Space-Time Multi-Resolution Banded Graph-Cut for Fast Segmentation. (Technical Report)

Abstract

Applying real-time segmentation is a major issue when processing every frame of image sequences. In this paper, we propose a modi cation of the well known graph-cut algorithm to improve speed for discrete segmentation. Our algorithm yields real-time segmentation, using graph-cut, by performing a single cut on an image with regions of di erent resolutions, combining space-time pyramids and narrow bands. This is especially suitable for image sequences, as segment borders in one image are re ned in the next image. The fast computation time allows one to use information contained in every image frame of an input image stream at 20 Hz, on a standard PC. The algorithm is applied to traf- c scenes, using a monocular camera installed in a moving vehicle. Our results show the segmentation of moving objects with similar results to standard graph-cut, but with improved speed.