Bias-minimizing filters for gradient-based motion estimation

Among the myriad of techniques used in estimating motion vector fields, perhaps the most popular and accurate methods are the so called gradient-based methods. A critical step in the gradient-based estimation process is the estimation of image gradients using derivative filters. It is well known that the gradient-based estimators contain significant deterministic bias related to the gradient calculation. In this paper, we describe the fundamental relationship between estimator bias and choice of derivative filters. From this, we propose an image adaptive method for designing bias-minimizing gradient filters. Simulations validate the superior performance of such filters for the many variants of gradient-based estimation including the widely used multiscale iterative methods.