Full text (gzip-compressed postscript)
We compare methods for choosing motion vectors for motion-compensated
video compression. Our primary focus is on videophone and
videoconferencing applications, where very low bit rates are necessary,
where the motion is usually limited, and where the frames must be coded
in the order they are generated. We provide evidence, using established
benchmark videos of this type, that choosing motion vectors to minimize
codelength subject to (implicit) constraints on quality yields
substantially better rate-distortion tradeoffs than minimizing notions
of prediction error. We illustrate this point using an algorithm within
the
standard. We show that using quadtrees to code the motion
vectors in conjunction with explicit codelength minimization yields
further improvement. We describe a
dynamic-programming algorithm for choosing a quadtree to minimize the
codelength.