Re: [Openexr-devel] newbie questions: alpha-only files and linear conversion on save

[Paul Schneider]

Also, don't forget that, according to the EXR spec, the RGB channels in 
an EXR are premultiplied by the alpha channel (if one exists), and this 
multiplication happened on linear RGB values.  If you're going to do 
image processing on 8-bit color channels, you'll probably be 
gamma-correcting them and then operating on the gamma-corrected pixels. 
 If you're going to do this, you'll want to disassociate the alpha from 
the color channels first (i.e. unmult them).  This is what the 
Photoshop plugin does.

So, lets say your application produces an 8-bit RGBA premultiplied 
color image, where the RGB channels are gamma corrected and the alpha 
channel is not.  If you want to save this buffer to an EXR, you'll need 
to:

1) unmult the image
2) linearize the color channels
3) premultiply the linearized color channels with the alpha channel
4) save those values to the EXR

You'd want to promote your buffer to half or float before doing any of 
this, so as not to lose data when linearizing.  If your application 
produces a disassociated 8-bit RGBA image (aka straight, unmatted, 
etc), you'd do the same thing, but skip step 1.  To read an EXR into 
such an 8-bit pipeline, reverse the steps.

For an alpha-only file, I would expect the alpha to already be linear, 
but as Florian says, it depends on where you got that channel from.  
Generally, alpha channels are linear, and since color fidelity and high 
dynamic range aren't (usually) issues, 8 bits are often enough.

- Paul


On Jan 20, 2005, at 12:04 PM, Florian Kainz wrote:

> 1) Using the RGBA-only interface, one way to write an alpha-only
>    image is to open the file like this:
>
>         RgbaOutputFile file (fileName, width, height, WRITE_A);
>
>    With the general (arbitrary set of channels) interface, you can
>    open an alpha-only file like this:
>
>         Header header (width, height);
>         header.channels().insert ("A", Channel (HALF));
>         OutputFile file (fileName, header);
>
> 2) 8-bit (0-255) pixel data are typically not linear in the sense
>    that the amount of light represented by the pixel is proportional
>    to the number stored in the pixel.  8-bit pixel data are usually
>    gamma-corrected; the amount of light represented by  a pixel with
>    value v is roughly proportional to pow(v,gamma), where gamma is
>    about 2.2.  (For details see Charle's Poynton's Gamma FAQ at
>    http://www.poynton.com/GammaFAQ.html.)
>
>    OpenEXR expects pixel data to be linear.  In order to convert
>    8-bit pixels to OpenEXR's "half" you should linearize the data:
>
>         half h = pow (v, 1/gamma) / 255.0;
>
>    Note that this may apply only to RGB data.  In many file formats
>    alpha is linear.  The opacity represented by the alpha channel
>    is proportional to the 8-bit value stored in the pixel.
>
>
> Paul Miller wrote:
> > Hi guys. I'm adding EXR support to our up-coming cross-platform 
> > rotoscoping system (Silhouette Roto).
> > I have two questions:
> > 1) what is the proper way to set up the header for an "alpha-only" 
> > image?
> > 2) when saving generic linear color data (ie. 0-255), is it 
> > sufficient to just normalize to 0-1 and stuff the components into the 
> > output half types? This seems to work ok when reading the image back 
> > in, as long as I adjust the gamma to 1.0. If I should apply the 
> > output gamma correction, can someone post a code snippet?
> > Thanks!
> > _______________________________________________
> > Openexr-devel mailing list
> > address@hidden
> > http://lists.nongnu.org/mailman/listinfo/openexr-devel
>
>
>
>
>
> _______________________________________________
> Openexr-devel mailing list
> address@hidden
> http://lists.nongnu.org/mailman/listinfo/openexr-devel