Computer Science · Computer Vision and Pattern Recognition
Fast motion estimation based on the redundant wavelet for 2D+T wavelet coder
Heechan Park, Graham R. Martin, Andy C. Yu
See on the mapAbstract
Scalable video coding utilizing the wavelet transform applied to both the spatial and temporal domains (3D-DWT) is of current interest in video coding. In the case of the 2D+t scheme, motion estimation is required to be conducted in the wavelet domain. We explore approaches that overcome the well-known shift-variant property of the wavelet transform so that motion is estimated accurately. Specifically, an algorithm using the redundant wavelet and phase-multi-hypothesis shows PSNR superiority but computational complexity is high. We propose a fast motion search method that significantly reduces computational complexity to as low as 2% of the existing algorithm yet still shows better performance than full-spatial motion estimation.