Newsgroups: comp.compression,sci.image.processing
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!oitnews.harvard.edu!news.dfci.harvard.edu!camelot.ccs.neu.edu!chaos.dac.neu.edu!usenet.eel.ufl.edu!news.mathworks.com!udel!news.sprintlink.net!howland.reston.ans.net!swrinde!sdd.hp.com!hplabs!hplntx!bhaskara
From: bhaskara@fc.hp.com (Vasudev Bhaskara)
Subject: Re: Motion Compensated Prediction Error/DCT
Sender: news@hpl.hp.com (HPLabs Usenet Login)
Message-ID: <D9vB04.Jr4@hpl.hp.com>
Date: Thu, 8 Jun 1995 18:50:28 GMT
References: <3qoi5q$e5h@net.auckland.ac.nz>
Nntp-Posting-Host: hplvab.hpl.hp.com
Organization: Hewlett-Packard Laboratories, Palo Alto, CA
X-Newsreader: TIN [version 1.2 PL2]
Followup-To: comp.compression,sci.image.processing
Lines: 77
Xref: glinda.oz.cs.cmu.edu comp.compression:20571 sci.image.processing:15114

Paul Wareham (wareham@eleceng.ee.queensu.ca) wrote:




: Hi I have a few questions regarding the MCPE used in interframe DPCM
: coding.  It seems that current day hybrid codecs tend to use the DCT
: to transform the MCPE before quantization.  Apparantly the concept is
: to decorrelate the MCPE signal.  However, it is well know(I think) that
: the statistical properties of the MCPE signal is quite different than that
: of the input image.  In fact the degree of inter-pixel correlation is quite
: low.
:  
: Given this, why do these codecs(MPEG, H.261, etc.) use the DCT in the
: interframe loop?  It seems that DCT would not be effective for coding
: almost uncorrelated inputs.
  MCPE will have a lower correlation and so your observation that a DCT is
  not efficient is correct. However, what MCP buys you is a coder where you
  can determine if it is better to code the block using MCP or intra. 
  In MPEG/H.261 MCP can be coded without any DCT (just use motion-vectors)
  or with DCT or the algorithm can determine if it is better to code the
  block in intra-mode (i.e. use DCT but not on MCP).
:  
: Exactly what effects the degree of spatial correlation in the MCPE?  I
: assume that the remaining correlation in a motion compensated frame would
: less than an uncompensated one.
:  
: Can one one say that the better the motion estimation - the lower degree of
: correlation is left in the frame?  That is, if the motion compensation is
: doing its job then the MCPE should not be spatially correlated?

  Yes. But the problem with a block based motion-estimation scheme and the
  translatory motion-model is that very good motion estimation is hard to
  come by within the MPEG/H.261 scheme.   
:  
: How does the type of motion compensation/prediction effect the correlation
: of the MCPE?
  Better motion-estimation will lower the correlation in MCPE; going to a
  general motion-model with more degrees of freedom will help but the standards
  don't support such things. Lowering noise in frames (by preprocessing) will
  also help out since true motion vectors can be found. 
  In an information-theoretic sense, for some assumed model of the video 
  source (say, isotropic covariance model) and model of motion-estimation
  uncertainity, it is possible to derive a rate-distortion function for
  the MCP case as well as the intra case and this can be parametrized by
  the variance of the motion-estimation error and perhaps this may answer your
  question concerning the efficacy of MCP.
  Bernd Girod's paper
  "The efficiency of motion-compensated prediction for hybrid coding",
  journal on selected articles in communication (IEEE), Aug. 1987,
  vol. SAC-5, no. 7, pages 1140-1154" addresses this.
  Another reference is
  Chi-Fa Chen and K.K. Pang "Hybrid coders with motion compensation",
  in Journal Multidimensional Systems and Signal Processing, Vol. 3,
  No. 12, Dec. 1992, pages 241-266.
 
:  
: And finally, what type of coding scheme might be more efficient for the MCPE
: and what type of statistical model would one use for the MCPE(i.e. Gaussian,
: Laplacian, ?)?
  Perhaps a generalized gaussian function is a better fit. 
:  
: Perhaps I'm getting mixed up somewhere in the spatial and temporal domains . . .
:  
: Regards,
:  
: -------------------------------------------------------------------------------
: Paul Wareham
: Image Processing and Communications Lab
: Dept. of Electrical and Computer Engineering             Kingston, Ontario
: Queen's University                                       K7L 3N6
: E-Mail:wareham@vision.ee.queensu.ca                      (613) 531-0660
: http://ipcl.ee.queensu.ca/wareham/wareham.html
: -------------------------------------------------------------------------------
 - Bhaskaran Vasudev
   HP Labs, Palo Alto
   email: bhaskara@hpl.hp.com
