Table 2. Rad rendering variables, their meanings, values and effects.
The most heavily used commercial rendering environments are graphics libraries. These libraries are often developed by computer manufacturers specifically for their graphics hardware. They tend to be very fast and eminently practical, but are not physically accurate or sufficiently general in terms of reflectance, transmittance and geometric modeling to be useful for lighting design. Accuracy and generality have been sacrificed for speed.
So-called "photo-realistic" rendering systems may be general and practical, but they are not accurate. One of the best examples of photo-realistic rendering software is RenderMan [Upstill89], which is based on the earlier REYES system [Cook87]. Although Renderman can generate some beautiful pictures, global illumination is not incorporated, and it is difficult to create accurate light sources. Shadows are usually cast using a z-buffer algorithm that cannot generate penumbra accurately [Reeves87]. However, the system does provide a flexible shading language that can be programmed to simulate some global illumination effects [Cook84b][Hanrahan90].
In recent years, many free rendering packages have been distributed over a network. RayShade, one of the best free ray-tracers, does not simulate diffuse interreflection, and uses a non-physical reflection model. As with most photo-realistic rendering software, accuracy is the key missing ingredient. Filling in this gap, some free radiosity programs are starting to show up on the network. Though the author has not had the opportunity to learn about all of them, most appear to use traditional approaches that are limited to diffuse surfaces and simple environments, and therefore are not general or practical enough for lighting design.
Systems for research in rendering and global illumination algorithms exist at hundreds of universities around the world. Few of these systems ever make it out of the laboratory, so it is particularly difficult to judge them in terms of practicality. However, from what research has been published, it appears that most of these systems are based on classic or progressive radiosity techniques. As we have noted, radiosity relies on diffuse surfaces and relatively simple geometries, so its generality is limited. Extensions to non-diffuse environments tend to be very expensive in time and memory, since directionality greatly complicates the governing equations of a view-independent solution [Sillion91]. Recent work on extending an adjoint system of equations for a view-dependent solution [Smits92] to non-diffuse environments appears promising, but the techniques are still limited to simple geometries [Christensen93][Aupperle93].
The basic problem with the radiosity method is that it ties illumination information to surfaces, and this approach runs into trouble when millions of surfaces are needed to represent a scene. Rushmeier et al addressed this problem with their "clumping" approach, which partially divorces illumination from geometry [Rushmeier93]. Baum et al [Baum91] developed techniques for automatically meshing large models, which saves on manual labor but does not reduce time and space requirements. The theater model shown in Figure 5 was rendered in [Baum91] using automatic meshing and progressive radiosity. Meshing the scene caused it to take up about 100 Mbytes of memory, and rendering took over 18 hours on an SGI R3000 workstation for the direct component alone, compared to 5 hours in 11 Mbytes using Radiance on the same computer.
Some of the larger architecture and engineering firms have the resources to create their own in-house lighting simulation and rendering software. Although it is difficult to speculate as to the nature and capabilities of these systems since they are hidden from public view, the author is aware of at least a half-dozen well-funded projects aimed at putting the state of the art in global illumination into practice. Most of these projects are based on progressive radiosity or something closely related. In at least two cases, Abacus Simulations in Scotland and Siemens Lighting in Germany, in-house software projects have been abandoned after considerable expense in favor of using Radiance. At least two other firms, Ove Arup in England and Phillips Lighting in the Netherlands, use Radiance side by side with in-house software. Of course, we cannot conclude from this that Radiance is the best, but the trend is encouraging.
By far the most relevant class of software to compare is commercial lighting simulation and rendering programs. Most of these systems are practical, or people would not buy them. Most are also accurate, or they would not qualify as lighting simulations. The problem is lack of generality. LumenMicro (Lighting Technologies, Boulder, Colorado) is the biggest-selling program among lighting designers, yet it is limited to modeling environments built from grey, diffuse, axis-aligned rectangles. A more promising product is called LightScape (LightScape Graphics Software, Toronto, Canada). This software uses progressive radiosity and graphics rendering hardware to provide real-time update capabilities. LightScape's ray tracing module may be used to improve shadow resolution and add specular effects, but this solution is expensive and incomplete. Also, memory costs associated with complicated geometries limit the practicality of this system. To be fair, LightScape is in its initial release, and has some very accomplished researchers working on it.
One program that shows great potential has recently been released to the U.S. market, Arris Integra (Sigma Design in Baltimore). This program uses a bidirectional ray tracing technique developed by Fujimoto [Fujimoto92], and its capabilities have been demonstrated in [Scully93]. The chief drawback of this system seems to be that it is somewhat difficult and expensive to use, costing roughly 15,000 dollars for the basic software and taking many long hours to perform its calculations.
As part of this larger standardization effort, we would like to see a common scene description format adopted by the rendering community. There are many candidates at this point, but none of them contain physically valid light source and material descriptions. We would welcome the use of the Radiance format, but extending a conventional scene description language might work better. We suggest the formation of a small committee of global and local illumination researchers to decide what should be included in such a format. We further suggest that one or two graphics hardware or software vendors could cover expenses for this task. In return, the vendors would get a new, physically valid foundation for building the next generation of rendering solutions.
The future of physically-based rendering depends on cooperation and agreement. We must agree on a starting point and work together towards a goal to bring science to this art.