CN110322453B - 3D point cloud semantic segmentation method based on position attention and auxiliary network - Google Patents

Abstract

The invention provides a 3D point cloud semantic segmentation method based on position attention and an auxiliary network, which mainly solves the problem of low segmentation precision in the prior art, and the implementation scheme is as follows: acquiring a training set T and a test set V; constructing a 3D point cloud semantic segmentation network, and setting a loss function of the network, wherein the network comprises a feature down-sampling network, a position attention module, a feature up-sampling network and an auxiliary network which are sequentially cascaded; and performing P rounds of supervised training on the segmentation network by using a training set T: adjusting network parameters according to a loss function in the training process of each round, and taking a network model with the highest segmentation precision as a trained network model after P rounds of training are completed; and inputting the test set V into the trained network model for semantic segmentation to obtain a segmentation result of each point. The method improves the semantic segmentation precision of the 3D point cloud, and can be used for automatic driving, robots, 3D scene reconstruction, quality detection, 3D mapping and smart city construction.

Description

3D point cloud semantic segmentation method based on position attention and auxiliary network

Technical Field

The present invention belongs to the field of data processing technology, and in particular relates to a 3D point cloud semantic segmentation method, which can be used for autonomous driving, robots, 3D scene reconstruction, quality inspection, 3D mapping and smart city construction.

Background Art

In recent years, with the widespread application of 3D sensors such as LiDAR and RGBD cameras in the fields of robotics and unmanned driving, the application of deep learning in 3D point cloud data has become one of the research hotspots. 3D point cloud data refers to: a set of vectors in a three-dimensional coordinate system. These vectors are usually expressed in the form of x, y, and z three-dimensional coordinates, and are generally used to represent the outer surface shape of an object. In addition, in addition to the geometric information represented by (x, y, z), it may also contain RGB color, intensity, grayscale value, depth or return times. Point cloud data is usually obtained by 3D scanning devices, such as LiDAR, RGBD cameras, etc. These sensors measure the information of a large number of points on the surface of an object in an automated way, and then output point cloud data in a certain data file. Point cloud data is disordered and unstructured and may have different densities in 3D space. This makes the application of deep learning in the study of 3D point cloud data face huge challenges.

3D point cloud semantic segmentation refers to assigning a category to each point in the input point cloud data. In early research work, 3D point cloud data is usually converted into manual voxel grid features or multi-view image features, and then sent to the deep learning network for feature extraction. This feature conversion method not only has a large amount of data, but also is computationally complex. If the resolution is reduced, the segmentation accuracy will decrease. Therefore, it is particularly important to use deep learning methods to directly process point cloud data.

In 2017, Charles R Qi et al. published a paper titled "PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation" at CVPR, which disclosed a deep learning framework for directly processing 3D point cloud data. This method uses the symmetric function of max-pooling to solve the problem of point cloud disorder, thereby extracting the global features of each point. However, this method only considers global features and ignores the local features of each point. Therefore, shortly after PointNet, Charles R Qi's team published a paper titled "PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space" at NIPS. PointNet++ is a hierarchical version of PointNet, and each layer has three stages: sampling, grouping, and feature extraction. First, some important points are selected as the center points of each local area, and then k neighboring points are selected around these center points according to the Euclidean distance. Then, the k neighboring points are used as a local point cloud to extract features using the PointNet network, and then the deep features are transmitted back to obtain the semantic segmentation results of 3D point cloud data. This method has improved accuracy compared to PointNet.

Compared with traditional methods, the above two methods directly process 3D point cloud data, are simple to calculate, effectively solve the problem of point cloud disorder and improve segmentation accuracy. However, PointNet++ does not consider the relationship between the features of each center point, that is, the context information, so the feature representation is relatively weak. In addition, PointNet++ follows the general framework of encoding and decoding and does not consider more underlying information. Therefore, the segmentation accuracy is not high and there is still room for improvement.

Summary of the invention

The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a 3D point cloud data semantic segmentation method based on position attention and auxiliary network, so as to improve the segmentation accuracy by associating the position attention of contextual features and the auxiliary network of reconstructing the underlying information.

