Computing Reviews
Today's Issue Hot Topics Search Browse Recommended My Account Log In
Review Help
Search
Compensating for sub-pixel shift in motion estimation
Konstantoudakis K., Machairidou E., Papanikolaou G.  MOCO 2016 (Proceedings of the 3rd International Symposium on Movement and Computing, Thessaloniki, Greece, Jul 5-6, 2016)Article-No. 34.2016.Type:Proceedings
Date Reviewed: Sep 28 2016

Motion estimation in imagery is the process of locating a subimage in two images, a reference and a test image, and precisely determining the displacement. When done to subpixel accuracy, the process generally involves matching an unchanged subimage to a set of synthetic subimages produced by subpixel shifts of the reference, to determine the subpixel displacement that produces the best match. As the authors of the paper point out, this comparison of an unchanged subimage to an interpolated (and therefore blurred) one is both conceptually uncomfortable and measurably less than optimal.

Konstantoudakis and colleagues begin with the insight to treat both subimages, which they call blocks, being compared as, in some sense, equal. They do this by creating a third block that can be shifted (by less than one pixel in each dimension) to approximate both of the blocks to be compared with the minimum error. In other words, imagine some underlying image patch, called the implied block, for which the representation in both the reference and test images has been obtained by shifting. The implied block chosen is the one with minimum square error when shifting in both X and Y directions. Both blocks, in the reference image and the test image, are compared to this implied block, over varying shifts. The final displacement is measured from the best matches of both blocks. One subtle point of the method is that a comparison of blocks of size N by N requires an implied block that is N + 1 by N + 1.

The actual method does not require calculation of the implied block and a subsequent matching step. Instead, the solution of a nonlinear least-squares problem will produce both the implied block and the subpixel displacements. The solution of this equality is done numerically. To demonstrate the method, the authors operate on a set of standard image sequences, with the first frame as the reference. They measure the total error in matching the images and express the error as peak signal-to-noise ratio (PSNR); higher values imply a better match between the reference and test images. Their method using implied blocks showed higher PSNR, with the increase varying from 2 to 4 dB.

This method has interesting applications to not only motion estimation but more general uses of pattern matching. The paper is written at a basic level; while the work is easy to read and comprehend, it would be useful to go further, with more results and more discussion of the specific implementation issues. It is not obvious that mean-square error is the only appropriate choice for measurement of error; there is room for more exploration here. This paper presents a clever and relevant idea for subpixel image pattern matching, in a sparse and clear style.

Reviewer:  Creed Jones Review #: CR144793 (1612-0927)
Bookmark and Share
  Reviewer Selected
 
 
Pattern Recognition (I.5 )
 
 
Enhancement (I.4.3 )
 
Would you recommend this review?
yes
no
Other reviews under "Pattern Recognition": Date
Classification and learning using genetic algorithms: applications in bioinformatics and Web intelligence (Natural Computing Series)
Bandyopadhyay S., Pal S., Springer-Verlag New York, Inc., Secaucus, NJ, 2007.  311, Type: Book (9783540496069), Reviews: (1 of 2)
Oct 24 2007
Classification and learning using genetic algorithms: applications in bioinformatics and Web intelligence (Natural Computing Series)
Bandyopadhyay S., Pal S., Springer-Verlag New York, Inc., Secaucus, NJ, 2007.  311, Type: Book (9783540496069), Reviews: (2 of 2)
Feb 8 2008
Computational intelligence: concepts to implementations
Eberhart R., Shi Y., Morgan Kaufmann Publishers Inc., San Francisco, CA, 2007.  496, Type: Book (9781558607590)
Feb 22 2010
more...

E-Mail This Printer-Friendly
Send Your Comments
Contact Us
Reproduction in whole or in part without permission is prohibited.   Copyright 1999-2024 ThinkLoud®
Terms of Use
| Privacy Policy