Skip to content

Latest commit

 

History

History

scene_seg

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
config/s3dis
config/s3dis
 
 
data/s3dis
data/s3dis
 
 
figure
figure
 
 
lib
lib
 
 
model
model
 
 
tool
tool
 
 
util
util
 
 
README.md
README.md
 
 

README.md

3D Semantic Segmentation

Dataset

  • Download S3DIS dataset and symlink the paths to them as follows (you can alternatively modify the relevant paths specified in folder config): 
     mkdir -p dataset
 ln -s /path_to_s3dis_dataset dataset/s3dis

Usage

  1. Requirement:

    • Hardware: 1 GPU to hold 6000MB for CUDA version, 2 GPUs to hold 10000MB for non-CUDA version.
    • Software: PyTorch>=1.5.0, Python3.7, CUDA>=10.2, tensorboardX, tqdm, h5py, pyYaml

  2. Train:

    • Specify the gpu used in config and then do training:

      sh tool/train.sh s3dis pointnet2_paconv                   # non-cuda version
sh tool/train.sh s3dis pointnet2_paconv_cuda              # cuda version

  3. Test:

    • Download pretrained models and put them under folder specified in config or modify the specified paths. Our CUDA-implemented PAConv achieves 66.01 mIoU (w/o voting) and vanilla PAConv without CUDA achieves 66.33 mIoU (w/o voting) in s3dis Area-5 validation set.

    • For full testing (get listed performance):

      CUDA_VISIBLE_DEVICES=0 sh tool/test.sh s3dis pointnet2_paconv        # non-cuda version
CUDA_VISIBLE_DEVICES=0 sh tool/test.sh s3dis pointnet2_paconv_cuda   # cuda version

    • For 6-fold validation (calculating the metrics with results from different folds merged):

      1. Change the test_area index in the config file to 1;
      2. Finish full train and test, the test result files of Area-1 will be saved in corresponding paths after the test;
      3. Repeat a,b by changing the test_area index to 2,3,4,5,6 respectively;
      4. Collect all the test result files of all areas to one directory and state the path to this directory here;
      5. Run the code for 6-fold validation to get the final 6-fold results: 
        python tool/test_s3dis_6fold.py

    • The code for visualizing the segmentation result is also incorporated in test_s3dis_6fold.py (called at line #55-#57), you can run it after generating the pickle file via finishing test.

Citation

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

@inproceedings{xu2021paconv,
  title={PAConv: Position Adaptive Convolution with Dynamic Kernel Assembling on Point Clouds},
  author={Xu, Mutian and Ding, Runyu and Zhao, Hengshuang and Qi, Xiaojuan},
  booktitle={CVPR},
  year={2021}
}

Contact

You are welcome to send pull requests or share some ideas with us. Contact information: Mutian Xu (mino1018@outlook.com) or Runyu Ding (ryding@eee.hku.hk).

Acknowledgement

The code is partially borrowed from PointWeb.