To achieve the above object, the technical solution of the present invention includes the following steps:

(1) Download the training and test files of 3D point cloud data from the ScanNet official website, perform category statistics and block processing on them, and obtain the training set T and test set V;

(2) Construct a 3D point cloud semantic segmentation network, which includes a cascaded feature downsampling network, a position attention module, a feature upsampling network, and an auxiliary network;

(3) Use the multi-classification cross entropy loss function as the loss function of the 3D point cloud semantic segmentation network;

(4) Using the training set T, perform P rounds of supervised training on the 3D point cloud data semantic segmentation network, where P ≥ 500;

(4a) In each round of training, the network parameters are adjusted according to the loss function of the semantic segmentation network to obtain the network model;

(4b) Every P ₁ rounds, use the samples of the test set to evaluate the segmentation accuracy of the current network model. If the segmentation accuracy of the current network model is higher than that of the previously saved network model, it is saved, P ₁ ≥ 2;

(4c) After P rounds of training are completed, the network model with the highest segmentation accuracy is taken as the trained network model;

(5) Input the test set V into the trained network model for semantic segmentation to obtain the segmentation result of each point.

Compared with the prior art, the present invention has the following advantages:

The present invention constructs a 3D point cloud semantic segmentation network, and calculates the correlation between the features represented by each centroid of its input data through the position attention module therein, thereby adding contextual information to the local centroid features of the 3D point cloud semantic segmentation network; at the same time, since the underlying features of the 3D point cloud semantic segmentation network are fed back through the auxiliary network therein, the underlying information of the 3D point cloud semantic segmentation network is reconstructed, thereby effectively improving the segmentation accuracy of the 3D point cloud semantic segmentation.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is an implementation flow chart of the present invention;

FIG2 is an overall structural diagram of a 3D point cloud semantic segmentation network constructed in the present invention;

FIG3 is a structural diagram of the position attention module in the present invention.

DETAILED DESCRIPTION

The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

1 , the implementation steps of this example include the following.

Step 1: Get the training set T and the test set V.

1.1) Download the training file and test file of 3D point cloud data from ScanNet official website, where the training file contains f ₀ point cloud scenes and the test file contains f ₁ point cloud scenes. In this embodiment, f ₀ =1201 and f ₁ =312;

1.2) Use the histogram to count the number of each category of point cloud data of all f ₀ scenes in the training file, and calculate the weight w _k of each category:

Wherein, G _k represents the number of point cloud data of the kth type, M represents the number of all point cloud data, L represents the number of segmentation categories, L ≥ 2, and in this embodiment, L = 21;

1.3) For each scene in the training file, randomly select a point as the center point with coordinates (x, y, z), and take points in the range of (x-0.75, x+0.75), (y-0.75, y+0.75), and (z-0.75, z+0.75) around it to form a data block;

1.4) Set the number of sampling points N ₀ , compare the number of points in the data block obtained in (1.3) with the number of sampling points N ₀ to determine whether it is reasonable:

If the number of points in the data block is greater than the number of sampling points N ₀ , the data block is judged to be reasonable, and N ₀ points are randomly sampled in the data block to form a sample data. Otherwise, the data block is discarded, thereby obtaining a training set T. In this embodiment, N ₀ =8192;

1.5) For each of all f ₁ scenes in the test file, use a cubic window of size 1.5×1.5×3 to perform sliding window cutting. For each data block, randomly sample N ₀ points to form a sample data to obtain the test set V.

Step 2: Build a 3D point cloud semantic segmentation network.

Referring to Figure 2, the 3D point cloud semantic segmentation network constructed in this step includes a feature downsampling network, a position attention module, a feature upsampling network and an auxiliary network that are cascaded in sequence.

2.1) Set up feature downsampling network:

The feature downsampling network includes n cascaded PointSA modules, each PointSA module includes sequentially cascaded point cloud centroid sampling and grouping layers, and point cloud feature extraction layers, wherein n≥2, and in this embodiment, the parameter is set to n＝4;

个点作为质心点，其次，以

个采样的质心点为中心，使用球形搜索算法，在其特定半径r^m的范围内搜索

个点，组成一个分组。本实施例中第1个PointSA模块，设置

r¹＝0.1；第2个PointSA模块，设置

r²＝0.2；第3个PointSA模块，设置

r³＝0.4；第4个PointSA模块，设置

r⁴＝0.8；For the centroid sampling and grouping layer of the m-th PointSA module, m = 1, 2, ..., n, the iterative farthest point sampling method is first used to sample from the input point set

points as the centroid, and secondly,

The centroid of the samples is used as the center, and the spherical search algorithm is used to search within the range of its specific radius r ^m.

