Fwd: [imgseek] Tips on color vs. shape adjustment

[Ricardo Niederberger Cabral]

Sorry for not being able to reply you earlier and thanks for your
interest on imgSeek.

At imgdb.cpp (https://imgseek.svn.sourceforge.net/svnroot/imgseek/trunk/net.imgseek.imgdb/src/net/imgseek/imgdb/core/imgdb.cpp
)
take a look at std::vector<double> queryImgDataFiltered()

For every computation done when comparing images, a final similarity
ratio is achieved by summing all wavelet coefficient indices where two
images are similar to each other, and this summation happens over the
three YIQ color components (the c variable on those loops, going from
0 to 2). 0 stands for the Y channel, 1 for I and 2 for Q.

The Y component represents the luma information. I and Q represent the
chrominance information. You may do some experiments by giving a
higher weight to agreements on the I and Q channels of each image
(when c=1,2) and see if color information becomes a stronger
discriminant between images.

Tweaking const float weights[2][6][3] at https://imgseek.svn.sourceforge.net/svnroot/imgseek/trunk/net.imgseek.imgdb/src/net/imgseek/imgdb/core/imgdb.h
may also achieve the same result explained above.

Anyway, it's up to you for doing some experiments.

Hope this extra info helps and sorry for not being more detailed,

--
Ricardo Niederberger Cabral
http://www.imgseek.net/
ricardo.cabral at imgseek.net
skype: rnc000

> On 13/05/2008, at 03:50, Cédric DELTHEIL wrote:
>>
>>> Hello Ricardo,
>>>
>>> I'm Cédric, a french developer.
>>>
>>> First of all congratulations for your work on imgSeek which is a
>>> nice software!
>>>
>>> I previously tried to contact you through the server at imgseek
>>> dot net email address regarding the ability to tune the importance
>>> of color vs. shape for the similarity computation. I saw that this
>>> possibility is pointed out into the FAQ :
>>>
>>> ===
>>>
>>> What is the order of image matching within the multiresolution
>>> wavelet decomposition (colour, shape…)?
>>>
>>> All these factors are taken into account at the same time, but the
>>> importance of shape versus the importance of colour could be fine
>>> adjusted. This option is to be added to the API on the next
>>> versions.
>>>
>>> ===
>>>
>>> I wanted to know if you could give me a tip on how and where to
>>> increase the color factor in the source code ? I currently work
>>> with the 0.6.2 distribution.
>>>
>>> Thank you very much for your help!
>>>
>>> Regards,
>>>
>>> Cédric
>>>
>>
>