Both the absolute and dynamic ranges of light we encounter in natural environments are vast. Over the course of the day the absolute level of illumination can vary by more than a trillion to 1 from bright sunlight down to starlight. The dynamic range of light energy in a single environment can also be large, on the order of 10,000 to 1 from highlights to shadows.
Although physically-based global illumination rendering methods now allow us to accurately simulate the distribution of light energy in environments, and soon-to-be-ubiquitous digital cameras will be able to capture scenes with widely varying light intensities, neither of these imaging technologies give provision for displaying their results on conventional CRT and print-based media which have useful luminance ranges of less than 100 to 1.
In computer graphics, work is just beginning on this problem. Tumblin (1993) and Ward (1994) introduced the community to the idea of a tone reproduction operator that maps from scene luminances to display luminances, attempting to produce a visual match between the scene and display. Nakame (1990) and Spencer (1995) developed techniques for producing convincing renderings of bright light sources based on optical models of diffraction and glare. Chiu (1993) and Schlick (1994) have introduced computational techniques for dynamic range compression and display of synthetic images based on non-linear local operators.
A critical and underdeveloped aspect of all this work is the visual model on which the algorithms are based. As we move through different environments or look from place to place within a single environment our eyes adapt to the prevailing conditions of illumination both globally and within local regions of the visual field. These adaptation processes have dramatic effects on the visibility and appearance of objects and on our visual performance. In order to produce realistic displayed images of synthesized or captured scenes, we need to develop a more complete visual model of adaptation. This model will be especially important for immersive display systems that occupy the whole visual field and therefore determine the viewer's visual state.
In our Siggraph '96 paper "A Model of Visual Adaptation for Realistic Image Synthesis" we introduced a perceptually-based algorithm for creating accurate visual displays of scenes illuminated at a wide range of absolute levels. The algorithm is based on a visual model derived from psychophysical data that takes into account the changes in sensitivity, color appearance, and acuity that occur with visual adaptation to different light levels. The algorithm then creates a mapping from scene luminances to display luminances that produces an image that is a visually faithful representation of the scene. Because the display algorithm is psychophysically-based and because the rendering algorithms it is used with are physically-based, the images are not just pleasing pictures, but are accurate visual representations that predict what an observer would be able to see if they were standing in the actual scene. This perceptually predictive attribute allows these images to be used in illumination engineering studies, ergonomics and safety design, and simulator technologies.
Having addressed the issue of accurately displaying images representing a wide range of absolute levels, in this proposal we want to address the dynamic range issue. Wide dynamic range scenes are not uncommon. Sunlight and shadow falling across a printed page produces an image whose dynamic range easily exceeds 1000:1. The dynamic range given by an unshaded lamp on a table top can be as much as 30,000:1 from the bulb to the shadow under the table. To accurately display these images on conventional display devices we need to develop non-linear mappings from scene luminances to display luminances that compress the dynamic range of the images while preserving visual fidelity.
We are not the first to address the dynamic range issue in graphics. In fact some of the earliest work was done by Stockham (1972) at the University of Utah. In his approach, a global mapping from scene to display was developed that preserved contrast while manipulating actual luminance levels. This work was one of the precursors to histogram equalization techniques (see (Pizer 1987) for a review), and adaptive histogram equalization in which local regions of the image are displayed using different mappings based on the luminance values found within a region. Land's Retinex theory (1964, 1986) has served as the basis of some current approaches (Jobson 1996) that generate adaptive mappings that vary with both the luminances found within a local region as well as the scale of scene features. There is also the aforementioned work by Chiu (1993) and Schlick (1994) on non-linear mappings. The problem with all these approaches is that they are based on ad-hoc or limited visual models. This means they neither accurately represent the appearance of high dynamic range scenes nor do they take good advantage of the limits of visual processing to constrain the mapping from scene to display.
The critical element of the visual model needed to accurately display high dynamic range images is a model of local adaptation in the visual field. Visual neurons can only signal about a 100:1 change in light intensity. To cope with wide dynamic range scenes these neurons adapt locally to the prevailing conditions of illumination within a limited region of the scene. Our visual impressions of high dynamic range scenes are constructed from the responses of these locally adapted neurons as we move our eyes from place to place in the scene. Realistic display algorithms for high dynamic range scenes must take local adaptation into account in producing the mapping from scene luminances to display luminances.
We propose to develop new display algorithms for high dynamic range scenes based on advanced psychophysical models of visual adaptation. To develop these algorithms we need to find answers to the following questions:
To develop our display algorithms we will need to integrate the experimental results found in the psychophysics literature into a comprehensive, computationally tractable, and practical model of vision that can be applied to the dynamic range problem. This will require a significant effort, because although psychophysical experiments are very exacting, they typically examine the visual system piecemeal, under reductionistic conditions. Fortunately we will be able to build on the visual model we have developed in our earlier work.
Generating images based on the model will be the next stage of work. Finally because we want these images to be accurate visual simulations, we intend to conduct experimental comparisons between the images and real high dynamic range scenes to assess the visual fidelity of the images both in terms of the quality of their appearance and in terms of quantitative measures of visual performance.
This work will have wide reaching impact on the field of computer graphics and digital imaging. First, on a practical level it will provide a sound scientific method for realistically displaying the results of physically based global illumination rendering algorithms. These methods will also be useful for displaying images captured by the coming generation of consumer grade high dynamic range digital cameras. Second, because the algorithm is based on psychophysical data, the images produced are accurate visual representations of the simulated or captured scenes. This will allow these images to be used quantitatively by the engineering and design professions. Finally, in computer graphics itself this work is important because knowledge of the capabilities and limits of visual processing will allow graphics researchers to develop more efficient algorithms that only compute the visual information actually picked up by an observer in a particular setting or performing a particular task. A particularly significant application of this work will be in immersive display technology where the display system occupies the entire visual field and determines the visual state of the observer.