Screen-Space Ambient Occlusion: Interpret GTAO as a Visibility Estimate

Define what the scalar represents

Ambient occlusion approximates how surrounding geometry blocks ambient illumination at a surface. A useful diffuse formulation measures cosine-weighted visibility over its hemisphere. Here, a visibility value of 1 means open and 0 means fully occluded. This scalar is an approximation to a particular lighting problem, not a general shadow multiplier for the finished image.

GTAO reformulates the visibility integral using horizon searches in slices and analytic integration within each slice. The original work also explains the assumptions behind separating occlusion from illumination. [1]

Understand the depth buffer's missing geometry

Screen-space methods reconstruct nearby positions from camera depth and use supplied or reconstructed normals. The buffer describes visible surfaces, not a complete solid scene. Hidden surfaces, off-screen occluders and object thickness are unavailable unless additional representations provide them.

A thin rail can therefore cast too much apparent occlusion, while missing geometry can cast too little. XeGTAO discusses this height-field assumption, thickness heuristics and comparison against ray-traced reference images. [2] Increasing the sample count reduces sampling error; it does not supply missing geometry.

Worked example: integrate the intended lighting term

Suppose a simplified scalar lighting calculation has ambient contribution 0.4 and already evaluated direct contribution 0.6. An AO visibility estimate of 0.65 changes the ambient term to 0.4 * 0.65 = 0.26. The combined result becomes 0.26 + 0.6 = 0.86.

Multiplying the finished value of 1 by 0.65 instead would also darken the direct contribution. That is a different artistic operation. This example illustrates integration, not a universal lighting equation. Directional environment lighting, specular reflection and indirect light already accounting for visibility need more careful treatment than blindly applying the same scalar everywhere. [1]

Calibrate geometry and scale first

Build a test scene with a wall corner, a thin pole near a floor and an isolated flat surface. Confirm depth reconstruction and normal orientation before tuning strength. Express the effect radius in consistent scene units. A larger radius changes the neighborhood being approximated; it is not simply a higher quality setting.

Inspect the raw AO, then its filtered result. Move the camera so test geometry crosses screen boundaries and depth discontinuities. Tune edge-aware filtering and thickness handling against a reference with the same radius. XeGTAO demonstrates why settings that improve thin objects can worsen genuinely deep occluders. [2]

Evaluate stability and artistic limits

Check contact detail, halos and camera-dependent changes during motion. Temporal filtering can reduce noise but needs validation when visibility changes. Avoid stacking several occlusion terms that darken the same contribution twice. AO does not simulate light bouncing between colored surfaces, and its blocked-light approximation can over-darken cavities. Keep the effect proportionate to the lighting model and measure the complete reconstruction, estimation and filtering cost.

Sources

• [1] Practical Realtime Strategies for Accurate Indirect Occlusion — Jorge Jimenez, Xian-Chun Wu, Angelo Pesce and Adrian Jarabo, original technical report.
• [2] XeGTAO — Intel GameTechDev, original implementation and technical explanation.