help-gift
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[help-GIFT] Searching for similarities


From: Andreas Enge
Subject: [help-GIFT] Searching for similarities
Date: Wed, 17 Oct 2001 09:36:48 +0200 (CEST)

Dear developers,

after playing around a little bit with gift, I realised that it does not
quite serve the purpose I had intended to use it for. Maybe I am just
overlooking something in the handling, maybe I am going to formulate a
feature wishlist. Let me present two scenarioes.

First, one might have a given image (say, of the Christmas tree of 2000)
and look for similar ones (Christmas trees from previous years) in the
indexed collection. Then it does not help too much to start with a
random choice and select tree like forms until the picture one has in
hand pops up. It would be nice to be able to specify one reference image
whose location on the disk one knows and to look for similar ones.
This is probably only a frontend issue and does not concern gift internals.

Second, one might wish to clean the hard drive from (almost) duplicate
images (the same image saved with different file formats, slightly
varying brightness, added or removed text, etc.). To do so, one would
need the list of the most resembling pairs of images in the collection,
and I am not sure if this is a problem easily solved with the techniques
used in gift. (Admittedly, I have no experience whatsoever with image
retrieval.) When looking for exact duplicates, the most efficient way
of doing so is to compute a hash value (md5sum) of each image, sorting
the list of hash values and looking for duplicates. This makes up for
a one-line shell script which is executed in basically linear time in
the number of images. To imitate this procedure when looking for
similar images, one would need to associate to each file some number so
that similar images get the same or close numbers. Is this a possibility,
or does gift only associate a degree of similarness to pairs of images?
This would result in a rather inefficient quadratic time algorithm to
compare everything with everything.

Thanks for your considerations,

Andreas


_______________________________________________________________________________

Andreas Enge
Laboratoire d'Informatique (LIX)
Ecole Polytechnique
91128 Palaiseau CEDEX
France

Phone:  +33 1 69 33 34 79
Fax:    +33 1 69 33 30 14
E-mail: address@hidden
URL:    http://www.math.uni-augsburg.de/~enge




reply via email to

[Prev in Thread] Current Thread [Next in Thread]