This is the official demo code for the paper. PDF
- Ubuntu
- python 3.6
- Pytorch 0.3.1
- installed with CUDA.
- Download DAVIS-2017.
- Edit path for
DAVIS_ROOT
in run.py.
DAVIS_ROOT = '<Your DAVIS path>'
- Download weights.pth and place it the same folde as run.py.
- To run single-object video object segmentation on DAVIS-2016 validation.
python run.py
- To run multi-object video object segmentation on DAVIS-2017 validation.
python run.py -MO
- Results will be saved in
./results/SO
or./results/MO
.
While our training script will not be released officially, xanderchf writes a great training script. Check it here:
https://github.com/xanderchf/RGMP
For pre-training, it is highly recommended to use recent large-scale Youtube-VOS dataset if you want to skip data synthesis from static images (Sect 3.2 in the paper) which is a headache.
If you use this code please cite:
@InProceedings{oh2018fast,
author = {Oh, Seoung Wug and Lee, Joon-Young and Sunkavalli, Kalyan and Kim, Seon Joo},
title = {Fast Video Object Segmentation by Reference-Guided Mask Propagation},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2018}
}
Please check out our NEW approach!
Video Object Segmentation using Space-Time Memory Networks
Seoung Wug Oh, Joon-Young Lee, Ning Xu, Seon Joo Kim
ICCV 2019