Awesome

Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation (ICCV 2023)

This is the official repository for Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation [arXiv][CVF]

Environmental Setup

• OS: Ubuntu 22.04
• CUDA: 11.7
• Conda environment (You can also refer to ReDAL for some details. Our code is modified on their codes. We thank the authors for sharing the codes.):

conda create -n HPAL python=3.6 -y
conda install pytorch==1.8.0 torchvision==0.9.0  cudatoolkit=11.1 -c pytorch -c conda-forge -y
conda install pytorch-scatter -c pyg -y
conda install scikit-learn=0.24.2 -y
conda install pyyaml=5.3.1 -y
conda install tqdm=4.61.1 -y
conda install pandas=1.3.2 -y
conda install pyntcloud -c conda-forge -y
conda install plyfile -c conda-forge -y
conda install cython -y
conda install h5py==2.10.0 -y
pip3 install open3d-python==0.3.0
pip3 install --upgrade git+https://github.com/mit-han-lab/torchsparse.git@v1.2.0 (depend on libsparsehash-dev)
install libsparsehash-dev (no sudo permissions):
      git clone https://github.com/sparsehash/sparsehash.git
      cd sparsehash
      ./configure --prefix=/path/you/want/local
      make
      make install
      gedit(vim) ~/.bashrc
      export CPLUS_INCLUDE_PATH=/path/you/want/local/include
      source ~/.bashrc
install libsparsehash-dev (with sudo permissions):
      sudo apt-get install libsparsehash-dev

• Compile cpp-related utils:

sh compile_op.sh

Data preparation

1. Fill the google form to get the dataset download link (download Stanford3dDataset_v1.2_Aligned_Version.zip)

2. S3DIS Data Preprocessing

   1. extract "Stanford3dDataset_v1.2_Aligned_Version.zip".
   2. Modify STANFORD_3D_IN_PATH and STANFORD_3D_OUT_PATH in data_preparation/data_prepare_s3dis.py and run it:

   python3 data_prepare_s3dis.py
   

After that, the file organization will be like:

S3DIS/
├── Area_1
│   ├── coords
│   │   ├── conferenceRoom_1.npy
│   │   ...
│   ├── labels
│   │   ├── conferenceRoom_1.npy
│   │   ...
│   ├── rgb
│   │   ├── conferenceRoom_1.npy
│   │   ...
│   └── proj
│       ├── conferenceRoom_1.pkl
│       ...
├── Area_2
...

Training and inference

A. Active semi-supervised training (Our method):

1. Build the initial labelled data (Randomly labeling a small part of data):
   1. Modify root in init_labeled_points_s3dis.py to processed data path.
   2. Modify save_path in init_labeled_points_s3dis.py to your desired result save path.
   3. Modify labeled_num_cur_pc in init_labeled_points_s3dis.py to the number of initial labelled points for each point cloud.
   4. Run init_labeled_points_s3dis.py:

   python init_labeled_points_s3dis.py
   

Then in your save_path, you can get a dictionary that each item saves a boolean array to stand whether each point is labelled.

You can also get the initial labelled data that we use for the reported experiments in our paper by downloading from google drive. The number of initial labelled data are 1/5 of the total labelling budget in each setting.

2. Modify configuration parameters under class ConfigS3DIS in config.py, change of the following parameters is necessary:

   data_path = '/your_data_path'  # Data root path after preparation
   init_labeled_data = '/initial_data_index'  # Path of initial labeled data dictionary
   base_path = '/results'  # Path to save the training results
   active_strategy = 'HMMU' # Scoring strategy for active learning
   chosen_rate_AL = 0.02 # The percentage of selected points in each iteration, which is 1/5 of the total labeling budget in our setting.
   

3. Training:

   python s3dis_main.py --test_area 5 --mode AL_train
   

B. Fully supervised Training (Upperbound):

python s3dis_main.py --test_area 5 --mode baseline_train 

C. Inference (Test any trained model):

You can test any trained model by following steps:

1. Get a trained model and save it to disk.
2. Run s3dis_main.py:

python s3dis_main.py --mode test --model_path the/path/of/the/trained/model

Pretrained Models

We provide pretrained models for reported results in our paper, you can download them through following links:

Visualization

Paper and Citation

If you find our paper is useful, please cite:

@InProceedings{Xu_2023_ICCV,
    author    = {Xu, Zongyi and Yuan, Bo and Zhao, Shanshan and Zhang, Qianni and Gao, Xinbo},
    title     = {Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {18098-18108}
}