Parametric warping for motion estimation

Abstract

Summary form only given. In warping (also known as mesh-based) motion estimation, motion vectors at individual pixels are computed through an interpolation of a subsampled set of motion vectors. A method for calculating optimal warping coefficients was introduced previously. This algorithm finds the interpolation coefficients, at each individual pixel location (within a block), such that the mean squared luminance errors are minimized. It has been observed that optimal coefficients vary widely with time and across different sequences. This observation motivates the optimization of the warping coefficients locally in time. However, doing so requires the encoder to transmit the coefficients to the decoder. Assuming a 16/spl times/16 block and four floating point coefficients per pixel, this would require a considerable overhead in bitrate. Especially in low bitrate regimes, such overhead is likely to be unacceptable. This paper proposes a parametric class of functions to represent the warping interpolation kernels. More specifically, we propose to use the two-parameter family of functions.