Points form a group. In this embodiment, the first PointSA module is set

r ¹ = 0.1; the second PointSA module, set

r ² = 0.2; The third PointSA module is set

r ³ = 0.4; the fourth PointSA module, set

^r4 = 0.8;

For the point cloud feature extraction layer of the mth PointSA module, it includes three cascaded 2D convolutional layers to extract the features of the centroid sampling and grouping layer output data, and uses the maximum pooling strategy to pool the extracted regional features. In this embodiment, the convolution kernel size of the three 2D convolution layers of the point cloud feature extraction layer of the first PointSA module is 1×1, the step size is 1, and the number of output channels is 32, 32, and 64 respectively; the convolution kernel size of the three 2D convolution layers of the point cloud feature extraction layer of the second PointSA module is 1×1, the step size is 1, and the number of output channels is 64, 64, and 128 respectively; the convolution kernel size of the three 2D convolution layers of the point cloud feature extraction layer of the third PointSA module is 1×1, the step size is 1, and the number of output channels is 128, 128, and 256 respectively; the convolution kernel size of the three 2D convolution layers of the point cloud feature extraction layer of the fourth PointSA module is 1×1, the step size is 1, and the number of output channels is 256, 256, and 512 respectively;

2.2) Set up a position attention module to calculate the correlation between the features represented by each centroid of its input data F, and obtain the feature E after position attention enhancement:

Referring to Figure 3, the working principle of this module is as follows:

第三1D卷积层V的输出特征通道数与输入数据F的特征通道数相同；2.2.1) The input data F passes through the first 1D convolutional layer Q to obtain the feature _Qi of the i-th centroid, i = 1, 2, ..., N, N represents the number of centroids of F; then passes through the second 1D convolutional layer U to obtain the feature _Uj of the j-th centroid, j = 1, 2, ..., N, and then passes through the third 1D convolutional layer V to obtain the feature _Vj of the j-th centroid; wherein the convolution kernel size of these three 1D convolutional layers Q, U, and V is 1, the step size is 1, and the number of output feature channels of the first 1D convolutional layer Q and the second 1D convolutional layer U is twice the number of feature channels of the input data F

The number of output feature channels of the third 1D convolutional layer V is the same as the number of feature channels of the input data F;

使用t_ij构成矩阵A：2.2.2) Calculate the attention influence value _tij between the features represented by each centroid:

Use t _ij to construct the matrix A:

2.2.3) Calculate position attention features

2.2.4) Output the feature E after attention enhancement:

E=[E ₁ ; E ₂ ;...; E _i ;...; E _N ],

Among them, E _i = αJ _i + _Fi represents the feature of the i-th centroid in E, α represents the weight of the position attention feature _Ji , and _Fi represents the feature of the i-th centroid of the input;

2.3) Set up the feature upsampling network:

The feature upsampling network includes a PointFP modules, a 1D convolution layer, a Dropout layer, and a 1D convolution layer for classification, which are cascaded in sequence. Each PointFP module includes a feature interpolation layer and a feature extraction layer, which are cascaded in sequence. Where a≥2, in this embodiment, the parameter is set to a＝4;

The structures of the feature interpolation layer and the feature extraction layer of the PointFP module are different, where:

For the first PointFP module, its feature interpolation layer interpolates the output features of the position attention module and cascades the interpolated features with the output features of the third PointSA module to obtain the output features of the feature interpolation layer; its feature extraction layer includes two cascaded 2D convolutional layers to further extract the output features. The convolution kernels of the two 2D convolutional layers are both 1×1 in size, 1 in stride, and the number of output channels is 256 and 256 respectively;

For the second PointFP module, its feature interpolation layer interpolates the output features of the first PointFP module and cascades the interpolated features with the output features of the second PointSA module to obtain the output features of the feature interpolation layer; its feature extraction layer includes two cascaded 2D convolutional layers to further extract the output features. The convolution kernels of the two 2D convolutional layers are both 1×1 in size, 1 in stride, and the number of output channels is 256 and 256 respectively;

