Fast superpixels for video analysis

The computational cost of video and motion analysis can be dramatically reduced by over-segmenting each frame of video into "superpixels". But most superpixel algorithms are themselves computationally expensive, and are thus inappropriate for use with real-time video. This paper advocates and analyzes the use of superpixels derived from minimum-cost paths that can be computed by dynamic programming. It is shown that superpixels can be computed comfortably in real time using such methods (30-40 times faster than the most efficient alternative), while sacrificing about 3% in the accuracy of the superpixels. The efficacy of the approach is demonstrated with a simple video analysis application.