Introduction

In the late 1980's, research in ray tracing benefited from the adoption of a standard set of models (the SPD) created by Eric Haines [2]. By having a common set of models and the resulting images used by many different people, it was fairly easy to tell if a new acceleration scheme was reasonable, and have some idea on how it compared to others. This set of models was highly effective, and continues to be used today.

The global illumination community could benefit greatly by having a similar set of models for testing and comparing algorithms. Greg Ward Larson has made a number of models available in his Material and Geometry Format (MGF) as part of Radiance [4]. These models have proven very useful for testing performance and overall image appearance for different rendering algorithms. Peter Shirley developed a set of models for comparison purposes.

In this paper we propose another set of simple test scenes. We decided to focus only on global energy transfers between surfaces. Each test scene has been designed to test different modes of transport and as such stress different parts of a given renderer. We have deliberately decided to ignore color (or spectral) information, complex reflection models, complex luminaires, and participating media. These are quantities that are represented differently by just about everyone, and new test scenes will probably not be used if they force people to make significant changes to their rendering system. Additionally, the set of models should ideally satisfy several constraints:

• each should test a single aspect of global transport algorithms.
• they should require a minimal rendering framework to compute a solution.
• they should allow easy translation between formats and reasonable replication.
• models and benchmark data should be in an easily accessible repository.

The goal of simplicity directly conflicts with testing geometric complexity, which was left out for two reasons. First is that it is hard to parameterize complexity. What is complex for one algorithm is often handled easily by a slightly different algorithm, potentially requiring many different models to cover the space well. Additionally, complex models would in general test many different aspects of the transport problem, complicating the results. Another form of complexity is in the description of the scenes. Our most complicated scene presented has 9 blocks and a square light. That and the regular layout should make it possible to create close approximations of them from the text and figures. While this may not make it possible to compare images pixel by pixel, it does make qualitative comparisons possible.

The reader is assumed to be familiar with various rendering techniques. A good place to start is Glassner[1]. First we will describe a class of environments that have analytic solutions. In Section 3 the set of scenes is described.