faultaddr / RSCNN

Relation-Shape Convolutional Neural Network for Point Cloud Analysis

This repository contains the author's implementation in Pytorch for the paper:

Relation-Shape Convolutional Neural Network for Point Cloud Analysis [arXiv] [CVF]
Yongcheng Liu, Bin Fan, Shiming Xiang and Chunhong Pan
CVPR 2019 Oral & Best paper finalist     Project Page: https://yochengliu.github.io/Relation-Shape-CNN/

Citation

If Liu's paper is helpful for your research, please consider citing:

        @inproceedings{liu2019rscnn,   
            author = {Yongcheng Liu and    
                            Bin Fan and    
                      Shiming Xiang and   
                           Chunhong Pan},   
            title = {Relation-Shape Convolutional Neural Network for Point Cloud Analysis},   
            booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},    
            pages = {8895--8904},  
            year = {2019}   
        }   

Usage: Preparation

Note

This repository is reproduction of RSCNN based on Pytorch version 1.0 or 1.1, borrowed from Liu's code (https://yochengliu.github.io/Relation-Shape-CNN/)

Requirement

• Ubuntu 16.04
• Python 3.6 (recommend Anaconda3)
• Pytorch 1.0+ (test on 1.0, 1.1 and 1.2)
• CMake > 2.8
• CUDA 10.2 + cuDNN 7.6
• GNU <= 7.5

Setup

The custom ops used by Pointnet++ are currently ONLY supported on the GPU using CUDA.

• Building _ext module (build and copy a .so file to utils/)

  ::

    python setup.py build_ext --inplace
  

Dataset

Shape Classification

Download and unzip ModelNet40 (415M). Replace $data_root$ in cfgs/config_*_cls.yaml with the dataset parent path.

ShapeNet Part Segmentation

Download and unzip ShapeNet Part (674M). Replace $data_root$ in cfgs/config_*_partseg.yaml with the dataset path.

Usage: Training

Shape Classification

sh train_cls.sh

You can modify relation_prior in cfgs/config_*_cls.yaml. We have trained a Single-Scale-Neighborhood classification model in cls folder, whose accuracy is 92.38%.

Shape Part Segmentation

sh train_partseg.sh

We have trained a Multi-Scale-Neighborhood part segmentation model in seg folder, whose class mIoU and instance mIoU is 84.18% and 85.81% respectively.

Usage: Evaluation

Shape Classification

Voting script: voting_evaluate_cls.py

You can use our model cls/model_cls_ssn_iter_16218_acc_0.923825.pth as the checkpoint in config_ssn_cls.yaml, and after this voting you will get an accuracy of 92.71% if all things go right.

Shape Part Segmentation

Voting script: voting_evaluate_partseg.py

You can use our model seg/model_seg_msn_iter_57585_ins_0.858054_cls_0.841787.pth as the checkpoint in config_msn_partseg.yaml.

License

The code is released under MIT License (see LICENSE file for details).

Acknowledgement

The code is heavily borrowed from Pointnet2_PyTorch.

Contact

If you have some ideas or questions about our research to share with us, please contact [e-mail]