Awesome

Learning Geometry-Disentangled Representation for Complementary Understanding of 3D Object Point Cloud.

This repository is built for the paper:

Learning Geometry-Disentangled Representation for Complementary Understanding of 3D Object Point Cloud (AAAI2021) [arXiv] <br> by Mutian Xu*, Junhao Zhang*, Zhipeng Zhou, Mingye Xu, Xiaojuan Qi and Yu Qiao.

Overview

Citation

If you find the code or trained models useful, please consider citing:

@inproceedings{xu2021gdanet,
  title={Learning Geometry-Disentangled Representation for Complementary Understanding of 3D Object Point Cloud}, 
  author={Mutian Xu and Junhao Zhang and Zhipeng Zhou and Mingye Xu and Xiaojuan Qi and Yu Qiao},
  booktitle={AAAI},
  year={2021}
}

Installation

Requirements

• Linux (tested on Ubuntu 14.04/16.04)
• Python 3.5+
• PyTorch 1.0+

Dataset

• Create the folder to symlink the data later:

  mkdir -p data

• Object Classification:

  Download and unzip ModelNet40 (415M), then symlink the path to it as follows (you can alternatively modify the path here) :

  ln -s /path to modelnet40/modelnet40_ply_hdf5_2048 data

• Shape Part Segmentation:

  Download and unzip ShapeNet Part (674M), then symlink the path to it as follows (you can alternatively modify the path here) :

  ln -s /path to shapenet part/shapenetcore_partanno_segmentation_benchmark_v0_normal data

Usage

Object Classification on ModelNet40

• Train:

  python main_cls.py

• Test:

  • Run the voting evaluation script, after this voting you will get an accuracy of 93.8% if all things go right:

    python voting_eval_modelnet.py --model_path 'pretrained/GDANet_ModelNet40_93.4.t7'

  • You can also directly evaluate our pretrained model without voting to get an accuracy of 93.4%:

    python main.py --eval True --model_path 'pretrained/GDANet_ModelNet40_93.4.t7'

Shape Part Segmentation on ShapeNet Part

• Train:

  • Training from scratch:

    python main_ptseg.py

  • If you want resume training from checkpoints, specify resume in the args:

    python main_ptseg.py --resume True

• Test:

  You can choose to test the model with the best instance mIoU, class mIoU or accuracy, by specifying eval and model_type in the args:

  • python main_ptseg.py --eval True --model_type 'insiou' (best instance mIoU, default)

  • python main_ptseg.py --eval True --model_type 'clsiou' (best class mIoU)

  • python main_ptseg.py --eval True --model_type 'acc' (best accuracy)

  Note: This works only after you trained the model or if you have the checkpoint in checkpoints/GDANet. If you run the training from scratch the checkpoints will automatically be generated there.

Performance

The following tables report the current performances on different tasks and datasets.

Object Classification on ModelNet40

Object Classification under Corruptions on OmniObject3D.

Object Classification under Corruptions on ModelNet-C.

Object Classification against Common Corruptions on ModelNet40-C.

Shape Part Segmentation on ShapeNet Part

Other information

Please contact Mutian Xu (mino1018@outlook.com) or Junhao Zhang (junhaozhang98@gmail.com) for further discussion.

Acknowledgement

This code is is partially borrowed from DGCNN and PointNet++.

Update

20/05/2022:

GDANet gains competitive performance on OmniObject3D, ModelNet-C and ModelNet40-C datasets for object classification under corruptions.