FP-ToddKosloff

From Visualization Sp06

1 Proposal
- 1.1 Group Members
- 1.2 Description
2 Midpoint Design Discussion
3 Final Deliverables

[edit]

Proposal

3D Visualization of High Dynamic Range Images

[edit]

Group Members

Todd Kosloff

[edit]

Description

Typical display devices, including monitors, projectors, and printouts, can display images with only a limited dynamic range. Luminance in the real world, however, has a much higher dynamic range. The difference in luminance between specular reflections of sunlight, vs. a shadowed region, will typically be much greater than any readily available display device can handle. Computer generated images, as well as digital photographs created using special photographic techniques, can numerically represent the full dynamic range present in nature.

Various techniques for displaying these images on ordinary display devices have been created in the past. The simplest technique is to map a certain scene luminance to a certain display luminance, and allow pixels that are too bright relative to that luminance to be displayed as pure white, whiel pixels that are too dim are displayed as black. This mimics the notion of choosing a certain exposure setting on a camera. Only a subset of the scene will be visible at any given exposure, but by either interactively changing the exposure, or instead viewing several exposures side by side, a sense of the scene as a whole can be gleaned. More sophisticated tone mapping operations bring the entire scene into visible range in a single image, with varying sorts of artifacts introduced as a byproduct of the range compression. Some tone mapping operators mimic the operation of the human visual systemm, attempting to recreate the sensation of viewing a real world scene by incorporating light and dark adaptation, chromatic adaptation, loss of acuity and loss of saturation at very dim light levels, and so forth.

Tone mapping operators are geared towards graphics applications; the goal is to produce pleasing images that look plausible. Tone mapping inherently loses the absolute luminance levels that were present in the original image. While the resulting images can look stunningly beautiful, looking at these images provides the viewer with no sense of how bright the various areas of the original image were in relation to one another. Subjectively we can assess that the shadow is certainly dimmer than highlight, but the precise magnitude of that difference is utterly lost.

The only existing method I am aware of for accurately displaying HDR luminances in a quantitative manner is the false coloring method. Similar to a weather map, different colors are assigned to different luminance values. A scale off to the side can indicate that red means 10,000 cd/m^2, where green means 100, for example.

False coloring quantitatively displays the luminance information, but it does not allow for visual comparisons of luminance. Red might mean 1000 times the brightness of green, but this mapping is arbitrary and not inherent in the colors themselves.

I propose to solve this problem by displaying HDR images as 3D height fields. My application will be similar in flavor to a fly-over of a landscape. Bright highlights will appear as mountains, dim shadows will appear as valleys. It is my hypothesis that viewing the height of a landscape will provide an easy understanding of the true luminance distribution in an image. This technique will advance the intuitive understanding of the underlying luminance structure of images.

The user of my system will reap the following benefits:

1. A better appreciation for high dynamic range imaging. 2. Insights gained may be useful for constructing better tone mapping operators. 3. Comparing the luminance landscape of a tone mapped image to the landscape of the original image may be a good way to illustrate just what the different tone mapping operators are really doing.

Challenges I might face:

Height ratios of many thousands to one might make for a very uneven landscape. A way to easily navigate through such a jagged landscape will be necessary. Taking the log of the height might map it to something better resembling the way it is perceived, as well as helping to flatten the landscape.

Occlusion. Too many mountains can make the valleys hard to see. Careful use of transparency could be of use here.

Numeric information. To visualize precisely how bright a given height is, isoluminance contours can be displayed on the height map, corresponding to user-selected luminance values. This is a special case of false coloring. False coloring in its more general form can also be of use here.

Mapping the landscape to the image itself. Standing in the middle of a vast landscape, the user is likely to forget the correspondence between the landscape and the image. This can be addressed by, in a separate window, having a tone-mapped version of the image with markings applied to it to indicate where in the landscape we are currently located, and what orientation the camera has.

[edit]