Re: [SIP-devel] new interf_demo + pbs

[Ricardo F. <rf...@if...>]

On Sat, 21 Jun 2003, DRUEL Jocelyn wrote:

> 2) 2 functions to take pictures from a video camera using xawtv. We use a
> cheap miroPCTV for this (about 50$). This works on linux only of course.

Cool. I suggest that you check if the xawtv command is available
before calling it. But this is not very important.

> 3) modifications to use gray_imread:
> I replaced many gray_imread by gray_imread/2555 because I had problems with
> displaying or tracing profiles. I don't exactly know why.

This is due to the fact that you read your .jpg images, which are
pseudo-color (indexed). What you probably did before imread was to use
the index of the image, which points to 256 different colors in the
colormap. The colors in the colormap goes from 0 to 65535, so you must
divide by __257__ to get the 0-255 range (65535 / 257 == 255).

>
> The gray_imread function is much much slower than imread when I read
> grayscale pictures. A test on my box gave: 2.4 and 0.26 (estimated by
> timer()). That's a pb for me (specially as I read several images to make some
> calculations).
>
gray_imread (because of im2gray) is really very slow, I haven't
noticed that before.  I optimized im2gray to eliminate a for loop (see
attachment). gray_imread is much faster now. In scilab any loop over
an image is prohibitive :-((. It is ok just for "thumbnail-sized"
images, when we need to prototype an algorithm.

I could implement im2gray in C, but the trick I done in gray_imread
saved me some development time. But if it is still slow, please tell me.

> Why not modify the gray_imread so that with a grayscale image, it does simply
> an imread ? Does it cause pbs ?

Yes, it causes problems, because there is no fast way (in scilab) to
know if a pseudo-color (indexed) image is grayscale or color. It
depends on the colormap. Even if it is grayscale, the intensity
distribution in the colormap may be different from the numbers in the
index image.  Although this is not common, may happen.

What you've been doing is using directly the index image and assume
that the colormap has a linear intensity distribution, so it can be
discarded. This will work only with your particular images, but may
fail with other images.

best regards,
Ricardo.

--
  Ricardo Fabbri, Cybernetic Vision Research Group, USP, Brazil.
  http://cyvision.if.sc.usp.br/~rfabbri