/IVOS-W

[CVPR'21] Learning to Recommend Frame for Interactive Video Object Segmentation in the Wild

Primary LanguagePythonMIT LicenseMIT

IVOS-W

Paper

Learning to Recommend Frame for Interactive Video Object Segmentation in the Wild

Zhaoyun Yin, Jia Zheng, Weixin Luo, Shenhan Qian, Hanling Zhang, Shenghua Gao.

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021.

[arXiv] [Paper] [Supp. Material]

Getting Started

Create the environment

# create conda env
conda create -n ivosw python=3.7
# activate conda env
conda activate ivosw
# install pytorch
conda install pytorch=1.3 torchvision
# install other dependencies
pip install -r requirements.txt

We adopt MANet, IPN, and ATNet as the VOS algorithms. Please follow the instructions to install the dependencies.

git clone https://github.com/yuk6heo/IVOS-ATNet.git VOS/ATNet
git clone https://github.com/lightas/CVPR2020_MANet.git VOS/MANet
git clone https://github.com/zyy-cn/IPN.git VOS/IPN

Dataset Preparation

  • DAVIS 2017 Dataset
    • Download the data and human annotated scribbles here.
    • Place DAVIS folder into root/data.
  • YouTube-VOS Dataset
    • Download the YouTube-VOS 2018 version here.
    • Clean up the annotations following here.
    • Download our annotated scribbles here.

Create a DAVIS-like structure of YouTube-VOS by running the following commands:

python datasets/prepare_ytbvos.py --src path/to/youtube_vos --scb path/to/scribble_dir

Evaluation

For evaluation, please download the pretrained agent model and quality assessment model, then place them into root/weights and run the following commands:

python eval_agent_{atnet/manet/ipn}.py with setting={oracle/wild} dataset={davis/ytbvos} method={random/linspace/worst/ours}

The results will be stored in results/{VOS}/{setting}/{dataset}/{method}/summary.json

Note: The results may fluctuate slightly with different versions of networkx, which is used by davisinteractive to generate simulated scribbles.

Training

First, prepare the data used to train the agent by downloading reward records and pretrained experience buffer, place them into root/train, or generate them from scratch:

python produce_reward.py
python pretrain_agent.py

To train the agent:

python train_agent.py

To train the segmentation quality assessment model:

python generate_data.py
python quality_assessment.py

Citation

@inproceedings{IVOSW,
  title     = {Learning to Recommend Frame for Interactive Video Object Segmentation in the Wild},
  author    = {Zhaoyuan Yin and
               Jia Zheng and
               Weixin Luo and
               Shenhan Qian and
               Hanling Zhang and
               Shenghua Gao},
  booktitle = {CVPR},
  year      = {2021}
}

LICENSE

The code is released under the MIT license.