Re: [imgSeek-devel] Percent

[Ricardo N. C. <rn...@gm...>]

Hi Plamen,

I got to this constant empirically by comparing several identical
images and adjusting the constant to reach 100%.

The core code changed a bit (huge optimizations) lately so this
constant may actually be a bit off.

If you come up with a better constant or a better way to convert the
signature difference into a human-understandable percentage, please
keep us posted.

Best regards,
--
rnc


On 2/27/06, Plamen Todorov <pla...@ya...> wrote:
>  Hello,
>
>  I am studying the source of imgSeek (really amazing program). I try to
> understand how it works. The core of it is based on the article "Fast
> Multiresolution Image Querying". The score of every image in the DB is
> updated from calculations with the queried image. Then the less the value=
 of
> the socre - the greater the similarity between the DB image and the query
> image.
>
>  What I miss is how the score is converted to percentage of similarity. I
> didn't find the answer in the research article. In the imgSeek source, th=
e
> conversion is:
>  percent =3D - 100 * score / 38.7
>
>  Where did the constant 38.7 come from? Is it accurate?  When I compare a=
n
> image with itself, the percentage varies with the pictures (not fixed to
> 100% as I expected).