Enhancing Sampling Protocol for Point Cloud Classification Against Corruptions

This repo contains Pytorch implementation for the paper Enhancing Sampling Protocol for Point Cloud Classification Against Corruptions(IJCAI 2025). This codebase is based on ModelNet40-C, and we thank the authors for their great contributions.

PointSP

Overview

Models:

PointNet, PointNet++, PCT, GDANet, CurveNet

Training Dataset:

ModelNet40

Test Datasets:

• ModelNet40-C (https://github.com/jiachens/ModelNet40-C): introduces 15 corruption types (occlusion, noise, etc.), each with 5 severity levels, applied to ModelNet40.
• PointCloud-C (https://github.com/ldkong1205/PointCloud-C): features 7 real-world inspired corruptions (each with 5 levels) on the same data.
• OmniObject-C (https://github.com/omniobject3d/OmniObject3D): applies PointCloud-C's 7 corruption types (5 levels each) to OmniObject3D classes matching ModelNet40 categories.

Getting Started

First clone the repository. We would refer to the directory containing the code as PointSP.

git clone --recurse-submodules [email protected]:tangsankou/PointSP.git

Requirements

progressbar2==3.20.0
tensorboardX==2.0
torch==2.0.1 
torchvision==0.15.2 
torchaudio==2.0.2
open3d
Ninja
python==3.8
CUDA >=11.7, 
CuDNN 
GCC

We recommend using these versions especially for installing pointnet++ custom CUDA modules.

Install

conda create --name pointsp python=3.8
conda activate pointsp
pip install -r requirements.txt

For PointNet++, we need to install custom CUDA modules. Make sure you have access to a GPU during this step. You might need to set the appropriate TORCH_CUDA_ARCH_LIST environment variable depending on your GPU model. The following command should work for most cases export TORCH_CUDA_ARCH_LIST="6.0;6.1;6.2;7.0;7.5". However, if the install fails, check if TORCH_CUDA_ARCH_LIST is correctly set.

Third-party modules pointnet2_pyt, PCT_Pytorch, emd, and PyGeM can be installed by the following script.

./setup.sh

Usage

This repository provides code for training and evaluating point cloud models (PointNet, PointNet++, PCT, GDANet, CurveNet) on ModelNet40 and testing their robustness against corruptions in ModelNet40-C, PointCloud-C, and OmniObject-C datasets.To train or test any model, we use the main.py script. The format for running this script is as follows.

Training

Train a model with upsampling and weighted sampling:

CUDA_VISIBLE_DEVICES=0 python main.py \
  --exp-config configs/<cfg_name>.yaml \
  --use_upsample lgp_or_lgd \
  --sample_type wrs  

• The --use_upsample parameter is designed for the full-point sampling module in PointSP, while the --sample_type parameter controls the key-point sampling module in PointSP.
• <cfg_name>: Config file (e.g., dgcnn_pointnet2_run_1.yaml).
• lgp_or_lgb: Choose lgp (Local Geometric Propagation)_or_lgd (Local Global Downsampling).
• wrs: Weighted Random Sampling.

Testing on Corrupted Datasets

CUDA_VISIBLE_DEVICES=0 python main.py \
  --entry mnc \(pcc\ominc)
  --model-path <cor_exp/runs>/<cfg_name>/<model_name>.pth \
  --exp-config configs/<cfg_name>.yaml \
  --use_upsample clgp \
  --sample_type ffps \
  --severity 1 \
  --corruption occlusion  

• clgp: Conditional Local Geometric Preserved.
• ffps: Filtered Furthest Point Sampling.
• severity: Corruption level (1–5 for ModelNet40-C and 0-4 for PointCloud-C、 OmniObject-C).
• corruption: Type (occlusion etc.).

Citation

Please cite our paper and ModelNet40-C if you find our work useful in your research. Thank you!

@article{Li2025enhancing,
      title={Enhancing Sampling Protocol for Point Cloud Classification Against Corruptions}, 
      author={Chongshou Li, Pin Tang, Xinke Li, Yuheng Liu, Tianrui Li},
      journal={https://arxiv.org/abs/2408.12062},
      year={2025}
}