[CVPR2021 Oral] End-to-End Video Instance Segmentation with Transformers

VisTR: End-to-End Video Instance Segmentation with Transformers

This is the official implementation of the VisTR paper:

Installation

We provide instructions how to install dependencies via conda. First, clone the repository locally:

git clone https://github.com/Epiphqny/vistr.git

Then, install PyTorch 1.6 and torchvision 0.7:

conda install pytorch==1.6.0 torchvision==0.7.0

Install pycocotools

conda install cython scipy
pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
pip install git+https://github.com/youtubevos/cocoapi.git#"egg=pycocotools&subdirectory=PythonAPI"

Compile DCN module(requires GCC>=5.3, cuda>=10.0)

cd models/dcn
python setup.py build_ext --inplace

Preparation

Download and extract 2019 version of YoutubeVIS train and val images with annotations from CodeLab or YoutubeVIS. We expect the directory structure to be the following:

VisTR
├── data
│   ├── train
│   ├── val
│   ├── annotations
│   │   ├── instances_train_sub.json
│   │   ├── instances_val_sub.json
├── models
...

Download the pretrained DETR models on COCO and save it to the pretrained path.

Training

Training of the model requires at least 32g memory GPU, we performed the experiment on 32g V100 card.

To train baseline VisTR on a single node with 8 gpus for 16 epochs, run:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --backbone resnet101/50 --ytvos_path /path/to/ytvos --masks --pretrained_weights /path/to/pretrained_path

Inference

python inference.py --masks --model_path /path/to/model_weights --save_path /path/to/results.json

Models

We provide baseline VisTR models, and plan to include more in future. AP is computed on YouTubeVIS dataset by submitting the result json file to the CodeLab system, and inference time is calculated by pure model inference time (without data-loading and post-processing).

License

VisTR is released under the Apache 2.0 license. Please see the LICENSE file for more information.

Acknowledgement

We would like to thank the DETR open-source project for its awesome work, part of the code are modified from its project.

Citation

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follow.

@inproceedings{wang2020end,
  title={End-to-End Video Instance Segmentation with Transformers},
  author={Wang, Yuqing and Xu, Zhaoliang and Wang, Xinlong and Shen, Chunhua and Cheng, Baoshan and Shen, Hao and Xia, Huaxia},
  booktitle =  {Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR)},
  year={2021}
}