Key-Segments Video Segmentation README File
-------------------------------------------
1. Using the code
There are 2 demo files:
a) hypotheses_demo.m: discovers keysegment hypotheses (i.e., clusters) in ranked-order of their object-like properties.
b) keySegs_demo.m: produces a binary segmentation of the object corresponding to the specified keysegment hypothesis.
2. Notes
The code assumes that the video frames are stored in order (i.e., frame 1 is the first file, frame 2 is the second file),
and that the files types are either jpg, png, or bmp. (The code can be easily changed to accomodate other file types.)
The code assumes that the BPLRs [1], region proposals [2], and optical flows [3] are pre-computed.
You can download the pre-computed data for the SegTrack dataset here: vision.cs.utexas.edu/projects/keysegments/data.tgz
Example code for computing them are provided in ./code/preprocessing/
[1] J. Kim and K. Grauman. Boundary Preserving Dense Local Regions. CVPR 2011.
[2] I. Endres and D. Hoiem. Category Independent Object Proposals. ECCV 2010.
[3] T. Brox and J. Malik. Large displacement optical flow: descriptor matching in variational motion estimation. TPAMI 2011.
The segmentation demo code (keySegs_demo.m) produces slightly different results to those reported in the paper ("Ours" in Table 1);
better for 3 videos, similar for 2 videos, and worse for 1 video. This is due to changes in the BPLR parameters.
The results for the color-only method ("Ours w/o partial shape match" in Table 2) are identical.
3. Citation
Please cite our paper if you use our code:
Y. J. Lee, J. Kim, and K. Grauman. Key-Segments for Video Object Segmentation.
Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2011.
4. Comments
(11/14/2013) I have fixed some hard-coded paths in the preprocessing code. Thanks to Anestis Papazoglou for pointing them out.