Rate-distortion-based combined motion estimation and segmentation
Yan Qin Yang, Sheila S. Hemami · 2002
This paper unifies the problems of variable block size motion estimation (VBSME) and region-based segmentation using a quadtree approach. The quadtree's nodes are each labeled providing region-based segmentation, and the same quadtree is also used for block motion estimation. The Lagrange multiplier method and a tree pruning technique are used to find the optimal tree which achieves the joint optimization of VBSME and the segmentation representation in the rate distortion sense. Perceptually-based bit allocation is achieved with a region-weighted distortion measure. A video coder employing this technique results 0.1-0.7 dB gains over MPEG-2. Moreover, the proposed coder combined with perceptual weighting achieves uniform visual quality across each frame. The proposed approach can be easily adapted to MPEG-4 based coders.