PointNet-master复现

1 准备工作

先激活虚拟环境，再安装h5py

# 安装h5py
sudo apt-get install libhdf5-dev
sudo pip install h5py

2 Classification分类

2.1 训练

• 下载ModelNet40和HDF5文件放置到pointnet-master/data文件夹下，进入data文件夹后再解压

# 解压modelnet40_ply_hdf5_2048.zip
unzip modelnet40_ply_hdf5_2048.zip

• 再退出data文件夹，执行python train.py
• 运行中结果

**** EPOCH 000 ****
----0-----
mean loss: 3.827456
accuracy: 0.192383
----1-----
mean loss: 2.796352
accuracy: 0.307598
----2-----
mean loss: 2.345549
accuracy: 0.385742
----3-----
mean loss: 2.171231
accuracy: 0.424316
----4-----
mean loss: 1.869357
accuracy: 0.485352
----0-----
----1-----
eval mean loss: 1.760690
eval accuracy: 0.489854
eval avg class acc: 0.418257
Model saved in file: log/model.ckpt
**** EPOCH 001 ****
----0-----
mean loss: 1.635424
accuracy: 0.538086
----1-----
mean loss: 1.596464
accuracy: 0.542892
----2-----
mean loss: 1.565687
accuracy: 0.545410
----3-----
mean loss: 1.608424
accuracy: 0.558105
----4-----
mean loss: 1.400728
accuracy: 0.592285
----0-----
----1-----
eval mean loss: 1.227646
eval accuracy: 0.619724
eval avg class acc: 0.530451
**** EPOCH 002 ****

• 使用TensorBoard，查看网络架构和其他的训练参数的变化曲线（最好是从服务器上下载到自己的电脑里查看）

#查看网络架构和其他的训练参数的变化曲线
tensorboard --logdir log

2.2 评估（我训练出来的结果85.6/88.6）

# 评估
python evaluate.py --visu

在上述培训之后，我们可以评估模型并输出错误情况的一些可视化,错误分类的点云将dump默认保存到文件夹中。

3 Part Segmentation物件分割

• 首先现在 ShapeNetPart数据集和HDF5文件

cd part_seg # 进入part_seg文件夹
sh download_data.sh # 下载ShapeNetPart数据集和HDF5文件

一般最好是自己下载好文件再上传到服务器，服务器上的下载速度是龟速。

python train.py
python test.py

#跑train.py的过程 
>>> Training for the epoch 0/200 ...
Loading train file /workshop/user_data/pointnet/part_seg/./hdf5_data/ply_data_train4.h5
	Training Total Mean_loss: 1.974480
		Training Label Mean_loss: 3.489420
		Training Label Accuracy: 0.046875
		Training Seg Mean_loss: 1.398422
		Training Seg Accuracy: 0.668733
Loading train file /workshop/user_data/pointnet/part_seg/./hdf5_data/ply_data_train5.h5
	Training Total Mean_loss: 0.896463
		Training Label Mean_loss: 3.628423
		Training Label Accuracy: 0.022246
		Training Seg Mean_loss: 0.728734
		Training Seg Accuracy: 0.795286
Loading train file /workshop/user_data/pointnet/part_seg/./hdf5_data/ply_data_train1.h5
	Training Total Mean_loss: 0.796229
		Training Label Mean_loss: 3.660106
		Training Label Accuracy: 0.029785
		Training Seg Mean_loss: 0.654158
		Training Seg Accuracy: 0.809802
Loading train file /workshop/user_data/pointnet/part_seg/./hdf5_data/ply_data_train2.h5
	Training Total Mean_loss: 0.773133
		Training Label Mean_loss: 3.381332
		Training Label Accuracy: 0.041016
		Training Seg Mean_loss: 0.633130
		Training Seg Accuracy: 0.811064
Loading train file /workshop/user_data/pointnet/part_seg/./hdf5_data/ply_data_train3.h5
	Training Total Mean_loss: 0.631645
		Training Label Mean_loss: 3.542044
		Training Label Accuracy: 0.023926
		Training Seg Mean_loss: 0.561357
		Training Seg Accuracy: 0.831671
Loading train file /workshop/user_data/pointnet/part_seg/./hdf5_data/ply_data_train0.h5
	Training Total Mean_loss: 261.794937
		Training Label Mean_loss: 3.207757
		Training Label Accuracy: 0.069336
		Training Seg Mean_loss: 0.966167
		Training Seg Accuracy: 0.712125

