Retrieving Material Properties of Implicit Surface Acceleration Structures - Colorizer: A library for retrieving Material Properties of Implicit Sur faces generated from Triangle Scenes

Sammanfattning

Distance field traversal offers a hardware-independent path for real-time ray–surfaceintersection, avoiding the specialized hardware that triangle-based ray tracing requires.Libraries such as AMD FidelityFX Brixelizer efficiently resolve these intersections,but provide no mechanism for retrieving the surface properties — albedo, emission,roughness — that illumination algorithms require.This thesis presents Colorizer, a C++/DirectX 12 static library that addresses thisgap by maintaining a world-space surface property cache. The cache organizes thescene into a camera-centred circular grid of chunks, each decomposed into a voxelgrid whose six cardinal faces form local inverse cubemaps. Surface properties arebaked into two-dimensional texture atlases via a scan-conversion rasterization pass,using an atomic depth sort to resolve per-voxel fragment priority. A spatial hashmaps voxel face tuples to atlas tiles, and three layered error corrections — shellexpansion, weighted bilinear reconstruction, and cardinal face fallback — absorb theimprecision inherent to distance field hit points. The cache is updated amortizedacross frames, targeting real-time budgets without requiring full per-frame coverage.The system is evaluated in the Amazon Lumberyard Bistro scene on an AMDRadeon RX 7800 XT. Library overhead is measured at 0.12 ms per frame, andsurface cache sampling is competitive with barycentric reconstruction against fullresolution textures. Bake jobs average 0.25 ms with a P95 of approximately 0.6 msat an 8 m chunk size. The cache produces plausible albedo and emission backingfor diffuse and high-roughness global illumination, with the error correction stackempirically absorbing predicted hash collisions at fill rates up to 37.9%. The primaryremaining artifact is intra-voxel leakage, addressable through a separate sparseemissive cache structure proposed in the discussion.