Investigating a different scene representation
A desktop GPU vendor asked us to evaluate ray-traced Gaussian rendering, comparing two published approaches on a current desktop GPU. The investigation looked at image quality, execution cost and the way each approach handles lighting.
The research went beyond displaying a reconstructed scene. It investigated how the representation participates in lighting, including shadows cast by Gaussians and shadows cast onto them.
Making the differences visible
The evaluation used image comparisons, heatmaps and timings to separate quality changes from execution cost. A representation that looks convincing in one view still has to hold up under the lighting and visibility conditions an application actually needs.
We examined the intersection and shadow behaviour of both representations inside a renderer, under the same scene and lighting conditions, so the comparison measured the technique rather than the setup.
Connected engineering
The Breda framework is where this kind of comparison gets built: one renderer, one scene, one set of lighting conditions, so what changes between two techniques is the technique.