An adaptive search range decision algorithm for parallel motion estimation

We introduce an adaptive search region decision algorithm to improve the speed of GPU-based motion estimation which generally adopts the full-search. The motion feature of the current block is estimated by MAP (Maximum A Posterior) estimation with the motion feature of the temporally adjacent block and its SR. In our MAP estimation, the prior distributions and the likelihoods are represented by the probabilistic modeling of the data extracted from the training set. The SR is adaptively decided depends on the estimated motion feature. In the motion estimation on GPU for an HEVC encoder, the proposed method with different change rates achieves 22.3% and 31.2% of average time reductions in the integer-pel ME with on average 0.25% and 1.08% BD-bitrate increases, respectively.