FANzhaoxin666 / mprm

Multi-Path Region Mining For Weakly Supervised 3D Semantic Segmentation on Point Clouds

Multi-Path Region Mining For Weakly Supervised 3D Semantic Segmentation on Point Clouds

This repository contains the implementation of our CVPR2020 paper Multi Path Region Mining For Weakly Supervised 3D Semantic Segmentation on Point Clouds(MPRM)(paper)

Installation

This project is developed based on Kernel Point Convolution KPConv (repo). You can follow the original KPConv installation guide for ubuntu 16.04 and 18.04.

If you want to use dense-crf for post-processing, install pydense-crf following this repo

Training

1. Download the Scannet dataset through the official webcite.

2. Modify the dataset path in datasets/Scannet_subcloud.py And start training:

python training_mprm.py

You can also specify the model saving path and change parameters in training_mprm.py.

3. You can plot the training details by speficy the saved model path and run:

python plot_convergence_mprm.py

4. Generate the pseudo label by running:

python generate_pseudo_label.py

5. You can choose to post-process the pseudo-label by running:

python crf_postprocessing.py

6. Finally, use the pseudo label to train a segmentation network by running:

python training_segmentation.py

Citing this work

If you find this work useful, please cite: Multi Path Region Mining For Weakly Supervised 3D Semantic Segmentation on Point Clouds

@inproceedings{wei2020multi,
  title={Multi-Path Region Mining For Weakly Supervised 3D Semantic Segmentation on Point Clouds},
  author={Wei, Jiacheng and Lin, Guosheng and Yap, Kim-Hui and Hung, Tzu-Yi and Xie, Lihua},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={4384--4393},
  year={2020}
}

Acknowledgement

This project is developed based on KPConv (repo). We also thanks dense-crf(repo) and nanoflann.

License

Our code is released under MIT License (see LICENSE file for details).