Motion-based object segmentation and estimation using the MDL principle

There are increasing demands for ultralow bit-rate video image transmission. We present a new formulation of moving object segmentation and motion estimation to quantify the potential gain in using object-oriented motion compensation in image sequence coding. Motivated by real-valued parameter estimation required in object-oriented motion-compensated video source coding, a framework motivated by Rissanen's (1983) minimum description length (MDL) principle is proposed to more tightly couple motion estimation and object segmentation algorithms to the overall objective of minimizing source bit rate. A new objective function is constructed, and a suboptimal procedure to segment and estimate moving objects in a scene is proposed. Each object is represented by chain-coded block boundaries, affine motion parameters, and motion-compensated prediction error. A number of experimental comparisons between block- and object-oriented coding schemes suggests a significant potential coding gain using object-oriented motion-compensated coding.< >