Pointnet / pointnet++ point cloud data set processing and training

One 、 Representation of 3D data
The expression forms of three-dimensional data are generally divided into four ：

① Point cloud ： from N individual D Point composition of dimension , When this D = 3 It usually means ( x , y , z ) Coordinates of , Of course, it can also include some normal vectors 、 Strength and other characteristics . This is the main data type today .

② Mesh： It consists of triangular patches and square patches .

③ Voxel ： Use the three-dimensional grid to use 0 and 1 characterization .

④ Multi angle RGB Image or RGB-D Images

And because the point cloud is closer to the original representation of the device （ That is, radar scans objects to directly generate point clouds ） At the same time, its expression is simpler , An object uses only one N × D The matrix representation of , Therefore, point cloud has become the most important of many 3D data representation methods . Mapping 、 Architecture 、 Electric power 、 Industry and even the most popular fields such as automatic driving are widely used .

Two 、 Point cloud segmentation data set processing

Our point cloud data may be inconsistent with the data required by the model , So you need to write your own script to standardize the data . The standard point cloud data processing format is as follows ：

It only contains point cloud （x,y,z） Coordinates and labels corresponding to each point . And maybe we start from CloudCompare There are four columns of data and one column of point cloud intensity information ：

So write the following script to remove the redundant fourth column data （ The specific path of the file is set by yourself ）：

# -*- coding:utf-8 -*-
import os
filePath = 'D:\\ Point cloud data processing team \\pointnet-my\\data\\shapenetcore_partanno_segmentation_benchmark_v0_normal\\03797390'
for i,j,k in os.walk(filePath):
    for name in  k:
        print(name)
        f = open(filePath+name)  #  open txt file 
        line = f.readline()  #  Read the file as a line 
        list1 = []
        while line:
            a = line.split()
            b = a[0:3]
            c = float(a[-1])
            print(c)
            if(float(a[-1])==36.0):
                c=2
            if(float(a[-1])==37.0):
                c=3
            b.append(c)
            list1.append(b)  #  Add it to the list 
            line = f.readline()
        f.close()
        print(list1)
        with open(filePath+name, 'w+') as file:
            for i in list1:
                file.write(str(i[0]))
                file.write(' '+str(i[1]))
                file.write(' ' + str(i[2]))
                file.write(' ' + str(i[3]))
                if(i!=list[-1]):
                    file.write('\n')
        file.close()
# path_out = 'test.txt' #  new txt file 
# with open(path_out, 'w+') as f_out:
# for i in list1:
# fir = '9443_' + i[0] #  Prefix the first column '9443_'
# sec = 9443 + int(i[1]) #  Add... To the values in the second column 9443
#         # print(fir)
#         # print(str(sec))
# f_out.write(fir + ' ' + str(sec) + '\n') #  Write the first two columns into the new txt file 

Then we need to divide the training set 、 Test set 、 Verification set of json Document processing , Because we use our own data set , every last txt The file name of must be different from that in the previous official data set , So you need to write a script to control the training set 、 Test set 、 Verify the three... Read in by the set json File modification . The specific code is as follows （ The path name also needs to be modified to its own path name ）：

import os
filePath = 'D:\\ Point cloud data processing team \\pointnet-my\\data\\shapenetcore_partanno_segmentation_benchmark_v0_normal\\03797390'
##### On the last line, there will be a ,  Report errors ！！！！！！！！！
###### Manually delete or improve programs 
#####
file = '1.txt'
with open(file,'a') as f:
    f.write("[")
    for i,j,k in os.walk(filePath):
         for name in k:
              base_name=os.path.splitext(name)[0]  # Remove the suffix  .txt
              f.write(" \"")
              f.write(os.path.join("shape_data/03797390/",base_name))
              f.write("\"")
              f.write(",")
    f.write("]")
f.close()

Finally, when the actual program is running, it is found that it will contain 0 Too much data leads to inaccurate model classification , as follows ：

According to the actual physical meaning of the specific project （ All the people here 0 That is, the position not detected by the laser ,x,y,z The coordinates are marked 0）, So writing scripts is right for all 0 Remove the part of ：

