12-12-2012, 05:10 PM
Limitations of Contentbased Image Retrieval
Limitations of Contentbased.pdf (Size: 502.31 KB / Downloads: 52)
The True Current State of the Art
• Title of Editorial in Special Issue of IEEE Proceedings (April 2008):
“The Holy Grail of Multimedia Information
Retrieval: So Close or Yet So Far Away?”
• Close if we take published results at face value.
• Far Away if we evaluate results from online test sites
or look closely at the published results.
• The answer also depends on what do we mean by
CBIR.
Two Kinds of CBIR
• General: We try to match a query image to an
arbitrary collection of images, such as those found on
the web. The goal of the query is to obtain images
with the same object as the query. Such CBIR
imitates web search engines for images rather than
for text.
– Given an image with a horse, find all images showing a
horse (at least as their main subject).
• Application Specific: We try to match a query image
to a collection of images of a specific type. For
example, fingerprints, X-ray images of a specific
organ, images of skin lesions, etc.
Results from Online Tests
of General CBIR
• Two types of tests are publicly available: image
retrieval and auto-tagging.
• Fewer sites are actually available than advertised.
– For example, Cortina (cited in a Nov. 2008 PAMI paper) is
not operational except for already tagged images.
• Results are generally poor. Only one site (GazoPa)
produced a good match and that was only once. See
[Appendix A] and [Recent Tests]. In some sites the
system failed to produce any results.
Human Discriminating Ability
• People seem to be able to discriminate 6x6
arrangements, so that 64 billion distinct
images seems a realistic lower bound.
• In the following two slides we show three
pairs of random patterns that differ in just one
location. Because of the lack of order this is
the most difficult case for discrimination.
• The second version of the patterns marks the
differences.