ObjectFlow

Project webpage: https://sites.google.com/site/yihsuantsai/research/cvpr16-segmentation
Contact: Yi-Hsuan Tsai (wasidennis at gmail dot com)

Paper

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.

Overview

• 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.

Installation

• 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).

Usage

• 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).

• Run demo_objectFlow.m and change settings if needed based on your video data (see the script for further details).

Note

• 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.

Hint

• The current implementation for generating optical flow is slow, so you can replace it with other optical flow methods to speed up the process.

Log

• 06/2016: code released
• 09/2016: evaluation method updated
• 10/2016: code updated for supervoxel extraction and online CNN model