PyTorch implementation of the paper:
Geometric Transformer for Fast and Robust Point Cloud Registration.
Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu.
We study the problem of extracting accurate correspondences for point cloud registration. Recent keypoint-free methods bypass the detection of repeatable keypoints which is difficult in low-overlap scenarios, showing great potential in registration. They seek correspondences over downsampled superpoints, which are then propagated to dense points. Superpoints are matched based on whether their neighboring patches overlap. Such sparse and loose matching requires contextual features capturing the geometric structure of the point clouds. We propose Geometric Transformer to learn geometric feature for robust superpoint matching. It encodes pair-wise distances and triplet-wise angles, making it robust in low-overlap cases and invariant to rigid transformation. The simplistic design attains surprisingly high matching accuracy such that no RANSAC is required in the estimation of alignment transformation, leading to
2022.03.30: Code and pretrained models on KITTI and ModelNet40 release.
2022.03.29: This work is selected for an ORAL presentation at CVPR 2022.
2022.03.02: This work is accepted by CVPR 2022. Code and Models on ModelNet40 and KITTI will be released soon.
2022.02.15: Paper is available at arXiv.
2022.02.14: Code and pretrained model on 3DMatch/3DLoMatch release.
Please use the following command for installation.
# It is recommended to create a new environment
conda create -n geotransformer python==3.8
conda activate geotransformer
# [Optional] If you are using CUDA 11.0 or newer, please install `torch==1.7.1+cu110`
pip install torch==1.7.1+cu110 -f https://download.pytorch.org/whl/torch_stable.html
# Install packages and other dependencies
pip install -r requirements.txt
python setup.py build develop
Code has been tested with Ubuntu 20.04, GCC 9.3.0, Python 3.8, PyTorch 1.7.1, CUDA 11.1 and cuDNN 8.1.0.
The dataset can be downloaded from PREDATOR. The data should be organized as follows:
--data--3DMatch--metadata
|--data--train--7-scenes-chess--cloud_bin_0.pth
| | |--...
| |--...
|--test--7-scenes-redkitchen--cloud_bin_0.pth
| |--...
|--...
The code for 3DMatch is in experiments/geotransformer.3dmatch.stage4.gse.k3.max.oacl.stage2.sinkhorn
. Use the following command for training.
CUDA_VISIBLE_DEVICES=0 python trainval.py
Use the following command for testing.
# 3DMatch
CUDA_VISIBLE_DEVICES=0 ./eval.sh EPOCH 3DMatch
# 3DLoMatch
CUDA_VISIBLE_DEVICES=0 ./eval.sh EPOCH 3DLoMatch
EPOCH
is the epoch id.
We also provide pretrained weights in weights
, use the following command to test the pretrained weights.
CUDA_VISIBLE_DEVICES=0 python test.py --snapshot=../../weights/geotransformer-3dmatch.pth.tar --benchmark=3DMatch
CUDA_VISIBLE_DEVICES=0 python eval.py --benchmark=3DMatch --method=lgr
Replace 3DMatch
with 3DLoMatch
to evaluate on 3DLoMatch.
Download the data from the Kitti official website into data/Kitti
and run data/Kitti/downsample_pcd.py
to generate the data. The data should be organized as follows:
--data--Kitti--metadata
|--sequences--00--velodyne--000000.bin
| | |--...
| |...
|--downsampled--00--000000.npy
| |--...
|--...
The code for Kitti is in experiments/geotransformer.kitti.stage5.gse.k3.max.oacl.stage2.sinkhorn
. Use the following command for training.
CUDA_VISIBLE_DEVICES=0 python trainval.py
Use the following command for testing.
CUDA_VISIBLE_DEVICES=0 ./eval.sh EPOCH
EPOCH
is the epoch id.
We also provide pretrained weights in weights
, use the following command to test the pretrained weights.
CUDA_VISIBLE_DEVICES=0 python test.py --snapshot=../../weights/geotransformer-kitti.pth.tar
CUDA_VISIBLE_DEVICES=0 python eval.py --method=lgr
Download the data and run data/ModelNet/split_data.py
to generate the data. The data should be organized as follows:
--data--ModelNet--modelnet_ply_hdf5_2048--...
|--train.pkl
|--val.pkl
|--test.pkl
The code for ModelNet is in experiments/geotransformer.modelnet.rpmnet.stage4.gse.k3.max.oacl.stage2.sinkhorn
. Use the following command for training.
CUDA_VISIBLE_DEVICES=0 python trainval.py
Use the following command for testing.
CUDA_VISIBLE_DEVICES=0 python test.py --test_iter=ITER
ITER
is the iteration id.
We also provide pretrained weights in weights
, use the following command to test the pretrained weights.
CUDA_VISIBLE_DEVICES=0 python test.py --snapshot=../../weights/geotransformer-modelnet.pth.tar
We evaluate GeoTransformer on the standard 3DMatch/3DLoMatch benchmarks as in PREDATOR.
Benchmark | FMR | IR | RR |
---|---|---|---|
3DMatch | 98.2 | 70.9 | 92.5 |
3DLoMatch | 87.1 | 43.5 | 74.2 |
We evaluate GeoTransformer on the standard Kitti benchmark as in PREDATOR.
Benchmark | RRE | RTE | RR |
---|---|---|---|
Kitti | 0.230 | 6.2 | 99.8 |
We evaluate GeoTransformer on ModelNet with two settings:
- Standard setting: [0, 45] rotation, [-0.5, 0.5] translation, gaussian noise clipped to 0.05.
- Full-range setting: [0, 180] rotation, [-0.5, 0.5] translation, gaussian noise clipped to 0.05.
We remove symmetric classes and use the data augmentation in RPMNet which is more difficult than PRNet.
Benchmark | RRE | RTE | RMSE |
---|---|---|---|
seen (45$^{circ}$) | 1.577 | 0.018 | 0.017 |
seen (180$^{circ}$) | 6.830 | 0.044 | 0.042 |
@misc{qin2022geometric,
title={Geometric Transformer for Fast and Robust Point Cloud Registration},
author={Zheng Qin and Hao Yu and Changjian Wang and Yulan Guo and Yuxing Peng and Kai Xu},
year={2022},
eprint={2202.06688},
archivePrefix={arXiv},
primaryClass={cs.CV}
}