home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!usc!usc!not-for-mail
- From: haddadi@sipi.usc.edu (Navid Haddadi)
- Newsgroups: comp.compression
- Subject: Re: Motion Compensation?
- Date: 15 Dec 1992 04:55:11 -0800
- Organization: USC, Los Angeles, CA
- Lines: 47
- Sender: haddadi@sipi.usc.edu
- Message-ID: <1gkkjfINN7c6@sipi.usc.edu>
- References: <1992Dec14.161316.38402@ns1.cc.lehigh.edu>
- NNTP-Posting-Host: sipi.usc.edu
-
- In article <1992Dec14.161316.38402@ns1.cc.lehigh.edu> cre1@ns1.cc.lehigh.edu (CHRISTOPHER R. EMERSON) writes:
- >Can anyone tell me (is this in a FAQ?) in plain english, what motion
- >compensation is, how it is done and what it does for you? I've heard it
- >mentioned in many papers, but I guess everyone assumes that you know what they
- >are talking about. Any help would be appreciated. Thanks.
- >
- >Chris
- >
- I don't know if it is in the FAQ or not, but here is a simple explanation.
-
- Say you have a sequence of images E1,E2,E3,... of an object in motion
- in front of a stationary background. You want to transmit or store
- these images and want to compress them. One way is to compress each
- image individually. Of course, the idea behind compression is to use
- the correlation in data to generate a non-correlated sequence (i.e.
- energy compaction). It turns out that, there is a lot of temporal correlation
- (as compared to spatial correlation). However, classical predictive
- coding method perform very poorly when applied along a temporal line.
- This is because when an object moves it creates sharp discont along
- temporal lines. However, if the motion of the object can be represented by
- a 2-D vector field (motion field) then each frame can be compensated.
- That is, by using the motion field, we perform interpolation in each frame
- E2,E3,... to make them look similar to E1 (and hence similar to
- one another). Now predictive coding can be applied along temporal lines
- since we have eleminated the discont.'s. The same approach can be
- used for motion compensated noise reduction where after the compensation
- all frames are averaged in order to reduce noise without bluring the image.
- There are many, many applications for motion field like navigation,
- segmentation, shape from motion,....
-
- Now the trick is that how do you compute the motion field from
- the sequence E1,E2,...? This is the topic of active research here
- in USC as well as many other places. One simple way is to sustitude
- optical flow field for the motion field. This gives satisfactory
- result for motion compensation, but not so good a result for 3-D
- applications. There are well known algorithms for computation
- of optical flow (e.g. Horn and Schunck, Heeger, etc.). There are
- two basic problems known as the aprature problem and oversmoothing
- of the flow field. Some newer methods such as (Haddadi and Kuo)
- offer very good results for motions as large as 10 pixels per frame
- for moderate size images (256x256 to 512x512).
-
- I'll give more explanation of the computation if there is a general
- interest in this topic.
-
- Navid
- haddadi@sipi.usc.edu
-