Re: [Rtftr-devel] some ray of hope after using vigra

[lakesh <lak...@gm...>]

Hi Abhishek,

I was tuning the parameters of the Gabor filters and I think I got some
improvements. I changed the window size and the threshold and some other
parameters. There is slight improvement; however we can tune the parameters
to any thing we want that is we can overfit it to a specific image. Also
other thing is that we don't need to get the high energy points to the
specific locations like  eyes,nose mouth etc. I got some papers where they
have used adaboost for feature selection and the features are located at
random locations like us. Also in most of the papers they have taken
frequency as sqrt(2). The ebgm paper has also taken the same frequency. Also
I think while comparing the responses of the images to get the feature point
we should use directly the magnitude. I mean when we shift through the image
using the window of size WxW to compare the results, we should use the
magnitude of the response for the comparison and not directly the complex
numbers. I tried to compare the complex number and it considers only the
real part while doing the comparison. So I think it is not the correct way
to compare. Though in Kepenekci  it seems like they have directly compared
the complex numbers, I think we should use the magnitude values directly.

Lets sit and discuss the parameter selection when you finish the C++
implementation.
Currently I am laying my hands with everything I can  find and checking the
results.


On Sat, Dec 20, 2008 at 8:46 PM, Abhishek Dutta <the...@gm...>wrote:

> Hi lakesh,
> I decided to use "vigra<http://kogs-www.informatik.uni-hamburg.de/%7Ekoethe/vigra/>"
> library instead of GSL. GSL was not sufficient and did not fulfill all our
> requirements.
>
> "vigra" allows to generate gabor filter family by just supplying the no. of
> orientations, scale and center frequency. According to the "vigra"
> documentation, center frequency should not exceed 0.375. However, I was
> using sqrt(2) = 1.414 as frequency in matlab implementation. Moreover, it
> selects the best gaussian function (the sigma that we needed to supply)
> based on a mathematical formula (that I didn't understand). The output (at
> image size 128x128) looks good. But this does not assure better final
> results. I will update you with final results after I complete the code to
> compute feature vectors.
>
> bye
>
> --
> Abhishek
> http://adutta.np.googlepages.com
>



-- 
Have a nice day