USM threshold tools needs better documentation in rawpedia

[Beep6581]

I think something might be off with the USM threshold. At the maximal setting http://i.imgur.com/W5KFlrH.png it seems to work fine because the image looks horrible. But move the top-right point one step to the left and it results in a major image change http://i.imgur.com/LqU5rgk.png Reproduced on a 900x900px TIFF image and on amsterdam.pef

Top-right 2000: http://i.imgur.com/VFfxygL.jpg Top-right 1994: http://i.imgur.com/Vy4Sw7y.jpg

[sguyader]

Reproduced this bug on a Fujifilm X-T1 raw file.

[heckflosse]

hmm, this one : https://github.com/Beep6581/RawTherapee/blob/master/rtengine/procparams.h#L111 ?

[Hombre57]

This code is correct. You can even see that the line doesn't go back to the bottom when the top left slider is completely to the right.

[Hombre57]

I'd rather try to find out if the upper bound of 2000.f is correct in rtengine/ipsharpen.cc, line 267, 357, 1522 and 1615. Or maybe that the processing algorithm is wrong...

[heckflosse]

Hombre, I didn't say that the code is not correct. I just wanted to point out that it's a special case (as you commented in code). I'm still searching for the reason of the behaviour...

[Hombre57]

ok.

The upper bound should (to be confirmed) be the max-value of the image data. Maybe it has changed over the time and my algo doesn't reflect this change.

[heckflosse]

Hombre, don't waste your time on this. You know, I'll find the reason ;-)

[Hombre57]

I just wanted to give some hints, as I did that Threshold slider. Thanks for taking this issue, I know you'll find it, as usual :). I'm off to bed now, and will continue my patches tomorrow.

[heckflosse]

I think there is a bug in this two lines:

https://github.com/Beep6581/RawTherapee/blob/master/rtengine/procparams.h#L144 https://github.com/Beep6581/RawTherapee/blob/master/rtengine/procparams.h#L148

If we swap 2 and 3 it should work as expected, but not as described in rawpedia. http://50.87.144.65/~rt/w/index.php?title=Sharpening#Threshold According to rawpedia the threshold tool controls sharpness in a function of Luminance but if we look at the code https://github.com/Beep6581/RawTherapee/blob/master/rtengine/ipsharpen.cc#L268 it's not a function of Luminance but a function of the difference between original and gaussian blurred Luminance. Maybe we should change the documentation in rawpedia...

Ingo

[heckflosse]

Here's what I meant by swapping 2 and 3

https://gist.github.com/heckflosse/44b7a7274a52d8aa75ef

[heckflosse]

I was wrong with my suggestion. I thought value[2] is top right and value[3] is bottom right, but it's vice versa. The point about wrong rawpedia documentation is still valid...

[heckflosse]

When top right == bottom right and top_right == 2000 the following happens: All pixels which meet the condition abs(L - gauss(L)) >= 2000 will get full sharpening

When top right = 1999 and bottom right = 2000 the following happens: All pixels which meet the condition 1999< abs(L - gauss(L)) < 2000 will get from full to zero sharpening. All pixels which meet the condition abs(L - gauss(L)) >= 2000 will get zero sharpening.

In my opinion all works as designed but it needs better documentation in rawpedia

[Beep6581]

Fixed in 9fec32cbfab6a9fe60010dd627be17757d7c1f72 Documentation on its way.