Qtpfsgui Notes - The Fattal TMO

Qtpfsgui Fattal Notes

Test Image: Test1.hdr (Hass Burn images 357, 359, 361, April 2009 copyright DARW)

Initial Set-Up: As per Qtpfsgui Drago Notes.

Fattal Tone Mapping Operator

The Fattal TMO has been created by R Fattal, D Lischinski and M Werman. It is reportedly (Flickr website) one of the most popular of the Qtpfsgui TMO's and it has a characteristic image rendition.

The TMO presents 5 parameters:

Alpha, default 0.100, range 0.001 – 2.000;
Beta, default 0.800, range 0.600 – 1.000;
Color Saturation, default 1.00, range 0.00 – 3.00;
Noise Reduction, default 0.000, range 0.000 – 1.000; and
“Old” Fattal – a radio button to enable a previous version, default - unchecked.

With the default settings, the Fattal TMO renders with the test HDR file a well saturated image with a characteristic look of contrast variations. Closer inspection shows areas that appear to have been painted with a grey transparent brush that gives an unbalanced out-of-focus feel to the image. The histogram, available via the Adjust Levels tool, reveals a symmetric, Gaussian-like distribution of tones, centred on ca. 128.

From previous work, I have noted that the final image is critically dependant on the Noise Reduction parameter. Setting this to 0.025 gives a slightly sharper image with the greyness of some areas reduced. The histogram is broader, slightly asymmetrical and centred above 128.

The parameter Beta is said to be the amount of the algorithm applied, but increasing the value decreases the amount - counter intuitive. A value of 1 apparently applies zero effect and the result is then dark and very contrasty, in effect of no use. The histogram is concentrated to the left-hand side. A value of 0.8, the default, to 0.9 is said to be the norm, but in my hands even a value of 0.9 starts to give a darker image, as seen in the histogram, with image detail and colour being lost. A value of 0.6, the lowest setting, provides unrealistic images with more noise, under saturated and over saturated areas. The histogram shows a very condensed tonal range centred above 128. However, applying a Noise Reduction of 0.25 to 0.5 gives realistic images with reasonable detail and colour saturation. The result is appended to these notes.

The Color Saturation parameter “does what it says on the tin”, and the default value appears reasonable to my eyes for this test file.

The Alpha parameter is said to be a threshold value. Increasing Alpha is said to make details more evident, but also to increase noise. In my hands, progressively increasing Alpha from 0.1, the default, to 2.0 makes the image darker, but with no obvious noise increase on a Noise Reduction of 0.025 and with the other parameters at default. On the whole, changing the Alpha value, at least with this test file, does not appear to yield any benefit.

“Old” Fattal - On the default settings a dark images results. No further investigations have been carried out on this parameter.

Technical Aspects:

This is another gradient field method. The authors state that, “We manipulate the gradient field of the luminance image by attenuating the magnitudes of large gradients.....Our results demonstrate that the method is capable of drastic dynamic range compression, while preserving fine details and avoiding common artefacts such as halos, gradient reversals or loss of local contrast.” I take this abstract to mean that the algorithm looks for the greatest changes in luminance in the HDR image and concentrates on reducing those to a manageable LDR image, leaving small luminace chances little touched. Colour, as opposed to luminance, does not appear to be treated.

As is usual, the original paper does not discuss the parameters that are presented to the user, but the results presented in the paper are visually impressive. See also the osp.wikidot.com reference in Sources below.

Conclusions:

This TMO is very useful for both realistic and surrealistic images. A key point for me is to set Noise Reduction to values between 0.025 and 0.5 (depending on the Beta setting) and not to leave it at the default of 0.000. Thus, with all parameters at default bar Noise Reduction, set at 0.025, the test HDR gives a realistic result. Nevertheless, there is a characteristic feel to the image and that can be seen in the histogram which is symmetric, centred on 128, and covers the whole tonal range.

For OTT or surrealistic images, reduction of Beta to 0.600 (which is actually increasing the amount of the algorithm applied), and changing the Noise Reduction back towards 0.000 gives increasingly odd image rendition.

Alert readers will have noted that the pre-gamma tool has not been used at all in these tests. Finally, processing speeds are moderate for this TMO.

Sources:


Gradient Domain High Dynamic Range Compression
R. Fattal, D. Lischinski, and M. Werman
In ACM Transactions on Graphics, 2002

http://www.cs.huji.ac.il/~danix/hdr/

http://osp.wikidot.com/parameters-for-photographers rces:


http://farm4.static.flickr.com/3613/...5bb6e28a_o.jpg

Cheers

David