Domain Agent Network

Overview

This code is for the paper "Delving Deep into Many-to-many Attention for Few-shot Video Object Segmentation" in CVPR2021.

The architecture of our Domain Agent Network:

Environment

conda create -n FSVOS python=3.6
conda activate FSVOS
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.2 -c pytorch
conda install opencv cython
pip install easydict imgaug

Usage

Preparation

1. Download the 2019 version of Youtube-VIS dataset.
2. Put the dataset in the ./data folder.

data
└─ Youtube-VOS
    └─ train
        ├─ Annotations
        ├─ JPEGImages
        └─ train.json

3. Install cocoapi for Youtube-VIS.
4. Download the ImageNet pretrained backbone and put it into the pretrain_model folder.

pretrain_model
└─ resnet50_v2.pth

5. Update the root_path in config/DAN_config.py.

Training

python train_DAN.py --group 1 --batch_size 4

Inference

You can download our pretrained model to test.

python test_DAN.py --test_best --group 1

References

Part of the code is based upon:

• PMMs: https://github.com/Yang-Bob/PMMs
• PFENet: https://github.com/Jia-Research-Lab/PFENet
• STM-Training: https://github.com/lyxok1/STM-Training