Awesome

Fast User-Guided Video Object Segmentation by Interaction-and-Propagation Networks

Seoung Wug Oh, Joon-Young Lee, Ning Xu, Seon Joo Kim

CVPR 2019

screenshot

Demo code with GUI interface.
paper

Requirements

python 3.6
pytorch 0.3.1
davisinteractive
matplotlib
PyQt5

How to Use

Environment setup

conda create --name ivs python=3.6
source activate ivs

pip install PyQt5 matplotlib opencv-contrib-python pillow Cython
pip install davisinteractive

conda install pytorch=0.3.1 cuda90 -c pytorch
conda install torchvision

Download weights

wget -O I_e290.pth "https://www.dropbox.com/s/khx9hmtnqbzg634/I_e290.pth?dl=1"
wget -O P_e290.pth "https://www.dropbox.com/s/89heglbglig0g04/P_e290.pth?dl=1"

Run!

python gui.py -seq camel

Test your own videos

Locate videos in ./sequences/[name] Run

python gui.py -seq [name]

Quantitative Evaluation

The GUI application in this repository is for demo and it is trained for scribbles from real users. The numbers in our paper will not be reproducable with the checkpoint included here. For the quantitative evaluation using DAVIS framework, evaluation summary is available [download link]. The timing is measured using a single 2080Ti GPU. For the further questions, please contact me by E-mail.

Reference

If you find our paper and repo useful, please cite our paper. Thanks!

Fast User-Guided Video Object Segmentation by Interaction-and-Propagation Networks
Seoung Wug Oh, Joon-Young Lee, Ning Xu, Seon Joo Kim
CVPR 2019

Terms of Use

This software is for non-commercial use only. The source code is released under the Attribution-NonCommercial-ShareAlike (CC BY-NC-SA) Licence (see this for details)