Forward/backward adaptive context selection with applications to motion vector field encoding

In low rate motion-compensated video coding, the rate required to encode the motion field can become a significant portion of the overall rate budget. This motivates us to investigate methods to efficiently, and losslessly, encode the motion field. Reductions in the required motion field bit rate allow increasing the rate for the residue images or may permit a higher density motion field to be used. We consider adaptive context modeling techniques, such as those proposed in image coding applications, and explore their effectiveness for coding the motion data. We rely on various forward/backward context selection algorithms, and demonstrate how forward adaptation, based on alphabet partitioning approaches, can result in performance improvements over purely backward adaptive methods. We observe substantial reductions in rate, especially for dense motion fields, when comparing with popular differential coding schemes using VLC tables, such as those in the H.263 standard.