Extension of the forward-backward motion compensation scheme for MPEG coded sequences: a sub-space approach

A different predictive view for the efficient coding of a given video. is proposed in this paper. Each image in a sequence can be seen as a vector in a hyperspace and the whole video as a curve, each point of the curve representing a given frame in the video. The whole video can be reconstructed from its video-samples: any image in the hyperspace can be obtained by means of a reconstruction algorithm, in analogy with the reconstruction of an analog signal from its samples. For this aim an appropriate interpolating kernel function should be used due to the multi-dimensional nature of the problem. To obtain the predicted image only the description of its position in the vector hyperspace is needed, i.e. by the vector of distances of the image at hand and the selected video samples (key-frames). A video key-frames codebook is used to synthesize video; it should allow the reconstruction of a good quality prediction of each image to be transmitted with poor requirements on the side motion information. Coding information reduces to the position in the hyperspace of the image to be predicted, the construction of the predicted image to the problem of the efficient multidimensional interpolation of key-frames. The focus of this paper is on the analysis phase of a given video sequence. Preliminary results are presented.