For the third PointFP module, its feature interpolation layer interpolates the output features of the second PointFP module and cascades the interpolated features with the output features of the first PointSA module to obtain the output features of the feature interpolation layer; its feature extraction layer includes two cascaded 2D convolutional layers to further extract the output features. The convolution kernels of the two 2D convolutional layers are both 1×1 in size, 1 in stride, and the number of output channels is 256 and 128 respectively;

For the fourth PointFP module, its feature interpolation layer interpolates the output features of the third PointFP module to obtain the interpolated features, which are used as the output features of its feature interpolation layer; its feature extraction layer includes three cascaded 2D convolutional layers to further extract the output features. The convolution kernels of the three 2D convolutional layers are all 1×1 in size, with a step size of 1, and the number of output channels are 128, 128, and 128 respectively.

The 1D convolution layer has a convolution kernel size of 1, a step size of 1, and an output feature channel number of 128;

The retention probability of the Dropout layer is set to 0.5;

The 1D convolution layer used for classification has a convolution kernel size of 1, a step size of 1, and the number of output feature channels is set to the number of segmented categories L.

2.4) Set up auxiliary network:

The auxiliary network includes b PointAux modules cascaded in sequence and a 1D convolution layer for classification, each PointAux module includes a 1D convolution layer and a feature interpolation layer, wherein b≥1, and in this embodiment, b＝2;

For the first PointAux module, its 1D convolution layer is used to extract the features of the output data of the second PointFP module. The convolution kernel size is 1, the step size is 1, and the output feature channel is the number of segmented categories L; its feature interpolation layer is used to interpolate the features extracted by the 1D convolution layer;

For the second PointAux module, its 1D convolution layer is used to extract the features of the output data of the first PointAux module, the convolution kernel size is 1, the step size is 1, and the output feature channel is the number of segmented categories L; its feature interpolation layer is used to interpolate the features extracted by the 1D convolution layer;

The 1D convolutional layer for classification is used to classify the output features of the second PointAux module. Its convolution kernel size is 1, the stride is 1, and the number of output feature channels is set to the number of segmented categories L.

Step 3: Set the loss function of the 3D point cloud semantic segmentation network.

In this example, the multi-classification cross entropy loss function is used as the loss function of the 3D point cloud semantic segmentation network. The expression formula is as follows:

Wherein, C represents the number of training sample points, L represents the total number of categories, _wk is the weight of the kth category, _wa is the weight of the auxiliary network loss, _wa∈ [0,1], and in this embodiment, _wa ＝0.5;

p _i,k represents the true probability that the i-th sample point belongs to the k-th class. If the i-th sample point belongs to the k-th class, the probability value is 1, otherwise, the probability value is 0;

和

分别表示特征上采样网络和辅助网络预测的第i个样本点属于第k类的概率，计算公式如下：

and

They represent the probability that the i-th sample point predicted by the feature upsampling network and the auxiliary network belongs to the k-th class, respectively. The calculation formula is as follows:

分别表示特征上采样网络和辅助网络输出的第i个样本点的第k个通道特征值，计算公式如下：in,

They represent the k-th channel feature value of the i-th sample point output by the feature upsampling network and the auxiliary network respectively. The calculation formula is as follows:

Among them, _xi represents the input feature of the i-th sample point, ^f1 represents the feature upsampling network, ^θ1 represents the parameters of the feature upsampling network, ^f2 represents the auxiliary network, and ^θ2 represents the parameters of the auxiliary network.

Step 4: Use the training set T to perform P rounds of supervised training on the 3D point cloud semantic segmentation network, where P ≥ 500.

In this embodiment, P=1000, and the training steps are as follows:

调整θ_q，得到用于第q+1轮训练过程的网络模型参数θ_q+1，由此得到第q轮训练过程后的网络模型；4.1) In the qth round of training, set _lq as the learning rate of the qth round of training, set _θq as the parameter of the network model of the qth round of training, and use the formula according to the loss function set in step 3

Adjust θ _q to obtain the network model parameter θ _q+1 for the q+1th round of training process, thereby obtaining the network model after the qth round of training process;

4.2) Every P ₁ rounds, the test set is input into the current network model to obtain the predicted categories of all point cloud data in the test set, P ₁ ≥ 2. In this embodiment, P ₁ = 5;

其中，R表示测试集中所有点云数据的预测类别与其真实类别相同的数目，H表示测试集中所有点云数据的数目；4.3) Count the number of predicted categories of all point cloud data in the test set that are the same as their true categories, and calculate the segmentation accuracy:

Among them, R represents the number of predicted categories of all point cloud data in the test set that are the same as their true categories, and H represents the number of all point cloud data in the test set;

4.4) Compare the segmentation accuracy acc of the current network model with the segmentation accuracy acc of the previously saved network model. If the segmentation accuracy acc of the current network model is higher than the segmentation accuracy acc of the previously saved network model, it indicates that the current network model is better and is saved. Otherwise, it is not saved.

