New adaptive pixel decimation for block motion vector estimation

It is observed that around the optimum point of the motionestimation process the error criterion function is well modeled as aquadratic function with respect to the motion vector offsets. This locallyquadratic functional model decomposes the motion estimation optimiza-tion at subpixel resolutions into a two-stage pipelinable processes: full-search at full-pixel resolution and interpolation at any subpixel resolution.Practical approximation formulas lead to the explicit computations ofboth motion vectors and error criterion functional values at subpixelresolutions.