ReDAL: Region-based and Diversity-aware Active Learning for Point Cloud Semantic Segmentation

Official pytorch implementation of "ReDAL: Region-based and Diversity-aware Active Learning for Point Cloud Semantic Segmentation. Wu et al. ICCV 2021." (presentation video)

In this work, we present ReDAL, a general active learning framework for point cloud semantic segmentation. By selecting only informative but diverse regions for label acquisition, the labeling effort can be hugely reduced. Here shows a demo video.

demo_video.mp4

Environmental Setup

• OS: Ubuntu 18.04
• CUDA: 11.1
• Install required packages (via conda)

  conda install pytorch==1.8.0 torchvision==0.9.0  cudatoolkit=11.1 -c pytorch -c conda-forge
  conda install pytorch-scatter -c pyg
  conda install scikit-learn=0.24.2
  conda install pyyaml=5.3.1
  conda install tqdm=4.61.1
  conda install pandas=1.3.2
  conda install pyntcloud -c conda-forge
  conda install plyfile -c conda-forge
  pip3 install --upgrade git+https://github.com/mit-han-lab/[email protected]

Note:

1. One might need to install libsparsehash-dev before installing torchsparse. For more information, please see the official website.
2. We only support [email protected].
3. torchsparse will automatically install the corresponding version of the package based on the the highest supported CUDA version of the nvidia-driver, where would be an alternative to install the torchsparse package.

Data Preparation

This repo supports S3DIS and SemanticKITTI datasets. (Support for scannetv2 dataset is still in beta.)

Please see this documentation.

Active Training Script

A. Supervised Training (Upperbound)

• The script supports supervised training on the whole dataset, which is the upper-bound in our experiment.

CUDA_VISIBLE_DEVICES=X python3 train_supervision.py -n <dataset> -d <dataset-path> -p <stored-ckpt-path> \
                        -m <model-backbone> --train_batch_size <train-batch-size> --val_batch_size <val-batch-size> \
                        --ignore_idx <invalid-category-idx> --max-epoch <epochs> [--distributed_training]

B. Scene-based Active Leraning (Baselines)

• The script supports multiple scene-based active learning strategies, which selects a batch of point cloud scans for label acquisition in an active iteration.
• Supported active_method flags: ["random", "softmax_confidence", "softmax_margin", "softmax_entropy", "mc_dropout", "segment_entropy", "core_set"].

CUDA_VISIBLE_DEVICES=X python3 train_scene_active.py -n <dataset> -d <dataset-path> -p <stored-ckpt-directory> \
                        -m <model-backbone> --train_batch_size <train-batch-size> --val_batch_size <val-batch-size> \
                        --ignore_idx <invalid-category-idx> --training-epoch <epochs> --finetune-epoch <epochs> \
                        --active_method <valid-active-method> --max_iterations <AL iterations> \
                        --active_selection_size <number of scene labels per query> [--distributed_training]

C. Region-based Active Learning

• The script supports multiple region-based active learning strategies, which selects a batch of divided sub-scene regions for label acquisition in an active iteration.
• Supported active_method flags: ["random", "softmax_confidence", "softmax_margin", "softmax_entropy", "mc_dropout", "ReDAL"].
• ReDAL is our proposed method.

CUDA_VISIBLE_DEVICES=X python3 train_region_active.py -n <dataset> -d <dataset-path> -p <stored-ckpt-directory> \
                        -m <model-backbone> --train_batch_size <train-batch-size> --val_batch_size <val-batch-size> \
                        --ignore_idx <invalid-category-idx> --training-epoch <epochs> --finetune-epoch <epochs> \
                        --active_method <valid-active-method> --max_iterations <AL iterations> \
                        --active_percent <percent of labels per query> [--distributed_training]

Citation

If you use this code, please cite the paper:

@inproceedings{wu2021redal,
  title={ReDAL: Region-based and Diversity-aware Active Learning for Point Cloud Semantic Segmentation},
  author={Wu, Tsung-Han and Liu, Yueh-Cheng and Huang, Yu-Kai and Lee, Hsin-Ying and Su, Hung-Ting and Huang, Ping-Chia and Hsu, Winston H},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={15510--15519},
  year={2021}
}