part_seg

3D Shape Part Segmentation

Installation

Requirements

• Hardware: GPUs to hold 14000MB
• Software: Linux (tested on Ubuntu 18.04) PyTorch>=1.5.0, Python>=3, CUDA>=10.1, tensorboardX, tqdm, pyYaml

Dataset

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

mkdir -p data
ln -s /path to shapenet part/shapenetcore_partanno_segmentation_benchmark_v0_normal data

Usage

• Build the CUDA kernel:

  When you run the program for the first time, please wait a few moments for compiling the cuda_lib automatically. Once the CUDA kernel is built, the program will skip this in the future running.

• Train:

  • Multi-thread training (nn.DataParallel) :

    • python main.py --config config/dgcnn_paconv_train.yaml (Embed PAConv into DGCNN)

  • We also provide a fast multi-process training (nn.parallel.DistributedDataParallel, recommended) with official nn.SyncBatchNorm. Please also remind to specify the GPU ID:

    • CUDA_VISIBLE_DEVICES=x,x python main_ddp.py --config config/dgcnn_paconv_train.yaml (Embed PAConv into DGCNN)

• Test:

  • Download our pretrained model and put it under the part_seg folder.

  • Run the voting evaluation script to test our pretrained models, after this voting you will get an instance mIoU of 86.1% if all things go right:

    python eval_voting.py --config config/dgcnn_paconv_test.yaml

  • You can also directly test our pretrained model without voting to get an instance mIoU of 86.0%:

    python main.py --config config/dgcnn_paconv_test.yaml

  • For full test after training the model:

    • Specify the eval to True in your config file.

    • Make sure to use main.py (main_ddp.py may lead to wrong result due to the repeating problem of all_reduce function in multi-process training) :

      python main.py --config config/your config file.yaml

  • You can choose to test the model with the best instance mIoU, class mIoU or accuracy, by specifying model_type to insiou, clsiou or acc in the test config file.

• Visualization: tensorboardX incorporated for better visualization.

  tensorboard --logdir=checkpoints/exp_name

Citation

If you find the code or trained models useful, 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

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