Motion estimation using adaptive blocksize observation model and efficient multiscale regularization

Bayesian motion estimation requires two PDF models: observation model and motion field (prior) model. The optimization process for this method uses a sequential approach, e.g. simulated annealing. This paper proposes an adaptive blocksize observation model and multiscale regularization for the prior model and the optimization process. The purposes are to increase the speed and to improve the result. The proposed framework can initialize the Bayesian method. The result in this paper shows one of the possibility of its usage. Many strategies can be derived from this framework to work for itself or to support the Markov random field modeling for motion estimation.