PyTorch implementation of the paper:
Adaptive Multi-step Refinement Network for Robust Point Cloud Registration
Point Cloud Registration (PCR) estimates the relative rigid transformation between two point clouds of the same scene. Despite significant progress with learning-based approaches, existing methods still face challenges when the overlapping region between the two point clouds is small. In this paper, we propose an adaptive multi-step refinement network that refines the registration quality at each step by leveraging the information from the preceding step. To achieve this, we introduce a training procedure and a refinement network. Firstly, to adapt the network to the current step, we utilize a generalized one-way attention mechanism, which prioritizes the last step's estimated overlapping region, and we condition the network on step indices. Secondly, instead of training the network to map either random transformations or a fixed pre-trained model's estimations to the ground truth, we train it on transformations with varying registration qualities, ranging from accurate to inaccurate, thereby enhancing the network's adaptiveness and robustness. Despite its conceptual simplicity, our method achieves state-of-the-art performance on both the 3DMatch/3DLoMatch and KITTI benchmarks. Notably, on 3DLoMatch, our method reaches 80.4% recall rate, with an absolute improvement of 1.2%. Our code will be made public upon publication.
Follow GeoTransformer to install the dependencies.
pip install -r requirements.txt
python setup.py build develop
3D prior data Unzip the file, you can get two folders, one is for training, the other is for testing/validating. The dataset is downloaded from PEAL.
Point cloud data The dataset is downloaded from PREDATOR.
Training prior supplements: Some of the prior information provided by PEAL for training is missing. In order to ensure the training, please extract prior_Geo_train.zip from the project.
Our method creates the training data with different levels of noise for training different models. The parameter --train_preset_T means the levels of noise, ranging from 999 to 0, and bigger train_preset_T refers to more noisy data. For example, if you want to train 5 models, then the five training commands can be set as below:
python ./experiments/geotransformer.3dmatch.stage4.gse.k3.max.oacl.stage2.sinkhorn/trainval.py --fix_train_T --train_preset_T=999 --Slerp
python ./experiments/geotransformer.3dmatch.stage4.gse.k3.max.oacl.stage2.sinkhorn/trainval.py --fix_train_T --train_preset_T=800 --Slerp
python ./experiments/geotransformer.3dmatch.stage4.gse.k3.max.oacl.stage2.sinkhorn/trainval.py --fix_train_T --train_preset_T=600 --Slerp
python ./experiments/geotransformer.3dmatch.stage4.gse.k3.max.oacl.stage2.sinkhorn/trainval.py --fix_train_T --train_preset_T=400 --Slerp
python ./experiments/geotransformer.3dmatch.stage4.gse.k3.max.oacl.stage2.sinkhorn/trainval.py --fix_train_T --train_preset_T=200 --SlerpUse the following command for testing.
bash eval_multi.sh 0 3DMatchor
bash eval_multi.sh 0 3DLoMatch