A Robust Loss for Point Cloud Registration

This repository is the implementation of our ICCV 2021 paper A Robust Loss for Point Cloud Registration.

Authors: Zhi Deng, Yuxin Yao, Bailin Deng and Juyong Zhang.

Setup

Requirements

Our metric is implemented with the Pytorch, and we test on the Pytorch [0.7, 1.7]. Besides, considering memory consumption, please keep the memory above 15G. You can refer to the requirements.txt for more details.

Data

You can download the Human dataset, Airplane datasets, and Real dataset.

You also can use our scripts to generate pre-processing data and retrain your network, and please refer to the training details in our paper.

Metric

The implemtement details of our metric are in the loss.py. Looking forward to extending our measurement to other frameworks or other areas. - Experiments - Optimization of a single example by embedding the metric into the traditional optimization based on Adam solver. Demo
- Embedding our metric into deep learning and transforming supervised frameworks into unsupervised frameworks, and we implement our experiments with RMP-Net, DCP, and FMR. We also provide the pretrained models. - Cost computation

• Energy visualization of optimizing a single example (refer to the More visualization)

BibTex

@inproceedings{dengzhi2021robust, 
title={A Robust Loss for Point Cloud Registration}, 
author={Zhi Deng, Yuxin Yao, Bailin Deng and Juyong Zhang},
journal={The IEEE International Conference on Computer Vision (ICCV)},
year={2021}}

Notes

If you have comments or questions, please contact Zhi Deng(zhideng@mail.ustc.edu.cn).

Acknowledgement

We would like to thank the authors of DCP_code, RPM-Net_code, FMR_code, FRICP, FGR for making their codes available, and we also thank the source of the data set, Human dataset, M40, Partial Real-datasets, 3D-Match, 7scenes, SLAM.