Guiding optical flow estimation using superpixels

Theodosios Gkamas, Christophoros Nikou

Abstract: In this paper, we show how the segmentation of an image into superpixels may be used as preprocessing paradigm to improve the accuracy of the optical flow estimation in an image sequence. Superpixels play the role of accurate support masks for the integration of the optical flow equation. We employ a variation of a recently proposed optical flow algorithm relying on local image properties that are taken into account only if the involved pixels belong to the same image segment. Experimental results show that the proposed optical flow estimation scheme significantly improves the accuracy of the estimated motion field with respect to other standard methods.