4.5) After P rounds of training are completed, the network model with the highest segmentation accuracy is used as the trained network model;

Step 5: Input the test set V into the trained network model obtained in step 4.5) for semantic segmentation to obtain the segmentation result of each point.

The following is a description of the technical effects of the present invention in conjunction with simulation experiments:

1. Simulation conditions

The simulation experiment of the present invention is carried out in the following environment.

Hardware platform: Intel(R) Xeon(R) CPU E5-2650v4@2.20GHz, 64GB RAM, Ubuntu16.04 operating system, GeForce GTX TITAN X;

Software platform: Tensorflow deep learning framework, Python3.5. The dataset used in the experiment is the point cloud dataset ScanNet.

ScanNet is a point cloud dataset of indoor scenes scanned and reconstructed by RGB-D cameras. It contains a total of 1513 scenes, using 1201 scenes as training sets and 312 scenes as test sets, and contains 21 categories.

2. Simulation experiment:

According to the present invention, a training set and a test set are obtained, a 3D point cloud semantic segmentation network is constructed, the training set is used to perform supervised training on the 3D point cloud semantic segmentation network, and then the trained network model is used to predict the points in the test set, and the segmentation accuracy of the 3D point cloud segmentation network for the test set V is calculated according to the method in step 4.3.

The accuracy of semantic segmentation of point cloud data by the present invention is compared with that of the existing PointNet++ method, and the segmentation accuracy is used as an evaluation index to evaluate the quality of the present invention and the existing technology. The results are shown in Table 1:

Table 1 ScanNet dataset segmentation accuracy comparison table

Evaluation indicators Prior art The present invention Segmentation accuracy 0.836 0.852

As can be seen from Table 1, the segmentation accuracy of the present invention on the ScanNet dataset exceeds that of the prior art Pointnet++ by 1.6%, indicating that the semantic segmentation effect of the present invention on 3D point clouds is stronger than that of PointNet++.

Claims (7)

1. A 3D point cloud semantic segmentation method based on position attention and auxiliary network, characterized by comprising the following: (1) Download the training and test files of 3D point cloud data from the ScanNet official website, perform category statistics and block processing on them, and obtain the training set T and test set V; (2) Construct a 3D point cloud semantic segmentation network, which includes a cascaded feature downsampling network, a position attention module, a feature upsampling network, and an auxiliary network; The position attention module includes three independent 1D convolutional layers Q, U, and V, which are used to extract the features of the input data F of the module and calculate the attention influence value _tij between the features represented by each centroid and the feature E after attention enhancement:

E=[E ₁ ; E ₂ ;...; E _i ;...; E _N ] Among them, U _i represents the feature of the i-th centroid extracted by the 1D convolution layer U from the input data F of the position attention module, Q _j ^T represents the transposition of the feature of the j-th centroid extracted by the 1D convolution layer Q from the input data F of the position attention module, N represents the number of centroids of F, and E _i represents the feature of the i-th centroid in E. The calculation formula is:

Among them, _Vj represents the feature of the jth centroid extracted by F through the 1D convolution layer V,

represents the feature of the i-th centroid after position attention, α represents the weight of the position attention feature, and F _i represents the feature of the i-th centroid of the input; The auxiliary network includes b PointAux modules cascaded in sequence and a 1D convolution layer for classification, each PointAux module includes a 1D convolution layer and a feature interpolation layer, wherein b≥1; (3) Use the multi-classification cross entropy loss function as the loss function of the 3D point cloud semantic segmentation network; (4) Using the training set T, perform P rounds of supervised training on the 3D point cloud data semantic segmentation network, where P ≥ 500: (4a) In each round of training, the network parameters are adjusted according to the loss function of the semantic segmentation network to obtain the network model; (4b) Every P ₁ rounds, use the samples of the test set to evaluate the segmentation accuracy of the current network model. If the segmentation accuracy of the current network model is higher than that of the previously saved network model, it is saved, P ₁ ≥ 2; (4c) After P rounds of training are completed, the network model with the highest segmentation accuracy is taken as the trained network model; (5) Input the test set V into the trained network model for semantic segmentation to obtain the segmentation result of each point. 2. The method according to claim 1 is characterized in that the point cloud data is subjected to category statistics and block processing in (1), which is implemented as follows: (1a) Use the histogram to count the number of each category of point cloud data of all _f0 scenes in the training file, and calculate the weight _wk of each category:

