features are input into the neighborhood relationship construction module respectively; 1.1) Get the neighborhood relationship graph In the neighborhood relationship construction module, the spatial distance between each point and other points is first calculated based on the input features, and then the neighborhood point set of each point is determined from the point cloud data according to the spatial distance, that is, the neighborhood relationship graph A is constructed by selecting the first k minimum spatial distances, and the asymmetric neighborhood relationship graph A is converted into a symmetric matrix and normalized. The normalization of the neighborhood relationship graph is: Among them, A represents the neighborhood relationship graph, and D represents the degree matrix of the neighborhood relationship graph; 1.2) Using self-diffusion mechanism to propagate and strengthen features In the neighborhood relationship diffusion module, the acquired neighborhood relationship graph is used to propagate and enhance features using the self-diffusion mechanism, so as to quickly expand to the surface contour points of the point cloud. Finally, the top k maximum values ​​and indexes on the specified dimension in the diffused neighborhood relationship graph are determined as the input features of the feature extraction network. The calculation formula of the self-diffusion mechanism is: in, The subscript n indicates the number of diffusions, the superscript n+1 and ni indicate the power, and α is a custom parameter used to adjust the local neighborhood field of view; 1.3) Extract features The feature extraction network is a multi-layer cascade neural network, each layer of which consists of two two-dimensional convolutional layers and a maximum pooling layer, and its update function is a neighborhood feature aggregation function with residual connections, specifically: Among them, h i (n) is the input feature of the ith point of the nth layer neural network, h i (n+1) is the output feature of the i-th point of the n-th layer neural network, p i is the point cloud coordinate of the i-th point, N(i) represents the neighborhood point set of the i-th point, j represents the j-th point in the neighborhood point set N(i), p j is the point cloud coordinate of the jth point in the neighborhood point set N(i), Θ and Φ are the parameters of the two 2D convolutional layers, max() is the maximum pooling function, and the comma in the brackets indicates concatenation; The input features of the feature extraction network are input into the first layer of the neural network, and the output features of the last layer of the feature extraction network are the extracted features; 1.4) Sampling at the farthest point The farthest point sampling method is used in the farthest point sampling module. According to the extracted features, the farthest point is selected iteratively to extract a representative point set from the point cloud data. The features corresponding to each point in the representative point set constitute the features of the point cloud data. (2) Point cloud classification / segmentation The classification / segmentation output head is used to classify / segment the point cloud data according to its features to obtain the point cloud classification / segmentation results.

2. The point cloud classification / segmentation method based on local geometric contours and global topological structure preservation according to claim 1, characterized in that: The custom parameter in step 1.2) is set to 0.

01.

3. The point cloud classification / segmentation method based on local geometric contours and global topological structure preservation according to claim 1, characterized in that: The feature extraction network described in step 1.3) is a five-layer cascade neural network. Among them: the first layer is a graph convolution layer with an input dimension of 3 and an output dimension of 32, the second layer is a graph convolution layer with an input dimension of 32 and an output dimension of 64, the third layer is a graph convolution layer with an input dimension of 64 and an output dimension of 128, the fourth layer is a graph convolution layer with an input dimension of 128 and an output dimension of 256, and the fifth layer is a graph convolution layer with an input dimension of 256 and an output dimension of 512.

4. The point cloud classification / segmentation method based on local geometric contours and global structure preservation according to claim 1, characterized in that: The classification / segmentation output head described in step (2) is a functional neural network output head including several layers of multi-layer perceptrons, and the output dimension is the number of categories in the classification task or the number of components to be segmented in the segmentation task.

Images