Awesome
Interactive Video Object Segmentation Using Global and Local Transfer Modules
Yuk Heo, Yeong Jun Koh, Chang-Su Kim
<img src = "Overall_Network.png" width="80%" height="80%">Implementation of ECCV2020 paper, "Interactive Video Object Segmentation Using Global and Local Transfer Modules"
Codes in this github:
- DAVIS2017 evaluation based on the DAVIS framework
- DAVIS2016 real-world evaluation GUI
Prerequisite
- cuda 10.0
- python 3.6
- pytorch 1.2.0
- davisinteractive 1.0.4
- corrlation package of FlowNet2
- numpy, cv2, PtQt5, and other general libraries of python3
Directory Structure
-
root/libs
: library of utility files. -
root/networks
: network codes.correlation_package.zip
: conserves GPU memory by appling the correlation package of FlowNet2.deeplab
: applies ASPP module in decoders. [original code]atnet.py
: consists A-Net and T-Net.ltm_transfer.py
: transfers previous segmentation with the local affinity of the local transfer module.
-
root/config.py
: configurations. -
root/eval_davis-framework.py
: DAVIS2017 evaluation based on the DAVIS framework. -
root/eval_real-world.py
: DAVIS2016 real-world evaluation GUI (to be released).
Instruction
DAVIS2017 evaluation based on the DAVIS framework
- Edit
config.py
to set the directory of your DAVIS2017 dataset and the gpu ID. - Unzip and build corrlation package by
cd ./networks
unzip correlation_package.zip
cd correlation_package
rm -rf *_cuda.egg-info build dist __pycache__
python3 setup.py install --user
If you have problems in this step, you can find more information in the repository.
- Download our network parameters and place the file as
root/ATNet-checkpoint.pth
. - Run with
python3 eval_davis-framework.py
.
DAVIS2016 real-world evaluation GUI
Multi-object GUI (for DAVIS2017) is available at our github page, [GUI-IVOS]
Reference
Please cite our paper if the implementations are useful in your work:
@Inproceedings{
Yuk2020IVOS,
title={Interactive Video Object Segmentation Using Global and Local Transfer Modules},
author={Yuk Heo and Yeong Jun Koh and Chang-Su Kim},
booktitle={ECCV},
year={2020},
url={https://openreview.net/forum?id=bo_lWt_aA}
}