Re: [help-GIFT] Effect of Normalization on Recall-Precision

[Henning Müller]

Hi,

I think that this normalization only makes sure that the similarity 
value for the pseudo-query image containing all feedback images is 1. 
This will get the similarity score for most images in between 0 and 1. 
Without normalization you get higher values but you should still get the 
exact same order of ranking, thus the same precision and recall values.

This is fairly different when using separate normalization of the four 
feature groups which does lead to different results as you can see in 
the following publication, where the two methods are compared with the 
GIFT system:
Henning   Müller, David   McG. Squire, Wolfgang  Müller, Thierry  Pun. 
Efficient access methods for content-based image retrieval with inverted 
files. In Multimedia Storage and Archiving Systems IV (VV02), Boston, 
Massachusetts, USA, 20--22~September, 1999 .

Cheers, Henning

Sailesh Suvarna wrote:

> Hi all,
>
>
>   I  have a set of Ranked results taking one ,two and three images as relevant. 
> Using the same queries for CIDF with normalization and without normalization ( 
> blocking-(*i).second/=fabs(lQueryScore);in  CQInvertedFile.cc) and then 
> calculating the Recall/Precision and Interpolated Recall Precision, I get the 
> same results for both (with and w/o    normalization). I have attached the .ps 
> graphs of these results.
>
>   Is the normalization supposed to have any effect on Recall and Precision?
>
> Sailesh Suvarna
>
>  
>
> ------------------------------------------------------------------------
>
> _______________________________________________
> help-GIFT mailing list
> address@hidden
> http://mail.gnu.org/mailman/listinfo/help-gift
>  

--
---------------------------------------
Dr. Henning Müller
University Hospitals of Geneva
Division of Medical Informatics
21, rue Micheli-du-Crest
CH-1211 Geneva 4, Switzerland
Tel  +41 22 372-6175
Fax  +41 22 372-8680
email address@hidden