Skip to content

Latest commit

 

History

History
66 lines (46 loc) · 1.53 KB

README.md

File metadata and controls

66 lines (46 loc) · 1.53 KB

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:

arch

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
  1. Install cocoapi for Youtube-VIS.
  2. Download the ImageNet pretrained backbone and put it into the pretrain_model folder.
pretrain_model
└─ resnet50_v2.pth
  1. 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: