Project webpage: https://sites.google.com/site/yihsuantsai/research/cvpr16-segmentation
Contact: Yi-Hsuan Tsai (wasidennis at gmail dot com)
Video Segmentation via Object Flow
Yi-Hsuan Tsai, Ming-Hsuan Yang and Michael J. Black
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
This is the authors' MATLAB implementation described in the above paper. Please cite our paper if you use our code and model for your research.
This code has been tested on Ubuntu 14.04 and MATLAB 2013b.
Download and unzip the code.
Install the attached caffe branch, as instructed at http://caffe.berkeleyvision.org/installation.html.
Download the CNN model for feature extraction here, then unzip the model folder under the caffe-cedn-dev/examples folder.
Install included libraries in the External folder if needed (pre-compiled codes are already included).
Put your video data in the Videos folder (see examples in this folder).
Set directories and parameters in
setup_all.m(suggest to use defaults).
demo_objectFlow.mand change settings if needed based on your video data (see the script for further details).
Currently this package only contains the implementation of object segment tracking without re-estimating optical flow and the performacne is a bit worse than the one reported in the paper.
For initialization, currently we use the ground truth of the first frame and propagate to following frames. If you prefer to use other initializations, please replace the ground truth data.
The model and code are available for non-commercial research purposes only.
- The current implementation for generating optical flow is slow, so you can replace it with other optical flow methods to speed up the process.
- 06/2016: code released
- 09/2016: evaluation method updated
- 10/2016: code updated for supervoxel extraction and online CNN model