Electrical Engineering
      and Computer Sciences

Electrical Engineering and Computer Sciences

COLLEGE OF ENGINEERING

UC Berkeley

How Much Does Globalization Help Segmentation?

Charless Fowlkes and Jitendra Malik

EECS Department
University of California, Berkeley
Technical Report No. UCB/CSD-04-1340
July 2004

http://www.eecs.berkeley.edu/Pubs/TechRpts/2004/CSD-04-1340.pdf

This paper quantifies the information gained in integrating local measurements using spectral graph partitioning. We employ a large dataset of manually segmented images in order to learn an optimal affinity function between nearby pairs of pixels. Region cues are computed as the similarity in brightness, color, and texture between image patches. Boundary cues are incorporated by looking for the presence of an "intervening contour", a large gradient along a straight line connecting two pixels. We then use spectral clustering to find an approximate minimizer of the normalized cut, partitioning the image into coherent segments.

We evaluate the power of local measurements and global segmentations in predicting the location of image boundaries by computing the precision and recall with respect to the human groundtruth data. The results show that spectral clustering is successful in suppressing noise and boosting weak signals over a wide variety of natural images.


BibTeX citation:

@techreport{Fowlkes:CSD-04-1340,
    Author = {Fowlkes, Charless and Malik, Jitendra},
    Title = {How Much Does Globalization Help Segmentation?},
    Institution = {EECS Department, University of California, Berkeley},
    Year = {2004},
    Month = {Jul},
    URL = {http://www.eecs.berkeley.edu/Pubs/TechRpts/2004/5468.html},
    Number = {UCB/CSD-04-1340},
    Abstract = {This paper quantifies the information gained in integrating local measurements using spectral graph partitioning. We employ a large dataset of manually segmented images in order to learn an optimal affinity function between nearby pairs of pixels. Region cues are computed as the similarity in brightness, color, and texture between image patches. Boundary cues are incorporated by looking for the presence of an "intervening contour", a large gradient along a straight line connecting two pixels. We then use spectral clustering to find an approximate minimizer of the normalized cut, partitioning the image into coherent segments. <p> We evaluate the power of local measurements and global segmentations in predicting the location of image boundaries by computing the precision and recall with respect to the human groundtruth data. The results show that spectral clustering is successful in suppressing noise and boosting weak signals over a wide variety of natural images.}
}

EndNote citation:

%0 Report
%A Fowlkes, Charless
%A Malik, Jitendra
%T How Much Does Globalization Help Segmentation?
%I EECS Department, University of California, Berkeley
%D 2004
%@ UCB/CSD-04-1340
%U http://www.eecs.berkeley.edu/Pubs/TechRpts/2004/5468.html
%F Fowlkes:CSD-04-1340