Re: [help-GIFT] Adding and removing images to/from gift

[Wolfgang Müller]

Hi again.

For the details of gift-add-collection.pl I would have to go into the code and 
currently I can't.

> - After getting a random set, I chose 3 images which i like, requery,
> select 3 more images gift found I could like. Now there is nothing more
> I like soi start "deselecting" the other 4 unwanted images, requeryand
> get 4 new images. But when I requery again, gift brings back the 4
> images i deselected before. After what I've seen this is not the gifts
> fault as such, it's more of a workflow problem because gift obviously
> doesn't know anymore what it server one request before. It would be nice
> if gift itself could track in the session, which images are queried
> wanted / unwanted and never show the unwanted in that session again. I
> think it could improve accuracy.

Well, we discussed this a lot in /199./ and /200[01]/ and came up with the 
thought that we rather want to add features like that in the client than in 
the query processor. 

In fact, from a HCI point of view it seems to be a good idea for the client to 
make the IR client a gathering space.

Check out 
http://www.dcs.gla.ac.uk/~jana/research/publications/ECIR2006.pdf
http://www.dcs.gla.ac.uk/~jana/research/publications.html

Clearly, things like that are far beyond the scope of what image retrieval the 
server should do (at least IMHO). Having some of this work to be done in the 
server and some in the client seems like double hassle to me. So you end up 
doing:

> The way it is currently working I would have to store the users query
> into a db, send the request to gift, store its answer and remerge
> something else out of the request I sent and the answer gift gave.

Which seems not so bad, after all.

> - The second thing is matching algorithm. Currently I get the feeling
> that my gift is fixed on color more than on shape / texture. sometimes I
> get perfect results ( e.g. gift finding all appropriate matches I know
> of ) but sometimes it doesn't even show an image it most definitely
> should be supposed to. How can I tweak my system to try different
> methods or make it more sensitive for shape / texture than color ?

I remember having written a document about configuring and hacking the gift, 
which is surely too short, but maybe a starting point. Have you tried the 
separate normalisation algorithm or which one did you use when running the 
GIFT?

Cheers,
Wolfgang
-- 
Dr. Wolfgang Müller
LS Medieninformatik
Universität Bamberg
Check out the SIG MM web site http://www.sigmm.org