Point Cloud Understanding with UniRepLKNet

November 28, 2023 ยท View on GitHub

Created by Xiaohan Ding, Yiyuan Zhang, etc.

This repository is an official implementation of UniRepLKNet.

This repository is built to explore the ability of RepLK-series networks to understand point cloud. We are mainly focused on the shape classification with ModelNet-40 and ScanObjectNN datasets. Besides fully training, we also explore the advantages of pretrained UniRepLKNet on few-shot learning tasks.

Preparation

Installation Prerequisites

  • Python 3.9
  • CUDA 11.3
  • PyTorch 1.11.1
  • timm 0.5.4
  • torch_scatter
  • pointnet2_ops
  • cv2, sklearn, yaml, h5py
conda create -n pt python=3.9
conda activate pt
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3

mkdir lib
cd lib
git clone https://github.com/erikwijmans/Pointnet2_PyTorch.git
cd Pointnet2_PyTorch
pip install pointnet2_ops_lib/.
cd ../..

pip install torch-scatter -f https://data.pyg.org/whl/torch-1.11.0+cu113.html
pip install timm==0.5.4 opencv-python scikit-learn h5py pyyaml tqdm tensorboardx einops

Data Preparation

  • Download the processed ModelNet40 dataset from [Google Drive][Tsinghua Cloud][BaiDuYun](code:4u1e). Or you can download the offical ModelNet from here, and process it by yourself.

  • Download the official ScanObjectNN dataset from here.

  • The data is expected to be in the following file structure:

    Point/
|-- config/
|-- data/
    |-- ModelNet40/
        |-- modelnet40_shape_names.txt
        |-- modelnet_train.txt
        |-- modelnet_test.txt
        |-- modelnet40_train_8192pts_fps.dat
        |-- modelnet40_test_8192pts_fps.dat
    |-- ScanObjectNN/
        |-- main_split/
            |-- training_objectdataset_augmentedrot_scale75.h5
            |-- test_objectdataset_augmentedrot_scale75.h5
|-- dataset/

(modelnet40_shape_names.txt, modelnet_train.txt, and modelnet_test.txt are provided in PointBERT )

Usage

bash tool/train_unireplknet.sh mv_unireplket-s ModelNet40 config/ModelNet40/multiview_UniRepLKNet-S.yaml

Acknowledgements

Our code is inspired by Meta-Transformer and P2P.