# -*- coding:utf-8 -*-
import os
filePath = 'D:\\ Point cloud data processing team \\pointnet-my\\data\\shapenetcore_partanno_segmentation_benchmark_v0_normal\\03797390'
for i,j,k in os.walk(filePath):
    for name in  k:
        list1 = []
        for line in open(filePath+name):
            a = line.split()
            #print(a)
            b = a[0:6]
            #print(b)
            a1 =float(a[0])
            a2 =float(a[1])
            a3 =float(a[2])
            #print(a1)
            if(a1==0 and a2==0 and a3==0):
                continue
            list1.append(b[0:6])
        with open(filePath+name, 'w+') as file:
            for i in list1:
                file.write(str(i[0]))
                file.write(' '+str(i[1]))
                file.write(' ' + str(i[2]))
                file.write(' ' + str(i[3]))
                file.write(' ' + str(i[4]))
                if(i!=list[-1]):
                    file.write('\n')
        file.close()

3、 ... and 、 model training
We open model Folder , Choose to pointnet2_part_seg_msg In the name of python file , Double click in ：

Next, we need to modify PointNetFeaturePropagation Of in_channel, The specific method is 128+4 Total number of classes divided , For example, our segmentation here uses two objects , Each object is 2 classification , So the number of channels is 128+42=136：

Next, we need to modify the part of the independent heat coding , take view The second parameter in the method is modified to the number of types of objects to be segmented ：

2.2.5 Modification of training model
In training the model , We need to modify the objects to be classified seg_classes For the category corresponding to our own dataset , For example, the classification in the example includes two categories , Each category has two components .（ For convenience , We keep the original ’Airplane’: [0, 1], ‘Mug’: [2, 3] These two names , At the same time, we should pay attention to it , The number should be from 0 Start numbering in sequence , If not, an error will be reported ）, At the same time, you should pay attention to whether your computer supports it cuda Speed up , without GPU, You can put all the following code ".cuda（）" delete , The program can run correctly .

Next, we need to modify the total number of segmented objects num_classes Number your categories , The total number of parts is your total number ：num_part.

Last , Right click to run the program , If a progress bar appears, it indicates that the model has been successfully improved , You can run through your own data set ！

Four 、 Print model prediction data
After training the model with our own data set , Will use the name test_partseg Of python File to test the model , In order to better understand the model and test the effect of the model , Sometimes we need to print out the point cloud coordinates predicted by the model . Here is given test_partseg File modification method and operation results .

4.1 Test model code modification
First we need to find the python file , Then make the same changes as above, as shown in the following two places ：

Then we need to write code to print out the coordinates with different labels after classification , In this example, we want to print the coordinate points of four categories of two objects , The code is as follows （ Part of the code ）：

                for j in aaa:
                    #print(points1[i,j])
 
                    res1= open(r'E:\03797390_0_'+str(i)+'.txt', 'a')
                    res1.write('\n' + str(points1[i,j]).strip('[]'))
                    res1.close()
                    xxxxxx=xxxxxx+1
 
                bbb = numpy.argwhere(cur_pred_val[i] == 1)
                for j in bbb:
                    #print(points1[i, j])
 
                    res2 = open(r'E:\03797390_1_' + str(i) + '.txt', 'a')
                    res2.write('\n' + str(points1[i,j]).strip('[]'))
                    res2.close()
                    xxxxxx = xxxxxx + 1
 
                ccc = numpy.argwhere(cur_pred_val[i] == 2)
                for j in ccc:
                    #print(points1[i, j])
 
                    res3 = open(r'E:\02691156_2_' + str(i) + '.txt', 'a')
                    res3.write('\n' + str(points1[i,j]).strip('[]'))
                    res3.close()
                    xxxxxx = xxxxxx + 1
 
                ddd = numpy.argwhere(cur_pred_val[i] == 3)
                for j in ddd:
                    #print(points1[i, j])
 
                    res4 = open(r'E:\02691156_3_' + str(i) + '.txt', 'a')
                    res4.write('\n' + str(points1[i,j]).strip('[]'))
                    res4.close()
                    xxxxxx = xxxxxx + 1

After the modification, run the program and find that the results can be displayed , Indicates that the modification was successful ！

The format of the printed point cloud coordinates is as follows ：

reference ：https://blog.csdn.net/weixin_44603934/article/details/123589948