Home

Awesome

PointCMT

arXiv GitHub Stars visitors

This repository is for PointCMT introduced in the following paper

Xu Yan*, Heshen Zhan*, Chaoda Zheng, Jiantao Gao, Ruimao Zhang, Shuguang Cui, Zhen Li*, "Let Images Give You More: Point Cloud Cross-Modal Training for Shape Analysis", NeurIPS 2022 (Spotlight) :smiley: [arxiv].

image

If you find our work useful in your research, please consider citing:

@InProceedings{yan2022let,
      title={Let Images Give You More: Point Cloud Cross-Modal Training for Shape Analysis}, 
      author={Xu Yan and Heshen Zhan and Chaoda Zheng and Jiantao Gao and Ruimao Zhang and Shuguang Cui and Zhen Li},
      year={2022},
      booktitle={NeurIPS}
}

@inproceedings{yan20222dpass,
  title={2dpass: 2d priors assisted semantic segmentation on lidar point clouds},
  author={Yan, Xu and Gao, Jiantao and Zheng, Chaoda and Zheng, Chao and Zhang, Ruimao and Cui, Shuguang and Li, Zhen},
  booktitle={European Conference on Computer Vision},
  pages={677--695},
  year={2022},
  organization={Springer}
}

Our another work for cross-modal semantic segmentation (ECCV 2022) is released HERE.

Update

2022/12/19: The codes for ModelNet40 are released :rocket:!<br> 2022/11/16: Our paper is selected as spotlight in NeurIPS 2022!

Install

The latest codes are tested on CUDA10.1, PyTorch 1.4.0 and Python 3.7.5. Please use the same environment with us!

conda install pytorch==1.4.0 cudatoolkit=10.1 -c pytorch
pip install -r requirements.txt

Compile the library through:

cd emdloss
python setup.py install
cd ../pointnet2/
pip install -e . && cd ..

Data Preparation

cd data
python modelnet40_processor.py

You can also download processed data directly from HERE.

Training on ModelNet40

Stage I

Train and evaluate MVCNN model through

# training
python main_mvcnn_modelnet.py --exp_name mvcnn_default

# evaluation
python main_mvcnn_modelnet.py --exp_name mvcnn_default --eval --model_path [CHECKPOINT_PATH]

Note that the large batch size for MVCNN is necessary!! If you use smaller batch size due to the memory limitation, you need increase the --accumulation_step during the training. After training, the performance of MVCNN should be around 97%.

Stage II

Training cross-model point generator through

python train_cmpg.py --exp_name cmpg_default --teacher_path [MVCNN_CHECKPOINT]

Stage III

Before conducting cross-modal training, we first generate the features of MVCNN offline to speedup the training:

python offline_feature_generation.py --model_path [MVCNN_CHECKPOINT]

The features will be saved in dataset/ModelNet40/data/. After that, we train PointNet++ with PointCMT through

python train_pointcmt.py --exp_name pointnet2_pointcmt --cmpg_checkpoint [CMPG_CHECKPOINT]

You can also train the vanilla model through --no_pointcmt option.

Inference

We conduct voting test on our trained model:

python voting_test.py --checkpoint [POINITNET2_CHECKPOINT]

Model Zoo

You can also use our pre-trained models from pretrained/modelnet40/, including MVCNN, CMPG and PointNet++. More models and codes for ScanObjectNN will be released very soon!

ModelAccuracy
MVCNN97.0%
PointNet++94.6%

Acknowledgements

Code is built based on PointNet++, SimpleView and EMD(Earth Mover's Distance).

License

This repository is released under MIT License (see LICENSE file for details).