Performance evaluation of MPEG-2 codec with accurate motion estimation

Conventional motion estimation algorithms that rely on displacement motion vectors do not take into account potential object transformations such as scaling, rotations and deformations in a video scene. Our prior work in isolated test environment where motion is estimated and compensated using original reference frames has demonstrated that transformation estimation algorithms based on Lie operators can significantly improve the motion estimation accuracy achievable by using only the conventional translation estimation. In this paper, we evaluate the improvement of the prediction performance after our transformation estimation algorithms are integrated into an MPEG-2 codec, where prediction errors tend to propagate to succeeding frames. The results show that transformation estimation can lead to higher PSNR (peak signal-to-noise ratio) values for predicted MPEG-2 frames in two standard test sequences.