04 · Execution · C++17 · BVH · SAH · traversal instrumentation

Ray-Scene Acceleration

Compression addresses what has to move through a system. This project looks at the other side of the problem: once scene data exists at runtime, how efficiently can the system organize geometric work?

C++17BVHbinned SAHuniform gridAmanatides & Woo DDAcorrectness tests

Open GitHub repository

Why this belongs in the same research story

A practical representation is not only a file format. It eventually has to execute inside a graphics system.

The earlier studies focus on representation quality, training cost, and payload size. This project addresses the systems layer from a classical graphics perspective. Instead of relying on an engine to hide scene-query cost, I implemented several spatial acceleration strategies from scratch and measured how much geometric work each one actually performed.

The point is not that triangle BVHs are the same thing as Gaussian Splatting. The point is that neural graphics work eventually becomes a systems problem: memory layout, spatial organization, traversal cost, profiling, and correctness all matter once the representation reaches a runtime.

Clustered-scene benchmark

Triangles99,972

Clustered synthetic scene used for traversal stress testing.

Rays50,000

Same ray set across acceleration structures.

SAH tri tests / ray12.8

Down from 99,972 brute-force tests per ray.

Correctness failures0

34 explicit correctness checks passed.

Measured performance

Method	Triangle tests / ray	Wall time	Speedup
Brute force	99,972	28,665 ms	1×
Uniform grid	99,961	33,598 ms	0.9×
Median BVH	32.5	143.7 ms	199×
Binned SAH BVH	12.8	64.6 ms	444×

SAH improved both traversal work and wall time.

The binned SAH BVH reduced average triangle tests per ray from 99,972 to 12.8 and produced a 444× speedup over brute force in this clustered benchmark. The improvement is explained by less geometric work, not only by a faster timing number.

What failed, and why that mattered

The uniform grid was technically an acceleration structure, but it was slower than brute force.

That result is useful because it demonstrates that choosing an “accelerated” data structure is not enough. Performance depends on how the structure matches the spatial distribution of the scene. In this clustered test, the grid did almost as many triangle tests per ray as brute force and added traversal overhead on top.

The median BVH performed much better. Binned SAH improved further by choosing splits that reduced expected traversal cost. This made the connection between algorithmic structure and measured work visible in the instrumentation.

Why this supports neural graphics work

The research program ends at systems behavior, not at a reconstruction screenshot.

Gaussian Splatting and neural rendering research often reaches a point where quality alone is no longer the bottleneck. The system has to move through scene data efficiently, schedule work, use memory well, and validate that optimizations remain correct.

This project demonstrates that I can reason about those lower-level constraints directly. It complements the ML-facing experiments with explicit data structures, traversal complexity, instrumentation, and performance validation.

Research loop

Representation → Training → Delivery → Execution

Together, the four studies form one systems-oriented investigation of 3D media. NeuralScene Bench asks whether the representation is promising. FastGS asks whether it can be created more efficiently. SplatStream asks whether it can be delivered more compactly. Ray-Scene Acceleration demonstrates the runtime systems reasoning that becomes necessary once scene data has to execute.

Return to the full research narrative →

Interpretation boundary

This is a synthetic clustered-scene benchmark for ray-triangle traversal. The 444× result should not be generalized to arbitrary production scenes or to Gaussian Splatting rendering. Its role in this portfolio is evidence of graphics-systems reasoning and measurement, not a claim that BVH traversal directly accelerates the 3DGS experiments above.