zhulf0804 / pcreg.pytorch Goto Github PK

Our recent registraion works:

• NgeNet: paper, code. We achieved SoTA RR (Registration Recall) in 3DMatch with 92.9%.
• ROPNet paper, code. Our solution based on ROPNet and OverlapPredator won the second place on the MVP Registration Challenge (ICCV Workshop 2021). [Technical Report]

A Simple Point Cloud Registration Pipeline based on Deep Learning. Detailed Information Please Visit this Zhihu Blog.

• requirements.txt pip install -r requirements.txt
• open3d-python==0.9.0.0 python -m pip install open3d==0.9
• emd loss cd loss/cuda/emd_torch & python setup.py install

• Download data from [here, 435M]

• evaluate and show(download the pretrained checkpoint [Complete, pwd: c4z7, 16.09 M] or [Paritial, pwd: pcno, 16.09] first)

  # Iterative Benchmark
  python modelnet40_evaluate.py --root your_data_path/modelnet40_ply_hdf5_2048 --checkpoint your_ckpt_path/test_min_loss.pth --cuda
  
  # Visualization
  # python modelnet40_evaluate.py --root your_data_path/modelnet40_ply_hdf5_2048 --checkpoint your_ckpt_path/test_min_loss.pth  --show
  
  # ICP
  # python modelnet40_evaluate.py --root your_data_path/modelnet40_ply_hdf5_2048 --method icp
  
  # FGR
  # python modelnet40_evaluate.py --root your_data_path/modelnet40_ply_hdf5_2048 --method fgr --normal
  
  

• train

  CUDA_VISIBLE_DEVICES=0 python modelnet40_train.py --root your_data_path/modelnet40_ply_hdf5_2048
  

• Point-to-Point Correspondences(R error is large due to EMDLoss, see here)

• Noise Data(infer_npts = 1024)

• Partial-to-Complete Registration(infer_npts = 1024)

Note:

• Detailed metrics information please refer to RPM-Net[CVPR 2020].

• Prepare the data in the following structure

  |- CustomData(dir)
      |- train_data(dir)
          - train1.pcd
          - train2.pcd
          - ...
      |- val_data(dir)
          - val1.pcd
          - val2.pcd
          - ...
  

• Train

  python custom_train.py --root your_datapath/CustomData --train_npts 2048 
  # Note: train_npts depends on your dataset
  

• Evaluate

  # Evaluate, infer_npts depends on your dataset
  python custom_evaluate.py --root your_datapath/CustomData --infer_npts 2048 --checkpoint work_dirs/models/checkpoints/test_min_loss.pth --cuda
  
  # Visualize, infer_npts depends on your dataset
  python custom_evaluate.py --root your_datapath/CustomData --infer_npts 2048 --checkpoint work_dirs/models/checkpoints/test_min_loss.pth --show
  

Thanks for the open source code for helping me to train the Point Cloud Registration Network successfully.