Re: [help-GIFT] A layman experiments on GIFT

[I. Wronsky]

On Thu, 6 Jun 2002, MUELLER Henning wrote:

Hi all,

> Is the database freely available that you did the tests with? 

Unfortunately, no. It would probably present an equivalent 
amount of copyright problems as there are images in it. :( 
I'm sorry. 

> As David pointed out, gift is only using global and local color and
> texture characteristics. There is no cheating with using filenames or
> scanner signatures for the queries and no object or shape features are
> used, either.

I'm ready to believe that. My previous message was sort of
a humorous "thought experiment" of a hypothetical "what if" 
case, where it was presumed that there were an individual 
unaware of the difficulties in image retrieval and who'd find
gift from e.g. freshmeat, and expect it to give useful
results on some particular style of query/data combination. ;)

> You can execute the queries with using exclusively one of the four
> feature groups or any combination to find out more about the origin of
> the similarity scores.

Does this mean the php gui? The java applet gui (which,
when I last looked, was claimed to be most complete <grin>), 
however, didn't allow changing those. Using php on redhat seemed
to require recompilation of apache, and I didn't bother to do
that. :(

> We would, of course, be happy to find developers for features such as
> shapes or object features based on segmentation. The features that are
> currently used are relatively simple and we know that they have
> shortcomings for certain query tasks.

Some months ago I spent some time looking for open source
cbir/computer_vision/pattern_recog/segmentation stuff from 
the net, but it seems that openness hasn't made a breakthrough 
on these areas. Instead of code, there's a lot of obsoleted binaries, 
plain hype or perhaps a web demo with a limited database. On 
the other hand, there's a lot of technical papers available 
about the subjects. When ones knowledge and intuition of the 
various proposals is quite limited, and the ideas do not seem
to be adequately ranked, its hard to choose which of these 
methods would be worthwhile for e.g. gift implementation.

While in theory I'd be interested to code something, betting
on a "wrong" or a "worthless" horse (idea) would cost several 
days of work, at the minimum. As I understand it, the problem
has not been solved, but I assume that after several decades 
of research there are better and worse methods. Which is which, 
though, is not clear to me. Each paper is trying to sell
its own approach and gloss over the deficiencies.

Oh well. Perhaps someday. 


Cheers,

 - Igor