Adaptive direct vector derivation for video coding

This paper proposes a new method for improving direct prediction scheme that has been employed in conventional video coding standards such as AVC/H.264. We extend direct prediction concept to achieve better adaptation to local statistics of video source with the assumption of the use of larger motion blocks than conventional macroblock size. Firstly, our direct prediction method introduces block adaptive spatio-temporal estimation of direct motion vector, in order to compensate loss of accuracy of motion vector estimation for large motion blocks. This estimation method is performed without explicit signaling by employing decode-side collaborative decision. Then, adaptive selection of two reference pictures is performed to improve direct prediction efficiency where reliability of estimated direct motion vector is poor. Experimental results show the proposed method provides up to 3.3% bitrate saving and 1.5% in average in low-bitrate coding.