Fast Video Segmentation

Anestis Papazoglou, Vittorio Ferrari


This software implements the full video segmentation pipeline of [1]. It produces foreground-background segmentations of a test video, without any manual initialisation. The software was developed under 64-bit Linux with Matlab R2012b/R2013a. There is no guarantee it will run on other operating systems or Matlab versions (though it probably will).

How fast is it?

The code typically takes 0.5 seconds/frame on a desktop PC, excluding the optical flow computation. For the optical flow, we have bundled the code of [2] in our release, which we have found that it takes 3.5 seconds/frame on average for an image size of 350x400. If you have a CUDA capable GPU, we also bundle a wrapper for the code of [3] which typically takes 1 second/frame for the same image size.


The software has been used by Han et al. placing second in the object detection from video category of ILSVRC2015.



[1] A. Papazoglou and V. Ferrari, Fast object segmentation in unconstrained video. In ICCV, 2013.
[2] T. Brox and J. Malik. Object segmentation by long term analysis of point trajectories. In ECCV, 2010.
[3] N. Sundaram, T. Brox, and K. Keutzer. Dense point trajectories by gpu-accelerated large displacement optical flow. In ECCV, 2010.