Motion Compensation by Adaptive Vector Quantization

We propose an efficient motion compensation scheme where adaptive vector quantization (AVQ) is used to optimally compress the dense motion field. The vector quantizer codebook is designed on-line by backward adaptation based on previous reconstructed frames. It thus takes advantage of the information already available at the receiver for adapting the codebook without recourse to side information. The method adapts the codebook to local motion statistics and thus exploits interframe motion correlations. It also eliminates the need for extensive prior codebook training, as well as problems of statistical mismatch due to scene variation. Preliminary experiments on benchmark video sequences show average improvement of about 0.8dB in SNR, and improved perceptual quality of the reconstructed video. The enhanced performance is achieved at the cost of only a fractional increase in complexity as compared to standard motion compensation methods.

Published

1996-10

Content type

Original Research

DOI

10.5594/M001238

ISBN

978-1-61482-947-8