3D Point Cloud Classification Method Based on Dynamic Coverage of Local Area

基于局部区域动态覆盖的3D点云分类方法

by Changshuo Wang*, Han Wang, Xin Ning, Weisheng Tian, and Weijun Li.

Introduction

This code is the pytorch 0.4 version of DC-CNN to reproduce 3D Point Cloud Classification Method Based on Dynamic Coverage of Local Area on the ModelNet40.

Requirement

Ubuntu 18.04
Python 3.7.10 (recommend Anaconda3)
Pytorch 0.4.1
CMake 3.10.2
CUDA 10.0 + cuDNN 7
Cudatoolkit V10.0.130

conda install pytorch=0.4.1 cuda92 -c pytorch

Note: If you want to run DC-CNN code in pytorch1.7+ environment, you can run DC-CNN_ScanObjectNN on ScanObjectNN dataset.

Download

git clone https://github.com/changshuowang/DC-CNN.git
cd changshuowang/DC-CNN

Building Kernel

mkdir build && cd build
cmake .. && make

Dataset

Download and unzip ModelNet40 (415M) in data directory.

Usage: Train

python train_cls.py

Note: We have trained a Single-Scale-Neighborhood classification model in log/seed_XXX-XXX/ folder.

Usage: Evaluation

python voting_evaluate_cls.py

Note: You can use your model log/seed_XXX-XXX/*.pth as the checkpoint in cfgs/config_ssn_cls.yaml, and with majority voting you will get an accuracy of 93.6% from the single-scale-neighborhood classification model. The performance on ModelNet40 of almost all methods are not stable, see (CVMI-Lab/PAConv#9 (comment)). If you run the same codes for several times, you will get different results.

Contact

You are welcome to send pull requests or share some ideas with us.

contact email: [email protected].

Acknowledgement

Our code base is partially borrowed from EllipsoidQuery, RS-CNN and Pointnet2_PyTorch.

LICENSE

DC-CNN is under the Apache-2.0 license. Please contact the authors for commercial use.