Home

Awesome

VPSNet for Video Panoptic Segmentation

Official implementation for "Video Panoptic Segmentation" (CVPR 2020 Oral) <br/>[Paper] [Dataset] [Project] [Slides] [Codalab]

Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon.

Cityscapes-VPS <b>test</b> set evaluation is now available at this <a href="https://competitions.codalab.org/competitions/26183">Codalab server.</a>

<img src="./image/landscape.png" width="800"> <br/> <img src="./image/panoptic_pair_240.gif" width="800"> <br/> Image-level baseline (left) / VPSNet result (right)

Disclaimer

This repo is tested under Python 3.7, PyTorch 1.4, Cuda 10.0, and mmcv==0.2.14.

Installation

a. This repo is built based on mmdetection commit hash 4357697. Our modifications for VPSNet implementation are listed here. Please refer to INSTALL.md to install the library. You can use following commands to create conda env with related dependencies.

conda create -n vps python=3.7 -y
conda activate vps
conda install pytorch=1.4 torchvision cudatoolkit=10.0 -c pytorch -y
pip install -r requirements.txt
pip install "git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI"
pip install "git+https://github.com/cocodataset/panopticapi.git"
pip install -v -e .

b. You also need to install dependencies for Flownet2 and UPSNet modules.

bash ./init_flownet.sh
bash ./init_upsnet.sh

c. You may also need to download some pretrained weights.

pip install gdown
bash ./download_weights.sh

Dataset

You can download Cityscapes-VPS here. It provides 2500-frame panoptic labels that temporally extend the 500 Cityscapes image-panoptic labels. There are total 3000-frame panoptic labels which correspond to 5, 10, 15, 20, 25, and 30th frames of each 500 videos, where all instance ids are associated over time. <img src="./image/dataset.png" width="1000">

It not only supports video panoptic segmentation (VPS) task, but also provides super-set annotations for video semantic segmentation (VSS) and video instance segmentation (VIS) tasks.

Necessary data for Cityscapes-VPS training, testing, and evaluation are as follows. Please refer to DATASET.md for dataset preparation.

mmdetection
├── mmdet
├── tools
├── configs
├── data
│   ├── cityscapes_vps
│   │   ├── panoptic_im_train_city_vps.json
│   │   ├── panoptic_im_val_city_vps.json
│   │   ├── panoptic_im_test_city_vps.json  
│   │   ├── instances_train_city_vps_rle.json (for training)
│   │   ├── instances_val_city_vps_rle.json 
│   │   ├── im_all_info_val_city_vps.json (for inference)
│   │   ├── im_all_info_test_city_vps.json (for inference)
│   │   ├── panoptic_gt_val_city_vps.json (for VPQ eval)
│   │   ├── train 
│   │   │   ├── img
│   │   │   ├── labelmap
│   │   ├── val
│   │   │   ├── img
│   │   │   ├── img_all
│   │   │   ├── panoptic_video
│   │   ├── test
│   │   │   ├── img_all

VIPER dataset is provided by "Playing for Benchmarks, ICCV 2017". We use the 'day' split of the dataset by converting it into the video panoptic segmentation format. The converted dataset is provided upon request.

Evaluation Metric

<img src="./image/vpq_measure.png" width="1000">

Testing

Our trained models are available for download here. Rename it to latest.pth and run the following commands to test the model on Cityscapes-VPS.

python tools/test_vpq.py configs/cityscapes/fusetrack.py \
  work_dirs/cityscapes_vps/fusetrack_vpct/latest.pth \
  --out work_dirs/cityscapes_vps/fusetrack_vpct/val.pkl \
  --pan_im_json_file data/cityscapes_vps/panoptic_im_val_city_vps.json \
  --n_video 50 --mode val \
python tools/eval_vpq.py \
  --submit_dir work_dirs/cityscapes_vps/fusetrack_vpct/val_pans_unified/ \
  --truth_dir data/cityscapes_vps/val/panoptic_video/ \
  --pan_gt_json_file data/cityscapes_vps/panoptic_gt_val_city_vps.json

python tools/test_vpq.py configs/cityscapes/fusetrack.py \
  work_dirs/cityscapes_vps/fusetrack_vpct/latest.pth \
  --out work_dirs/cityscapes_vps/fusetrack_vpct/test.pkl \
  --pan_im_json_file data/cityscapes_vps/panoptic_im_test_city_vps.json \
  --n_video 50 --mode test \

Files containing the predicted results will be generated as pred.json and pan_pred/*.png at work_dirs/cityscapes_vps/fusetrack_vpct/test_pans_unified/.

Cityscapes-VPS test split currently only allows evaluation on the codalab server. Please upload submission.zip to <a href="https://competitions.codalab.org/competitions/26183">Codalab server</a> to see actual performances.

submission.zip
├── pred.json
├── pan_pred.zip
│   ├── 0005_0025_frankfurt_000000_001736.png
│   ├── 0005_0026_frankfurt_000000_001741.png
│   ├── ...
│   ├── 0500_3000_munster_000173_000029.png

Training

# Multi-GPU distributed training
bash ./tools/dist_train.sh configs/cityscapes/fusetrack.py ${GPU_NUM}
# OR
python ./tools/train.py configs/cityscapes/fusetrack.py --gpus ${GPU_NUM}

Citation

If you use this toolbox or benchmark in your research, please cite this project.

@inproceedings{kim2020vps,
  title={Video Panoptic Segmentation},
  author={Dahun Kim and Sanghyun Woo and Joon-Young Lee and In So Kweon},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2020}
}

Terms of Use

This software is for non-commercial use only. The source code is released under the Attribution-NonCommercial-ShareAlike (CC BY-NC-SA) Licence (see this for details)

Acknowledgements

This project has used utility functions from other wonderful open-sourced libraries. We would especially thank the authors of:

Contact

If you have any questions regarding the repo, please contact Dahun Kim (mcahny01@gmail.com) or create an issue.