A Multiscale Model of Adaptation and Spatial Visionfor Realistic Image DisplaySumanta N. Pattanaik James A. Ferwerda Mark D. Fairchild∗Donald P. GreenbergProgram of Computer Graphics†, Cornell UniversityAbstractIn this paper we develop a computational model of adaptation andspatial vision for realistic tone reproduction. The model is basedon a multiscale representation of pattern, luminance, and color pro-cessing in the human visual system. We incorporate the model intoa tone reproduction operator that maps the vast ranges of radiancesfound in real and synthetic scenes into the small fixed ranges avail-able on conventional display devices such as CRT’s and printers.The model allows the operator to address the two major problems inrealistic tone reproduction: wide absolute range and high dynamicrange scenes can be displayed; and the displayed images matchour perceptions of the scenes at both threshold and suprathresh-old levels to the degree possible given a particular display device.Although in this paper we apply our visual model to the tone re-production problem, the model is general and can be usefully ap-plied to image quality metrics, image compression methods, andperceptually-based image synthesis algorithms.CR Categories: I.3.0 [Computer Graphics]: General;Keywords: realistic imaging, visual perception, tone reproduc-tion, adaptation, spatial vision1 INTRODUCTIONThe range of light we encounter in natural scenes is vast. The ab-solute level of illumination provided by direct sunlight can be 100million times more intense than starlight. The dynamic range oflight energy can also be large, on the order of 10,000 to 1 fromhighlights to shadows, or higher if light sources are visible.Although physically-based rendering methods and new tech-niques that utilize the output of digital cameras [Debevec97] nowallow us to produce radiance maps that accurately represent thewide variations of light energy in scenes, neither of these methodsspecify how to realistically display these images on conventional∗On sabbatical leave from: Munsell Color Science Laboratory, Cen-ter for Imaging Science, Rochester Institute of Technology, 54 LombMemorial Drive, Rochester, NY 14623-5604, USA. Web address:http://www.cis.rit.edu/people/faculty/fairchild†580 Rhodes Hall, Cornell University, Ithaca, NY 14853, USA.E-mail addresses: {sumant,jaf,mdf,dpg}@graphics.cornell.eduWeb address: http://www.graphics.cornell.edu.electronic and print-based media which have only moderate outputlevels and typical dynamic ranges of less than 100 to 1.Recently graphics researchers have started to address this issueby developing tone reproduction operators that map scene radi-ances to display outputs with the goal of producing a visual matchbetween the scene and the display. There are two major problemsto be solved in realistic tone reproduction:• to find an operator that maps the vast ranges of radiancesfound in scenes into the range that can be produced by a givendisplay device.• to be certain that this operator produces images that match ourperceptions of the scenes.The critical element that links these two problems is the visualmodel used in the tone reproduction operator. Visual models areused to relate the perceptual responses of a scene observer to theresponses of the display observer in order to specify a mapping thatproduces a visual match between the scene and the display. A cen-tral issue is that different tone reproduction operators have made useof different visual models to determine what constitutes a match.Tumblin and Rushmeier’s [1993] operator is based on Stevens’[1961] model of brightness and contrast perception illustrated inFigure 1b. The operator attempts to produce images that capturethe changes insuprathreshold brightness and apparent contrastthatoccur with changes in the level of illumination. Ward [1994] intro-duced an operator based on a model of contrast sensitivity derivedfrom threshold vs. intensity (TVI) functions similar to those shownin Figure 1a. Its goal is to match the threshold visibility of featuresin the image to features in the scene. Ferwerda [1996] developed anoperator based on a model of adaptation that like Ward’s matchesthreshold visibility, but also accounts for the changes in visual acu-ity and color discriminability that occur with the changes in thelevel of illumination.Both threshold and suprathreshold models of vision capture im-portant aspects of our visual experience, and a realistic tone repro-duction operator should produce a mapping that matches both as-pects. Unfortunately, threshold models don’t scale well to predictsuprathreshold appearance, and suprathreshold models don’t accu-rately predict visual thresholds.Recently much effort has been devoted to developing tone re-production operators for high dynamic range scenes. Chiu [1993],Schlick [1995], and Jobson [1996] introduced spatially-varying op-erators that compress high dynamic range scenes into the limitedrange available on display devices, but the ad-hoc visual modelsthey incorporate limits what can be said about the visual fidelity ofthe mappings. Tumblin [1997] has recently introduced an opera-tor for high dynamic range scenes based on a model of perceptualconstancy. Although this operator produces attractive images, themodel it uses is not quantitative, and therefore the operator can’tpredict whether an image will be a visual match to a scene. Fi-nally, Ward-Larson [1997] has introduced an operator that extendsthe work of Ward [1994] and Ferwerda [1996] with a model oflocaladaptation, to produce a threshold-based operator that can handleFigure 1: Threshold and suprathreshold models of vision: a) Threshold vs. intensity (TVI) functions for the rod and cone systems. The curves plot thesmallest threshold increment ∆L necessary to see a spot against a uniform background with luminance L. b) Stevens’ model of suprathreshold brightness andapparent contrast. The curves plot the changes in brightness and apparent contrast of gray targets and a white surround as the level of illumination rises (1 Bril= apparent brightness of a target with a luminance of 1µLambert). Adapted from [Ferwerda96, Stevens61].high dynamic range scenes, and also match the changes in thresholdvisibility, visual acuity and color discriminability that occur withchanges in the level of illumination.Although the innovations introduced in each of these operatorsrepresent significant advances toward addressing the two
View Full Document
Unlocking...