Download files
Abstract
The persistent growth of video-based applications is heavily dependent on the advancements in video coding systems. Modern video codecs use the motion model itself to describe the geometric boundaries of moving objects in video sequences and thereby spend a significant portion of their bit rate refining the motion description in regions where motion discontinuities exist. This explicit communication of motion introduces redundancy, since some aspects of the motion can at least partially be inferred from the reference frames. In this thesis work, a novel bi-directional motion hints based prediction paradigm is proposed that moves away from the traditional redundant approach of careful partitioning around object boundaries by exploiting the spatial structure of the reference frames to infer appropriate boundaries for the intermediate ones.
Motion hint provide a global description of motion over specific domain. Fundamentally this is related to the segmentation of foreground from background regions where the foreground and background motions are the motion hints. The appealing thing about motion hints is that they are continuous and invertible, even though the observed motion field for a frame is discontinuous and non-invertible.
Experimental results show that at low bit rate applications, the motion hints based coder achieved a rate-distortion (RD) gain of 0.81 dB, or equivalently 13.38% savings in bit rate over the H.264/AVC reference. In a hybrid setting, this gain increased to 0.94 dB and 20.41% bit rebate is obtained. If both low and high bit rate scenarios are considered then the hybrid coder showed a RD performance of 0.80 dB, or equivalently 16.57% savings in bit rate. The usage of higher fractional pixel accurate motion hint, predictive coding of motion hint, a memory-based initialization for motion hint estimation improved the RD gain to 0.85 dB and 17.55% of bit rebate.
The prediction framework is highly flexible in the sense that the motion model order for the hints can be content adaptive i.e. it can accommodate different motion models like affine, elastic, etc. Detecting motion discontinuity macroblocks (MBs) is a challenging task and the prediction paradigm managed to detect a significant number of such MBs. If the motion hints based prediction is used as a prediction mode for MBs, at low bit rates almost 50% of the motion discontinuity MBs chose to use affine hint mode and this number increased to 60% if elastic hint is used.