Block Adaptive Quantization of Multiple Frame Motion Field

This paper presents a method for the adaptive quantization of the motion field obtained from multiple reference frames. The motion estimation is obtained through a block-matching technique, with the assumptions of pure translational motion and uniform motion within a block. The regular assumption of constant intensity along the motion trajectory in the spatio- temporal path is relaxed. On one hand, this allows for illumination changes between the reference frames and the frame to be motion-compensated. On the other hand, it allows for additional freedom to reduce the prediction error after motion-compensation when the translational and rigid body motion assumptions are violated. The matching function represents the difference for each block of the luminance signal in the frame to be encoded with respect to a linear combination of 2 displaced blocks of same size in the reference frames. The adaptive quantization mechanism is based on evaluating, on a block basis, the local sensitivity of the displaced frame difference signal to a quantization of the motion field parameters. It is shown how such sensitivity depends only on the reference frame signals, which allows to keep it below a desired threshold without additional information to be sent to a receiver. Simulations are carried out on standard CIF (240 X 360) source material provided to the ISO/MPEG. Results are discussed to show the improvement with respect to the strategy suggested in the draft recommendation of the ISO MPEG for interactive video at 1.5 Mbps.