Low-bandwidth dynamic aspect ratio region-based motion estimation

Abstract

In the domain of motion estimation based applications, in order to keep the bandwidth requirements low, the usage of multiple levels of memory hierarchy is a necessity. We analyze and optimize the two-level memory hierarchy system for motion estimation where the first level (L0 scratchpad) holds the search area of the estimator and the second level (L1 scratchpad) holds the region wherein the estimation is performed. In our system, the L1 scratchpad is reconfigurable and the aspect ratio of the region dynamically changes per video field. The aspect ratio is changed such that physically available region (L1) memory is maximally utilized. We extend this idea to the extreme case where the aspect ratio of the region changes from horizontal stripe to vertical column. This idea keeps the bandwidth requirements towards the off-chip image memory minimal, one access per pixel, regardless of the number of motion estimation scans. Further, switching the aspect ratio of the region to extreme values enables fast convergence of the motion estimator. To demonstrate our idea, experiments were performed on the test set of video sequences using the state-of-the-art de-interlacing as the application. The results are encouraging regarding both, objective quality metric as well as visual perception.