Joint motion estimation and segmentation based on the MDL principle

Abstract

We consider the problem of joint motion segmentation and estimation. An algorithm based on region-merging and the minimum description length (MDL) principle is presented here. The proposed technique is an unsupervised bottom-up merging method which could be separated into two parts: initial segmentation and further segmentation. The error map of the motion-compensated prediction error (MCPE) of each region is used to check the motion correlation of the contiguous regions and to obtain the motion vectors of the newly merged one. The MDL-based criteria are used to evaluate each possible fusion in further segmentation. The criteria are also used as a test for stopping the merging process. Experimental results demonstrate the efficiency of the proposed method.