Re: [help-GIFT] Asymmetric Similarity Matrix

[Henning Müller]

Hi,

this is true for the simple tf/idf weighting that is usually used in 
GIFT but it will not work out for all weightings as some of them take 
into account the term frequency in the query itself and not only in the 
documents and in this case it will not be symmetric. If you use separate 
normalization this will not work either as you have to normalize for it. 
The separate normalisation has much better results as otherwise small 
scale textures can become dominant.

Cheers, Henning

Juan C. Caicedo a écrit :
> We were working on analysing both, papers and code. We wonder if the 
> similarity s(x,y) between two images x and y should be the same as 
> s(y,x). And of course it is. We found that the asymmetry is a 
> normalisation effect, in which the results list is scaled by the score 
> of the query applied on itself. So, when we obtained the scores 
> normalised using the auto-score of the image x, they are slightly 
> different to the scores normalised using the auto-score of the image y.
>
> We just modify the source file CQInvertedFile.cc to avoid this 
> normalization, and now we obtain a symmetric similarity matrix.
>
> Thank you very much for your response.
>
> Juan C. Caicedo
>
>
> On Thu, Apr 16, 2009 at 4:48 AM, Henning Müller 
> <address@hidden <mailto:address@hidden>> wrote:
>
>     Indeed, the paper:
>     Tversky, A. (1977). Features of similarity. Psychological Review,
>     84(4), 327-352.
>     shows through epxeriments that our visual similarity percetion does
>     not at all correspond to a metric.
>
>     Cheers, Henning
>
>
>     Wolfgang Müller a écrit :
>
>         I think the Squire et al. papers from 1999 accessible from the
>         Viper site in Geneva cite a paper of Tversky's which justifies
>         asymmetric similarity matrices: What you are looking for
>         influences your notion of similarity.
>         Cheers,
>         Wolfgang
>
>         On Thu, Apr 16, 2009 at 8:14 AM, Henning Müller
>         <address@hidden
>         <mailto:address@hidden>
>         <mailto:address@hidden
>         <mailto:address@hidden>>> wrote:
>
>            Dear Juan,
>
>            this is normal as the similarity measure of GIFT is not a
>         metric as
>            it is based on a tf/idf weighting form text retrieval.
>            Image similarity is calculated in the space spanned by the
>         features
>            present in the query, only. Each image has around 1500
>         features out
>            of 87000 possible (most of them binary features), so each image
>            potentially spans a different sub-space in which similarity is
>            calculated.
>
>            Cheers, Henning
>
>            Juan C. Caicedo a écrit :
>
>                Hello everybody,
>
>                We are building a similarity matrix of an image
>         collection using
>                GIFT. However, we notice that this matrix is not a
>         symmetric one.
>                Could anybody tells us what is the reason of this
>         behaviour and
>                some hints to obtain a symmetric similarity matrix?
>
>                Thanks in advance to all you.
>
>                Juan C. Caicedo
>
>
>              
>          ------------------------------------------------------------------------
>
>                _______________________________________________
>                help-GIFT mailing list
>                address@hidden <mailto:address@hidden>
>         <mailto:address@hidden <mailto:address@hidden>>
>
>                http://lists.gnu.org/mailman/listinfo/help-gift
>
>
>
>
>            _______________________________________________
>            help-GIFT mailing list
>            address@hidden <mailto:address@hidden>
>         <mailto:address@hidden <mailto:address@hidden>>
>
>            http://lists.gnu.org/mailman/listinfo/help-gift
>
>
>
>         ------------------------------------------------------------------------
>
>         _______________________________________________
>         help-GIFT mailing list
>         address@hidden <mailto:address@hidden>
>         http://lists.gnu.org/mailman/listinfo/help-gift