REGTR: End-to-end Point Cloud Correspondences with Transformers
This repository contains the source code for REGTR. REGTR utilizes multiple transformer attention layers to directly predict each downsampled point's corresponding location in the other point cloud. Unlike typical correspondence-based registration algorithms, the predicted correspondences are clean and do not require an additional RANSAC step. This results in a fast, yet accurate registration.
If you find this useful, please cite:
@inproceedings{yew2022regtr,
title={REGTR: End-to-end Point Cloud Correspondences with Transformers},
author={Yew, Zi Jian and Lee, Gim hee},
booktitle={CVPR},
year={2022},
}
Dataset environment
Our model is trained with the following environment:
- Python 3.8.8
- PyTorch 1.9.1 with torchvision 0.10.1 (Cuda 11.1)
- PyTorch3D 0.6.0
- MinkowskiEngine 0.5.4
Other required packages can be installed using pip: pip install -r src/requirements.txt.
Data and Preparation
Follow the following instructions to download each dataset (as necessary). Your folder should then look like this:
.
├── data/
├── indoor/
├── test/
| ├── 7-scenes-redkitchen/
| | ├── cloud_bin_0.info.txt
| | ├── cloud_bin_0.pth
| | ├── ...
| ├── ...
├── train/
| ├── 7-scenes-chess/
| | ├── cloud_bin_0.info.txt
| | ├── cloud_bin_0.pth
| | ├── ...
├── test_3DLoMatch_pairs-overlapmask.h5
├── test_3DMatch_pairs-overlapmask.h5
├── train_pairs-overlapmask.h5
└── val_pairs-overlapmask.h5
└── modelnet40_ply_hdf5_2048
├── ply_data_test0.h5
├── ply_data_test1.h5
├── ...
├── src/
└── Readme.md
3DMatch
Download the processed dataset from Predator project site, and place them into ../data.
Then for efficiency, it is recommended to pre-compute the overlapping points (used for computing the overlap loss). You can do this by running the following from the src/ directory:
python data_processing/compute_overlap_3dmatch.py
ModelNet
Download the PointNet-processed dataset from here, and place it into ../data.
Pretrained models
You can download our trained models here. Unzip the files into the trained_models/.
Demo
We provide a simple demo script demo.py that loads our model and checkpoints, registers 2 point clouds, and visualizes the result. Simply download the pretrained models and run the following from the src/ directory:
python demo.py --example 0 # choose from 0 - 4 (see code for details)
Press 'q' to end the visualization and exit. Refer the documentation for visualize_result() for explanation of the visualization.
Inference/Evaluation
The following code in the src/ directory performs evaluation using the pretrained checkpoints as provided above; change the checkpoint paths accordingly if you're using your own trained models. Note that due to non-determinism from the neighborhood computation during our GPU-based KPConv processing, the results will differ slightly (e.g. mean registration recall may differ by around +/- 0.2%) between each run.
3DMatch / 3DLoMatch
This will run inference and compute the evaluation metrics used in Predator (registration success of <20cm).
# 3DMatch
python test.py --dev --resume ../trained_models/3dmatch/ckpt/model-best.pth --benchmark 3DMatch
# 3DLoMatch
python test.py --dev --resume ../trained_models/3dmatch/ckpt/model-best.pth --benchmark 3DLoMatch
ModelNet
# ModelNet
python test.py --dev --resume ../trained_models/modelnet/ckpt/model-best.pth --benchmark ModelNet
# ModelLoNet
python test.py --dev --resume ../trained_models/modelnet/ckpt/model-best.pth --benchmark ModelNet
Training
Run the following commands from the src/ directory to train the network.
3DMatch (Takes ~2.5 days on a Titan RTX)
python train.py --config conf/3dmatch.yaml
ModelNet (Takes <2 days on a Titan RTX)
python train.py --config conf/modelnet.yaml
Acknowledgements
We would like to thank the authors for Predator, D3Feat, KPConv, DETR for making their source codes public.
176 Dec 30, 2022
64 Dec 20, 2022
45 Dec 13, 2022
45 Dec 15, 2022
113 Nov 27, 2022
564 Jan 02, 2023
7 Sep 26, 2022
5 Jan 03, 2023
6 Dec 27, 2022
28 Dec 09, 2022
2 Mar 24, 2022
3 Jun 26, 2022
1 Feb 13, 2022
94 Dec 25, 2022
1 Jan 20, 2022
2 Aug 08, 2022
2.6k Jan 06, 2023
58 Jan 02, 2023
312 Jan 08, 2023
115 Dec 21, 2022