Three-dimensional overlapped spatial transformations for motion compensation

We propose the application of overlapped spatial transformation motion estimation with a search strategy that extends motion vectors to past and future frames in a framework of three-dimensional overlapped warping prediction. This arrangement combines the prediction efficiency of multihypothesis motion compensation in the spatial and temporal domains while allowing some degree of compensation for occlusion that occurs in subblock areas. Experimental tests show that multi-frame reference without overlap can achieve a prediction PSNR improvement of more than 1.5 dB over the conventional single-frame reference non-overlap technique. Considering the three-dimensional overlapped approach with multiple reference frames, a prediction PSNR improvement of up to 2 dB can be obtained. A practical implementation with one past and one future frame can achieve a gain of more than 1.5 dB or an equivalent 30% reduction in the mean-square error signal.