Re: [help-GIFT] Processing multiple-example-queries

[David Squire]

Sailesh Suvarna wrote:
> Hello all,
>
> How does GIFT process multiple-example-queries? Does it combine the features to 
> form a single composite query?

There is no such thing as a "multiple example query", as the term is 
used by some at RMIT. I know that folk there will not agree with me, but 
I have read their publications in the area, and what they are describing 
is identical to what everyone else knows as relevance feedback 
mechanisms, the vast majority of which handle multiple feedback images 
(or indeed regions). They are creating confusion by using new 
terminology for a concept that has been around for decades.

There are two basic approaches to relevance feedback with multiple 
feedback images/regions:

	1. Merge features from the feedback images (with appropriate weighting) 
so that the query in some sense corresponds to a pseudo-image.
        2. Query with each example individually, and then merge results.

GIFT uses the first approach, as is described in the Viper publications. 
The approach is based on the decades-old Rocchio scheme. For details, 
see, for example:

Henning Müller, Wolfgang Müller, Stéphane Marchand-Maillet, Thierry Pun 
and David McG. Squire, Strategies for positive and negative relevance 
feedback in image retrieval, In Proceedings of the 15th International 
Conference on Pattern Recognition, Barcelona, Spain, September 3-8 2000.
http://www.csse.monash.edu.au/~davids/publications/postscript/2000/MuellerHMuellerWMarchandPunSquire_icpr2000.ps.gz

David McG. Squire, Wolfgang Müller, Henning Müller and Jilali Raki, 
Content-based query of image databases, inspirations from text 
retrieval: inverted files, frequency-based weights and relevance 
feedback, In The 11th Scandinavian Conference on Image Analysis 
(SCIA'99), pp. 143-149, Kangerlussuaq, Greenland, June 7-11 1999.
http://www.csse.monash.edu.au/~davids/publications/postscript/1999/SquireMuellerMuellerRaki_scia99.ps.gz


> > From some preliminary experiments, using single-example queries always gave 
> > better results than using 2 and 3-example queries in terms of Interpolated
> Recall-Precsion and using only the  Global Colour Histogram feature.

Which is what you would expect, since a merged global colour histogram 
is not likely to make a lot of sense unless the feedback images are 
*very* similar.

Regards,

David


--
Dr. David McG. Squire, Postgraduate Research Coordinator (Caulfield),
Computer Science and Software Engineering, Monash University, Australia
Monash Provider No. 00008C       http://www.csse.monash.edu.au/~davids/