<<< Testing on the test dataset ...
Loading test file /workshop/user_data/pointnet/part_seg/./hdf5_data/ply_data_val0.h5
	Testing Total Mean_loss: 2509.134123
		Testing Label Mean_loss: 3.620634
		Testing Label Accuracy: 0.094289
		Testing Seg Mean_loss: 1.736152
		Testing Seg Accuracy: 0.562174

		Category Airplane Object Number: 386
		Category Airplane Label Accuracy: 0.000000
		Category Airplane Seg Accuracy: 0.650140

		Category Bag Object Number: 8
		Category Bag Label Accuracy: 0.000000
		Category Bag Seg Accuracy: 0.000000

		Category Cap Object Number: 5
		Category Cap Label Accuracy: 0.200000
		Category Cap Seg Accuracy: 0.000000

		Category Car Object Number: 79
		Category Car Label Accuracy: 0.000000
		Category Car Seg Accuracy: 0.726321

		Category Chair Object Number: 394
		Category Chair Label Accuracy: 0.307107
		Category Chair Seg Accuracy: 0.593638

		Category Earphone Object Number: 6
		Category Earphone Label Accuracy: 0.000000
		Category Earphone Seg Accuracy: 0.000000

		Category Guitar Object Number: 78
		Category Guitar Label Accuracy: 0.102564
		Category Guitar Seg Accuracy: 0.000000

		Category Knife Object Number: 35
		Category Knife Label Accuracy: 0.000000
		Category Knife Seg Accuracy: 0.000000

		Category Lamp Object Number: 142
		Category Lamp Label Accuracy: 0.316901
		Category Lamp Seg Accuracy: 0.626561

		Category Laptop Object Number: 44
		Category Laptop Label Accuracy: 0.000000
		Category Laptop Seg Accuracy: 0.023748

		Category Motorbike Object Number: 26
		Category Motorbike Label Accuracy: 0.000000
		Category Motorbike Seg Accuracy: 0.000000

		Category Mug Object Number: 16
		Category Mug Label Accuracy: 0.000000
		Category Mug Seg Accuracy: 0.000000

		Category Pistol Object Number: 30
		Category Pistol Label Accuracy: 0.000000
		Category Pistol Seg Accuracy: 0.436865

		Category Rocket Object Number: 8
		Category Rocket Label Accuracy: 0.000000
		Category Rocket Seg Accuracy: 0.000000

		Category Skateboard Object Number: 15
		Category Skateboard Label Accuracy: 0.000000
		Category Skateboard Seg Accuracy: 0.590267

		Category Table Object Number: 584
		Category Table Label Accuracy: 0.000000
		Category Table Seg Accuracy: 0.666422

#跑test.py的过程
0/2874 ...
100/2874 ...
200/2874 ...
300/2874 ...
400/2874 ...
500/2874 ...
600/2874 ...
700/2874 ...
800/2874 ...
900/2874 ...
1000/2874 ...
1100/2874 ...
1200/2874 ...
1300/2874 ...
1400/2874 ...
1500/2874 ...
1600/2874 ...
1700/2874 ...
1800/2874 ...
1900/2874 ...
2000/2874 ...
2100/2874 ...
2200/2874 ...
2300/2874 ...
2400/2874 ...
2500/2874 ...
2600/2874 ...
2700/2874 ...
2800/2874 ...
Accuracy: 0.931741
IoU: 0.828935

• 跑完test.py后可以通过CloudCompare或者MeshLab来显示test_results文件夹中的点云(但是不知道如何才能调出不同的部件不同的颜色)
  

4 Semantic Segmentation语义分割

暂时还没跑

参考的博客
[1]: https://blog.csdn.net/qq_40234695/article/details/86223577?depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-1&utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-1

本文地址：https://blog.csdn.net/weixin_44581536/article/details/107638165