One of the basic goals of realistic rendering is to create images which are perceptually indistinguishable from real scenes. Since the fidelity and quality of the resulting images are judged by the human observer, the perceivable differences between the appearance of a virtual world (reconstructed on a computer) and its real world counterpart should be minimized. Thus, perception issues are clearly involved in realistic rendering and should be considered at various stages of computation such as global illumination computation, rendering, and image display. In this document we would like to focus on the problem of perception-guided global illumination computation and perception-based error metrics of resulting images.

Current global illumination algorithms usually rely on energy-based metrics of solution errors, which do not necessarily correspond to the visible improvements of the image quality. One of the research directions towards the perception-driven improvement of the global illumination computation performance relies on direct embedding of some simple error metrics to find the adequate level of light interactions between surfaces. Using such perceptual error metrics on the atomic level (e.g., every light interaction between patches, every mesh element subdivision) introduces a significant overhead on procedures that are repeated thousands of times in the course of the global illumination solution. This imposes severe limitations on the complexity of human spatial vision models, which in practice are restricted to models of brightness and contrast perception.

The scenario of embedding advanced HVS models into global illumination and rendering algorithms is very attractive, because more robust perception-driven guidance of the global illumination computation can be achieved. A practical problem arises that the computational costs incurred by the HVS models introduce an overhead to the actual lighting computation, which may become the more significant the more rapid the lighting computation becomes. The potential gains of such perception-driven computation can be easily cancelled out by this overhead, depending on many factors, such as scene complexity, performance of a given lighting simulation algorithm for a given type of scene, image resolution, and so on. The HVS models can be simplified to reduce the overhead, but then underestimation of visible image artifacts becomes more likely. To prevent such problems and to compensate for ignored perceptual mechanisms, more conservative (sensitive) settings of the HVS models should be applied, which may also reduce gains in the lighting computation driven by such models.

It seems that keeping the HVS models at some high level of sophistication and embedding them into rendering algorithms, which are supposed to provide a meaningful response rapidly, e.g., in tens of seconds or single minutes, may be a difficult task. The question arises what approach should be taken:

• Using advanced HVS models at the design stage of the global illumination algorithms and the tuning of their parameters. In such a case, the resulting algorithms can spend 100% of their computation time in lighting simulation, and the costs of HVS processing (which is performed off-line) are of secondary importance.
• Embedding advanced HVS models directly into the global illumination computation, which incurs significant overhead, but on the other hand better adaptation to the scene specific characteristics can be achieved.
• Using a hybrid of inexpensive energy-based metrics and more costly perception-based metrics as a trade-off solution.

Another question arises how to generalize the approaches chosen for static image generation toward the global illumination computation for dynamic environments, which becomes important for the high-quality animation rendering. I am looking forward discussing those issues during the campfire.