Motion compensation with minimal residue dispersion matching criteria

With the ever growing demand for video services, video compression techniques have become a technology of central importance for communication systems. Industry standards for video coding have emerged, allowing the integration between these services and the most diverse devices to access them. The almost entirety of these standards adopt a hybrid coding model combining di erential and transform coding methods, with block-based motion compensation (BMC) at the core of its prediction step. The BMC method have become the single most important technique to exploit the strong temporal redundancy typical of most video sequences. In fact, much of the improvements in video coding e ciency over the past two decades can be attributed to incremental re nements to the BMC technique. In this work, we propose another such re nement. A key issue to the BMC framework is motion estimation (ME), i.e., the selection of appropriate motion vectors (MV). Coding standards tend to strictly regulate the coding syntax and decoding processes for MV's and residual information, but the ME algorithm itself is left at the discretion of the codec designers. However, though virtually any MV selection criterion will allow for correct decoding, judicious MV selection is critical to the overall codec performance, providing the encoder with a competitive edge in the market. Most ME algorithms rely on the minimization of a cost function for the candidate prediction blocks given a target block, usually the sum of absolute di erences (SAD) or the sum of squared di erences (SSD). The minimization of any of these cost functions will select the prediction that results in the smallest residual, each in a di erent but well de ned sense. In this work, we show that the prediction of minimal residue dispersion is frequently more e cient than the usual prediction of minimal residue size. As proof of concept, we propose the double matching criterion algorithm (DMCA), a simple two-pass algorithm to exploit both of these MV selection criteria in turns. Dispersion minimizing and size minimizing predictions are carried out independently. The encoder then compares these predictions in terms of rate-distortion performance and outputs only the most e cient one. For the dispersion minimizing pass of the DMCA, we also propose the total absolute deviation from the mean (TADM) as the measure of residue dispersion to be minimized in ME. The usual SAD is used as the ME cost function in the size minimizing pass. The DMCA with SAD/TADM was implemented in a modi ed version of the JM reference software encoder for the widely popular H.264/AVC coding standard. Absolute compliance to the standard was maintained, so that no modi cation on the decoder side were necessary. Results show signi cant improvements over the unmodi ed H.264/AVC encoder.