Where G _k represents the number of point cloud data of the kth category, M represents the number of all point cloud data, L represents the number of segmentation categories, f ₀ ≥ 1000, L ≥ 2; (1b) For each scene in the training file, randomly select a point as the center point with coordinates (x, y, z), and take points in the range of (x-0.75, x+0.75), (y-0.75, y+0.75), and (z-0.75, z+0.75) around it to form a data block. The number of points in the data block is compared with the number of sampling points _N0 to determine whether it is reasonable: If the number of points in the data block is greater than the number of sampling points N ₀ , the data block is judged to be reasonable, and N ₀ points are randomly sampled in the data block to form a sample data. Otherwise, the data block is discarded, thereby obtaining the training set T, where N ₀ ≥ 4096; (1c) For each of all f ₁ scenes in the test file, a cubic window of size 1.5×1.5×3 is used for sliding window dicing. For each data block, N ₀ points are randomly sampled to form a sample data to obtain a test set V, f ₁ ≥300. 3. The method according to claim 1 is characterized in that the feature downsampling network described in (2) includes n cascaded PointSA modules, each PointSA module includes a point cloud centroid sampling and grouping layer, and a point cloud feature extraction layer cascaded in sequence, where n≥2. 4. The method according to claim 1 is characterized in that the feature upsampling network described in (2) includes a PointFP modules, a 1D convolution layer, a Dropout layer and a 1D convolution layer for classification that are cascaded in sequence, and each PointFP module includes a feature interpolation layer and a feature extraction layer that are cascaded in sequence, where a≥2. 5. The method according to claim 1, characterized in that the loss function of the 3D point cloud semantic segmentation network in step (3) is calculated as follows:

Where C represents the number of training sample points, L represents the total number of categories, _wk is the weight of the kth category, _wa is the weight of the auxiliary network loss, _wa∈ [0,1]; pi _,k represents the true probability that the ith sample point belongs to the kth category. If the ith sample point belongs to the kth category, the probability value is 1, otherwise, the probability value is 0;

and

They represent the probability that the i-th sample point predicted by the feature upsampling network and the auxiliary network belongs to the k-th class, respectively.

and

The calculation formula is as follows:

in,

They represent the k-th channel feature value of the i-th sample point output by the feature upsampling network and the auxiliary network respectively. The calculation formula is as follows:

Among them, _xi represents the input feature of the i-th sample point, ^f1 represents the feature upsampling network, ^θ1 represents the parameters of the feature upsampling network, ^f2 represents the auxiliary network, and ^θ2 represents the parameters of the auxiliary network. 6. The method according to claim 5, characterized in that in (4a), adjusting the network parameters according to the loss function of the semantic segmentation network is performed by the following formula:

Among them, _lq represents the learning rate of the qth round of training process, _θq represents the parameters of the 3D point cloud semantic segmentation network in the qth round of training process, and _θq+1 represents the parameters used for the q+1th round of training process after adjusting _θq . 7. The method according to claim 1, characterized in that, in (4b), the segmentation accuracy of the current network model is evaluated every P ₁ rounds, which is implemented as follows: (4b1) Every P ₁ rounds, the test set is input into the current network model to obtain the predicted categories of all point cloud data in the test set; (4b2) Count the number of predicted categories of all point cloud data in the test set that are the same as their true categories, and calculate the segmentation accuracy:

Among them, R represents the number of predicted categories of all point cloud data in the test set that are the same as their true categories, and H represents the number of all point cloud data in the test set; (4b3) Compare the segmentation accuracy of the current network model with the segmentation accuracy of the previously saved network model. If the segmentation accuracy of the current network model is higher than that of the previously saved network model, it indicates that the current network model is better and is saved. Otherwise, it is not saved.

Images