Re: [help-GIFT] gift algorithms - separate normalization and CIDF

[David Squire]

Mika Rummukainen wrote:

> Hi there,
>
> I've been wondering for a while now for couple of questions.
>
> Question 1:
> How does GIFT perform the weighting of features with different algorithms?
>
> > From articles (Content-based query of image databases,inspirations from text
> retrieval: inverted files, frequency-based weights and relevance feedback
> and
> Content-based query of image databases: inspirations from text retrieval
> (Pattern Recognition Letters 21))
>
> I managed to find some information but I assume the equations how scores are
> calculated (after relevance feedback) for every image are for CIDF only. 
>
> Now I'd like to know how Separate Normalization performs its weighting. From
> "Strategies for positive and negative relevance feedback in image retrieval" I
> found a "Separately weighted feedback" - is this how separate normalization is
> performaing? 
>
> Question 2:
> This makes me think that when GIFT uses separate normalization, it first uses
> CIDF algorithm to weight the features and then uses separate normalization to
> weight even more the weighted features calculated by CIDF. Am I completely on a
> wrong path here?
>  
All that "separate normalization" means is that the similarity score is 
first calculated for each feature group separately, each feature group 
score is normalized, and then they are added to get the final score, e.g.

FinalSimilarity = 0.25*ColourHistogramSimilarity + 
0.25*ColourBlockSimilarity + 0.25*TextureHistogramSimilarity + 
0.25*TextureBlockSimilarity

where all of ColourHistogramSimilarity, ColourBlockSimilarity, 
TextureHistogramSimilarity and TextureBlockSimilarity have been 
normalized to the range [0,1]. This prevents feature groups that have 
more features completely dominating the total score.

> Question 3: 
> How is the similarity of images calculated with both algorithms if there would
> be no relevance feedback?
>  
All query similarities are calculated the same way, whether there is 
relevance feedback or not. A single image query simply has one relevant 
image.

Cheers,

David


--
Dr. David McG. Squire, Postgraduate Research Coordinator (Caulfield),
Computer Science and Software Engineering, Monash University, Australia
http://www.csse.monash.edu.au/~davids/