Camera response functions as a tonemapping option

[Daniel90]

Like the title suggests I'm trying to implement camera response functions as a tonemapping option (like described in issue 632).
The problem is that the tonemappers that we have now (reinhard, linear, contrast and maxwhite) use the XYZ values to apply their algorithms to. The conversion from XYZ to RGB takes place later in the imaging pipeline (not in the tonemapper itself).
But the camera response functions have to be applied to the RGB values.
One option would be to convert XYZ -> RGB, apply the response function and convert back to XYZ. Then the imaging pipeline would again convert from XYZ -> RGB.

Because this is wasted computing time LordCrc suggested that the tonemappers themselves have to apply the conversion from XYZ to RGB. That requires not much change to the code and seems like a good solution to me.

Another suggestion of LordCrc is to apply the camera response function *after* (and additionally to) the tonemapping. But I don't know if this is correct. Does anybody know how Indigo uses CRFs? As a real tonemapping option or as an addition to the tonemapping?

[dougal2]

Indigo implements the CRF function as a tonemapper, however I suspect that in reality that tonemapper consists of:
[(XYZ) linear tonemapper (XYZ->XYZ) colourspace converter (XYZ->RGB) CRF processor (RGB)]

[jeanphi]

Hi,

Please don't reference commercial renderers when discussing algorithm implementation. I don't think this matters in this particular case, but better safe than sorry.
To the point, I think that currently the best thing would probably be to move XYZ->RGB to the tonemapping. They are already done just one after the other and will make it easier to add those fancy options. However you'll have to keep the current xyzpixels/rgbpixels in ApplyImagingPipeline to make it clear what data you have in the buffer. It will also require a protoype change for the tonemappers (have both an input buffer of XYZColor and an output buffer of RGBColor) so that you don't do fancy type casting in them too.

Jeanphi

[Daniel90]

I've got some examples for camera response function "corrected" pics
They're all tonemapped with the linear tonemapper at a sensitivity of 50, an exposure of 0.11, an f-stop of 16.5 and a gamma of 1.

The original image (no film used)

Advantix 100CD

Agfacolor OptimaII 100CD

Agfapan APX 100CD

F125CD

FP2900Z

Kodachrome 200CD

Portra 400NCCD

Some films show an obvious bluish or reddish tint, but that seems to be right. Maybe it is so obvious because the background of my test scene is so uniformly grey.

[tomb]

Very nice!

[SATtva]

Holy-moly! I'm in love. ^__^

[jeanphi]

Hi,

The gamma of the linear tonemapper should ALWAYS match the output gamma, because the ISO definition for digital images depends on the mapping of the gamma corrected image. So if you use a camera response, you should even factor that in the tonemapping conversion factor. The only other interpretation I can see is that the gamma should be hardcoded to 2.2, but not 1.

Jeanphi

[Daniel90]

argh, the above example images are already out of date

LordCrc said that it would be more plausible to apply the inverse gamma correction after and not before the crf.
Now also the bluish and reddish tints vanished (or at least got weaker; depending on the film used).
I'll post new example images later.

[jeanphi]

Hi,

Applying the gamma correction after tonemapping makes perfect sense as this is a display correction, tonemapping is a sensor correction.

Jeanphi

[Daniel90]

I revised my code and created new previews with the help of LordCrc's nice color chart scene.
In the top left you can see the original image without a CRF.

Results after applying some different CRFs