03/05/03© 2003 University of Wisconsin Last Time Tone Reproduction If you don’t use perceptual info, some people call it contrast reduction.

03/05/03 © 2003 University of Wisconsin

Last Time

• Tone Reproduction

• If you don’t use perceptual info, some people call it contrast reduction


Today

• Perceptually driven tone-reproduction


Local Tone Reproduction

• Last time I introduced the idea of local tone reproduction methods– Filter to get a high-pass and low pass version of the image

– Reduce the contrast in the low pass

– Add the reduced low-pass to the high-pass to get a new image

• The filter you use should be edge preserving– A filter that smoothes out regions with low gradient, but does not

smooth across high gradients


Edge Preserving Filters(Durand and Dorsey, SIGGRAPH 2002)

• The idea is to take a smoothing filter, say a Gaussian, and multiply it by a function that reduces the weight at sharp edges (and influence function)– It turns out, this is closely related to robust statistical

estimation – the problem of ignoring outliers in statistical data


Robust Estimation

• Instead of least squares estimation, use something that reduces the weight of outliers from the mean– Least squares minimizes sum of squared differences


Results

• Right, a full image• Below, still some artifacts at edges

– Fix by detecting unreliable pixels and smoothing


Gradient Compression(Fattal, Lischinski, Werman, SIGGRAPH 2002)

• Instead of reducing the dynamic range of the data directly, reduce the size of the gradients– Ultimate effect is similar, but it’s easier to make it local

• Reduce large gradients more than small ones– Get most reduction in dynamic range– Small gradients are probably due to texture, large ones due to shadows,

occlusion, different surfaces


Gradient Compression in 2D

• In 1D, you can reduce the gradients and then simply integrate to extract the new signal

• In 2D, the gradient field has to be conservative

• To work around this, seek a final image that has the closest legal gradient field to the desired one– Results in a Poison equation over the image

– Set appropriate boundary conditions and solve to get the image


Details

• Work on luminance channel from CIE XYZ

• Do multi-resolution gradient attenuation– Generate a Gaussian image pyramid

– Use central differences to estimate gradient at each level

– Attenuate by multiplying by an attenuation function:

– Use linear filtering to push low-res results up to high-res

0.98.0 gradient, average times1.0

,

,,

yxH

yxHyx k

kk


Dealing with Color

• Work in RGB space for final image

• For C in (R,G,B) compute:

– s between 0.4 and 0.6 worked well

• Not really the best thing to do

out

s

in

inout L

L

CC


Results


Local Filtering Methods


Ward’s Method


Attenuation Image


Psychophysics

• The study of the perception of physical quantities– Does not try to explain anything, just observes what we observe

• Vital to many areas of graphics– Tone-reproduction: generating the same per perceptual sense with

different dynamic range– Rendering Control: Stopping rendering algorithms when the results

are no longer perceptible– Error metrics: Comparing images with a sense of what is visually

important

• Ferwerda, Pattanaik, Shirley and Goldberg, SIGGRAPH 1996, is a good reference– The next several slides were borrowed from the paper


Dynamic Range (Again)

• Note the names for various ranges of perception– Related to which part of the vision system (rods or cones) is

functioning effectively


Spectral Sensitivity

• Incoming radiance is integrated against these curves to get visual response– To match visual response with a different radiance, have to take these

curves into account• Note in particular the change in color sensitivity

– Rods are NOT color sensitive


Detection Thresholds

• Experiment: Flash a light against a background. How much brighter does the light have to be in order to be noticed?– Sizes measured in degrees of arc

• Not so important in graphics– Most events we deal with

change illumination well beyond the detection threshold


Contrast Thresholds

• Experiment: Show people a grating. What contrast must be present (difference in foreground/background luminance) for the grating to be distinguishable?

• Very important to graphics– Places a limit on the maximum useful display resolution

– Places a limit on the amount of detail that should be distinguishable in a tone-reproduction algorithm

– Useful in measuring the effect of various aliasing and noise effects

• We might expect the curve on the next slide to flatten out at high luminance. Why?


Contrast Thresholds (Acuity)


Adaptation

• Measures the effect of transition from light to dark, and vice versa

• This is light to dark


Adaptation Dark-to-Light


Perceptual Tone Reproduction

• Various tone reproduction algorithms can be modified to exploit these effects

• Deliberately reduce contrast if the light level is low

• Reduce color saturation at low light levels

• Model adaptation level of viewer over time by manipulating contrast and color


Adjusting for Acuity and Color


Adjusting for Adaptation

• Not so easy – you can’t make the monitor appear “painfully bright”


Histograms and Perception

• Ward’s histogram method can be adapted to place limits on resolvable features– Another bound on the magnitude of histogram bins


Other Effects

• To get color washout, reduce (x,y) by an amount that depends on the average luminance– Formulas in Ward’s and Ferwerda’s papers

• Veiling is due to light scattering inside the eye– Responsible for the halos around truly bright objects

– Ward’s method can handle this

– Glare is something different – an adaptation problem


Histograms and Perception

03/05/03© 2003 University of Wisconsin Last Time Tone Reproduction If you don’t use perceptual info, some people call it contrast reduction.

Documents

university of wisconsindealing

university of wisconsinedge

low gradient

high gradients

low pass version

dynamic rangesmall gradients

low passadd

smoothing filter