Interpolation for Motion Compensated Prediction

In motion compensated prediction (MCP) the current image is predicted from already reconstructed images and the prediction error is coded along with the motion information. MCP using fractional sample accurate displacements is an established technique in video coding. When the accessed position does not fall on the integer-sample grid, it is interpolated using the neighboring samples. In the H.264/AVC standard, a 2D separable architecture is employed, in which the horizontal and vertical filtering stages are performed sequentially, depending on the required fractional position.

Spline-Based Motion Compensated Prediction

An ideal filter for fractional displacement of a reference signal retains all frequencies and introduces constant phase shift for all frequencies. However, due to various effects in signal acquisition and noise, the efficiency of MCP is limited. In this work, we propose a spline based framework for MCP. The output pictures are first converted into spline coefficients using a prefilter and stored in the reference picture buffer. Then, during the MCP stage, fractional displacements are generated using interpolation filters with short support. A detailed analysis of the coding efficiency, complexity and memory bandwidth is provided. Average bit-rate savings of around 7 - 8% compared to the scheme in H.264/AVC and around 2 - 3% compared to the scheme in H.265 / MPEG-HEVC are observed.

Conditional Motion Vector Refinement

In this work, adapting the resolution of MCP in a video codec is considered. We increase the number of MCP filters to support 1/6th sample accuracy but control the motion search in a way such that the number of points to be tested is the same as 1/4th sample accuracy. In order to reduce the signaling overhead, common information available to both encoder and decoder, e.g. current slice type, texture in reference pictures, etc., are used as a cue to control motion vector accuracy. An efficient reuse of motion information for predicting subsequent blocks further reduces the signaling overhead. An average bit rate reduction of around 2.5% for P pictures and 0.5% for B pictures in a H.265 / MPEG-HEVC framework is observed with no extra search compared to a fixed quarter-sample resolution.

Adaptive Motion Modeling

In this investigation, a block based video coder that supports multiple motion models is proposed. Apart from the typical translational motion model, we employ parametric models to more accurately represent complex motions that occur in video sequences. A quad-tree based partitioning scheme, where the encoder selects a motion model to minimize the RD cost for each block, is employed. A cubic spline framework is utilized to obtain samples at fractional positions according to the selected motion model. Efficient motion vector prediction schemes are developed to maintain the continuity of the predictor in spite of different motion models.