Leveraging Inlier Correspondences Proportion for Point Cloud Registration

Paper "Neighborhood-aware Geometric Encoding Network for Point Cloud Registration" was renamed to "Leveraging Inlier Correspondences Proportion for Point Cloud Registration" (NgeNet -> GCNet).

Results (saved in reg_results/3DMatch*-pred)

• Recall on 3DMatch and 3DLoMatch (correspondences RMSE below 0.2)

  Dataset	npairs	Scene Recall (%)	Pair Recall (%)
  3DMatch	1279	92.9	93.9
  3DLoMatch	1726	71.9	74.5

• Recall on 3DMatch and 3DLoMatch (under 0.3m && 15 degrees)

  Dataset	npairs	Pair Recall (%)
  3DMatch	1623	95.0
  3DLoMatch	1781	75.1

• Results on Odometry KITTI

  Dataset	RTE(cm)	RRE(°)	Recall (%)
  Odometry KITTI	6.1	0.26	99.8

• Results on MVP-RG

  Dataset	RRE(°)	RTE	RMSE
  MVP-RG	7.99	0.048	0.093

Environments

• All experiments were run on a RTX 3090 GPU with an Intel 8255C CPU at 2.50GHz CPU. Dependencies can be found in requirements.txt.

• Compile python bindings

  # Compile
  
  cd NgeNet/cpp_wrappers
  sh compile_wrappers.sh
  

[Pretrained weights (Optional)]

Download pretrained weights for 3DMatch, 3DLoMatch, Odometry KITTI and MVP-RG from GoogleDrive or BaiduDisk (pwd: vr9g).

[3DMatch and 3DLoMatch]

1.1 dataset

We adopt the 3DMatch and 3DLoMatch provided from PREDATOR, and download it here [936.1MB]. Unzip it, then we should get the following directories structure:

| -- indoor
    | -- train (#82, cats: #54)
        | -- 7-scenes-chess
        | -- 7-scenes-fire
        | -- ...
        | -- sun3d-mit_w20_athena-sc_athena_oct_29_2012_scan1_erika_4
    | -- test (#8, cats: #8)
        | -- 7-scenes-redkitchen
        | -- sun3d-home_md-home_md_scan9_2012_sep_30
        | -- ...
        | -- sun3d-mit_lab_hj-lab_hj_tea_nov_2_2012_scan1_erika

1.2 train

## Reconfigure configs/threedmatch.yaml by updating the following values based on your dataset.

# exp_dir: your_saved_path for checkpoints and summary.
# root: your_data_path for the 3dMatch dataset.

cd NgeNet
python train.py configs/threedmatch.yaml

# note: The code `torch.cuda.empty_cache()` in `train.py` has some impact on the training speed.
# You can remove it or change its postion according to your GPU memory. 

1.3 evaluate and visualize

cd NgeNet

python eval_3dmatch.py --benchmark 3DMatch --data_root your_path/indoor --checkpoint your_path/3dmatch.pth --saved_path work_dirs/3dmatch [--vis] [--no_cuda]

python eval_3dmatch.py --benchmark 3DLoMatch --data_root your_path/indoor --checkpoint your_path/3dmatch.pth --saved_path work_dirs/3dlomatch [--vis] [--no_cuda]

[Odometry KITTI]

2.1 dataset

Download odometry kitti here with [velodyne laser data, 80 GB] and [ground truth poses (4 MB)], then unzip and organize in the following format.

| -- kitti
    | -- dataset
        | -- poses (#11 txt)
        | -- sequences (#11 / #22)
    | -- icp (generated automatically when training and testing)
        | -- 0_0_11.npy
        | -- ...
        | -- 9_992_1004.npy

2.2 train

## Reconfigure configs/kitti.yaml by updating the following values based on your dataset.

# exp_dir: your_saved_path for checkpoints and summary.
# root: your_data_path for the Odometry KITTI.

cd NgeNet
python train.py configs/kitti.yaml

2.3 evaluate and visualize

cd NgeNet
python eval_kitti.py --data_root your_path/kitti --checkpoint your_path/kitti.pth [--vis] [--no_cuda]

[MVP-RG]

3.1 dataset

Download MVP-RG dataset here, then organize in the following format.

| -- mvp_rg
    | -- MVP_Train_RG.h5
    | -- MVP_Test_RG.h5

3.2 train

## Reconfigure configs/mvp_rg.yaml by updating the following values based on your dataset.

# exp_dir: your_saved_path for checkpoints and summary.
# root: your_data_path for the MVP-RG.

python train.py configs/mvp_rg.yaml

# note: The code `torch.cuda.empty_cache()` in `train.py` has some impact on the training speed.
# You can remove it or change its postion according to your GPU memory. 

3.3 evaluate and visualize

python eval_mvp_rg.py --data_root your_path/mvp_rg --checkpoint your_path/mvp_rg.pth [--vis] [--no_cuda]

[Demo]

4.1 3DMatch

python demo.py --src_path demo_data/cloud_bin_21.pth --tgt_path demo_data/cloud_bin_34.pth --checkpoint your_path/3dmatch.pth --voxel_size 0.025 --npts 5000

4.2 Personal data (with the same voxel size as 3DMatch)

python demo.py --src_path demo_data/src1.ply --tgt_path demo_data/tgt1.ply --checkpoint your_path/3dmatch.pth  --voxel_size 0.025 --npts 20000

4.3 Personal data (with different voxel size from 3DMatch)

python demo.py --src_path demo_data/src2.ply  --tgt_path demo_data/tgt2.ply --checkpoint your_path/3dmatch.pth --voxel_size 3 --npts 20000

Set an appropriate voxel_size for your test data. If you want to test on point cloud pair with large amount of points, please set a large voxel_size according to your data.

Citation

@article{zhu2022leveraging,
  title={Leveraging Inlier Correspondences Proportion for Point Cloud Registration},
  author={Zhu, Lifa and Guan, Haining and Lin, Changwei and Han, Renmin},
  journal={arXiv preprint arXiv:2201.12094},
  year={2022}
}

Acknowledgements

Thanks for the open source code OverlapPredator, KPConv-PyTorch, KPConv.pytorch, FCGF, D3Feat.pytorch, MVP_Benchmark and ROPNet.