Decoder-Side Motion Vector Derivation for Hybrid Video Coding

In this thesis, a Decoder-Side Motion Vector Derivation (DMVD) algorithm is developed and studied. Based on the assumption of piecewise homogeneous motion, decoded signal samples in the immediate spatial neighbourhood of a block are used for a distortion-based motion estimation among the previously decoded frames. It is found that such a scheme cannot completely replace the explicit coding of motion parameters. Instead, signalling at the block level is used for selecting between explicit motion coding and decoder inference. The system design aspects of DMVD are discussed on the basis of a software implementation of H.264/AVC. A basic theoretical model of weighted multi-hypothesis prediction within the DMVD framework is given and confirmed by experiments. In order to address complexity concerns, a fast DMVD algorithm is developed, maintaining coding efficiency with significantly reduced processing requirements. Compression performance and complexity of the proposed system are evaluated in comparison to state-of-the-art H.264/AVC configurations for low-delay and high efficiency scenarios. Also, aspects of the resilience to transmission errors are briefly discussed. In response to the Call for Proposals (CfP) for HEVC, a coding technology proposal containing DMVD has been submitted, which has gained considerable interest in the video research community, as confirmed by the definition of a Tool Experiment for evaluating DMVD and related algorithms for potential inclusion into the upcoming HEVC standard.