DuMLP-Pin

This repo is the official implementation of "DuMLP-Pin: A Dual-MLP-dot-product Permutation-invariant Network for Set Feature Extraction" (AAAI 2022). It includes codes and models for the following tasks:

Task	Dataset	Evaluation Metric	Value (%)
Point Cloud Classification	ModelNet40	Overall Accuracy	92.26
Point Cloud Part Segmentation	ShapeNetPart	Mean IoU	S: 83.43 L: 84.92

Getting Started

All experiments are done with one RTX 3090.

Install

• Clone this repo.

• Create a conda virtual environment and activate it:

conda create -n dmpp python=3.8
conda activate dmpp

• Install CUDA==11.1 with cudnn8 following the official installation instructions
• Install pytorch:

conda install pytorch==1.8.1 cudatoolkit=11.1

• Install h5py, tqdm and tensorboard:

conda install h5py tqdm==4.59.0 tensorboard==2.4.0

Quick Start

We provide several examples in my_script.sh, which includes some commands to reproduce our results in the paper. Please fill in the value of MODEL_LOAD_PATH for your usage.

The template is:

./my_script.sh [MODE] [TASK] [GPU_INDEX] [S=small, L=large, optional for ShapeNetPart]

Some examples:

chmod +x my_script.sh
./my_script.sh train ModelNet40 0
./my_script.sh train ShapeNetPart 0 S
./my_script.sh eval ShapeNetPart -1 L # -1 for CPU

Structure

Models and datasets are grouped as follows. For ShapeNetPart, we have two models of different sizes: small in S and large in L. We do each single experiment twice to validate the reproducibility.

 DuMLP_Pin
    ├─ModelNet40
    │  ├─datasets
    │  └─models
    │      ├─2021-10-09-18-58-28
    │      └─2021-10-09-20-57-57
    └─ShapeNetPart
        ├─datasets
        └─models
            ├─L
            │  ├─2021-10-17-11-29-54
            │  └─2021-10-18-15-33-17
            └─S
                ├─2021-10-18-15-34-37
                └─2021-10-19-04-49-45

Citation

@article{Fei_Zhu_Liu_Deng_Li_Deng_Zhang_2022,
  title={DuMLP-Pin: A Dual-MLP-Dot-Product Permutation-Invariant Network for Set Feature Extraction},
  volume={36},
  url={https://ojs.aaai.org/index.php/AAAI/article/view/19939},
  DOI={10.1609/aaai.v36i1.19939},
  number={1},
  journal={Proceedings of the AAAI Conference on Artificial Intelligence},
  author={Fei, Jiajun and Zhu, Ziyu and Liu, Wenlei and Deng, Zhidong and Li, Mingyang and Deng, Huanjun and Zhang, Shuo},
  year={2022},
  month={Jun.},
  pages={598-606}
}

Contact

Feel free to contact Jiajun Fei feijj20@mails.tsinghua.edu.cn if you have some questions.