luciennnnnnn / spa-then-temp Goto Github PK

This is the official code for "Spatial-then-Temporal Self-Supervised Learning for Video Correspondence" (CVPR'23).

If you find this repository useful for your research, please cite our paper:

@inproceedings{li2023spatial,
  title={Spatial-then-Temporal Self-Supervised Learning for Video Correspondence},
  author={Li, Rui and Liu, Dong},
  booktitle={CVPR},
  pages={2279--2288},
  year={2023}
}

• Python 3.7.11
• PyTorch 1.8.0
• mmcv-full == 1.3.18
• davis2017-evaluation

To get started, first please clone the repo

git clone https://github.com/qianduoduolr/Spa-then-Temp

Then, please run the following commands:

conda create -n spa_then_temp python=3.7.11
conda activate spa_then_temp

pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install mmcv-full==1.3.18 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.8.0/index.html
pip install -r requirements.txt
pip install future tensorboard

# setup for davis evaluation
git clone https://github.com/davisvideochallenge/davis2017-evaluation.git && cd davis2017-evaluation
python setup.py develop

We have rerun part of the main experiments in our paper. The reproduced performance in this repo may be slightly different from the paper.

Install ImageNet dataset following the official PyTorch ImageNet training code.

Please download the zip file train.zip from the official website. Then, unzip and place it todata/ytv. Besides, please move the youtube2018_train.json in data/data_info/ to data/ytv.

DAVIS-2017 dataset could be downloaded from the official website. We use the 480p validation set for evaluation. Please move the davis2017_val_list.json in data/data_info/ to data/davis/ImageSets.

Please download the data (Videos, Joint positions) from official website, unzip and place them in data/jhmdb. Please move the val_list.txt in data/data_info/ to data/jhmdb.

Please download the data (only the VIP_Fine) from official website, unzip and place them in data/vip.

The overall data structure is as followed:

├── data
│   ├── imagenet
│   │   ├── train
│   │   │   ├── n01440764/
│   │   │   │── ...
│   ├── ytv
│   │   ├── train
│   │   │   ├── JPEGImages
│   │   │   │   ├──003234408d/
│   │   │   │   ├──...
│   │   │   ├── youtube2018_train.json
│   ├── davis
│   │   ├── Annotations
│   │   │   ├── 480p
│   │   │   │   ├── bike-packing/
│   │   │   │   ├── ...
│   │   ├── ImageSets
│   │   │   ├── davis2017_val_list.json
│   │   │   ├── ...
│   │   ├── JPEGImages
│   │   │   ├── 480p
│   │   │   │   ├── bike-packing/
│   │   │   │   ├── ...
│   ├── jhmdb
│   │   ├── Rename_Images
│   │   │   ├── brush_hair/
│   │   │   ├── ...
│   │   ├── joint_positions
│   │   │   ├── brush_hair/
│   │   │   ├── ...
│   │   ├── val_list.txt
│   ├── vip
│   │   ├── VIP_Fine
│   │   │   ├── Annotations/
│   │   │   ├── Images/
│   │   │   ├── lists/

The evaluation is conducted on three correspondence-related tasks, including semi-supervised video object segmentation, keypoint tracking, and human part propagation. Here we show some evaluation results.

We follow the prior studies to leverage label propagation for inference, which can be achieved by:

bash tools/dist_test.sh ${CONFIG}  ${GPUS}  ${TASK}  ${CKPT}

Note the TASK consists of davis (for VOS), jhmdb (for keypoint tracking), and vip (for human part propagation).

We give a inference cmd example:

# testing ResNet-18 with a stride of 8 for video object segmentation
bash tools/dist_test.sh configs/test/res18_d8.py 4 davis ckpt/res18_d8_temp.pth

The results will be saved to eval/. If you have enough memory, you can set the save_np=False and increase the step of test_cfg in config for faster inference.

Our training pipeline consists of two seperate steps, including spatial and temporal feature learning.

In the first step, we learn the spatial-discriminative features with contrastive learning. Please refer to these official implementations (MoCo, DetCo) for image-level self-supervised learning.

Then we perform temporal feature learning on YouTube-VOS:

bash tools/dist_train.sh configs/train/train.py 4

Please modify the pretrained for backbone_t in configs/train/train.py, in order to regard the model learned in the first step as teacher.

The codebase is implemented based on the MMCV and VFS. Thanks for these excellent open source repositories.