Re: [help-GIFT] Understanding MRML

[Ralf Juengling]

On Mon, 2002-07-01 at 07:15, Wolfgang Müller wrote:
> Good to meet you again :-)

thanks, nice to meet you again here :)


> > query engine/algorithm, the client respectively). Thus, as I understand
> > it, the legal values of the "type" attribute should be specified by the
> > MRML protocol.
> 
> Yes. And I find this is typically something where you need several people. Up 
> in the list there are some discussions between Pruet and me, in which we 
> write something about that topic. It could be the base to our discussion. And 
> yes, this should definitely go to MRML net.

Okay, let's jump into and start the discussion. At the moment, I see
two and a half query-paradigms: ;)

1. Query by keyword (qbk)
2. Query by example (qbe)
3. Posing constrains to define subsets of the image collection
   (constraint query, cq)

The latter might need some explanation: With constrain I mean a
simple proposition about an image which can be "decided" by an
algorithm, e.g.:

The image belongs to a class "indoor"
The image is larger than 320x320 pixels
The image is a grayscale image
The annotation of the image containts the word "america".

Constraints must be combined logicaly. 
(For now, there is no means to communicate constraints via MRML, 
I'll make a proposition if you agree it's important.)

When you think about it, you can imagine to combine (or "apply")
all these "query-paradigms" within on query session. Thus, every
"query-step" element should contain an element which somehow
denotes the query-paradigm it belongs to. For instance, there is
a "cui-text-query" element, which surely represents a qbk.
On the other hand, there is no "query-by-example" element, but
a "user-relevance-element-list" element. Here, I think, the real
trouble begins: 
How can we specify the treatment of relevance feedback in an 
extensible and -- if that's reasonable -- query-paradigm-independent 
way? 
I'm currently aware of two schemes to request and treat user 
relevance information: 
1. the "classical" scheme where the user judges the current best
   results (this is implemented in the GIFT)
2. the "active learning" scheme, where images to judge about a 
   carefully selected by the learning algorithm. 

(Do you know more?).


Cheers,
Ralf


-- 
-------------------------------------------------------------------------
Ralf Jüngling
Institut für Informatik - Lehrstuhl f. Mustererkennung &
Bildverarbeitung
Georges-Köhler-Allee
Gebäude 52                                       Tel:
+49-(0)761-203-8215
79110 Freiburg                                   Fax:
+49-(0)761-203-8262
-------------------------------------------------------------------------