Building conservative frustum and occlusion culling

Preserve visible contributions

Visibility culling rejects work before expensive drawing. Frustum tests reject objects outside the camera volume; occlusion tests reject objects hidden behind other geometry. Bounding volumes provide cheaper proxies for these decisions. [1] A safe rejection must not remove any contribution needed by that view.

Choose the view explicitly. An object outside the main camera may still matter to a shadow map or reflection, so one visibility list does not automatically serve every pass.

Understand delayed evidence

Testing every object and immediately waiting for each result can cost more than drawing. Hierarchical grouping and interleaving useful rendering with queries reduce this scheduling problem. The original coherent hierarchical approach uses previous visibility to schedule current work. [2]

Availability is separate from a zero visible-sample result. Vulkan exposes query availability and warns that reused query slots can still contain results from an earlier use if resetting has not executed. [3] Associate every result with its intended frame or generation.

Count an illustrative workload

Suppose a scene contains 10,000 candidate objects. A frustum stage rejects 7,000, leaving 3,000. Fresh occlusion evidence rejects another 1,800, leaving 1,200 candidates for drawing. The count reduction alone says nothing about frame-time savings: testing tiny cheap objects can outweigh their avoided work.

Now group a hidden building interior under one conservative bound. If that parent is fully occluded, many descendants can be skipped together. If it intersects a visible doorway, descend or draw conservatively. Compare this grouped case with per-object testing under the same camera path.

Verify bounds and result handling

Visualize bounds over animated, skinned, and procedurally displaced meshes. Exercise maximum motion rather than checking only the rest pose. Ensure any occlusion proxy test does not accidentally write its box into the scene's depth buffer. [1]

Test camera teleports, opening doors, fast-moving occluders, and newly spawned objects. Treat unknown or invalidated evidence as visible until valid rejection is available. Log rejected groups, candidate counts, query age, CPU scheduling time, and GPU test cost. Compare against a culling-disabled reference to find missing objects.

Decide when testing pays

Large hidden groups are promising; open aerial views may provide little useful occlusion. [2] Keep a way to disable or reduce the expensive stage for workloads where it adds overhead.

Conservative bounds may submit invisible geometry, but undersized bounds can erase visible content. Optimize bounds only after correctness is established. Revisit the grouping strategy when level layout or deformation changes, and evaluate total frame time rather than celebrating a high rejection percentage.

Sources

• [1] Efficient Occlusion Culling — Dean Sekulic, Croteam; NVIDIA GPU Gems.
• [2] Hardware Occlusion Queries Made Useful — Michael Wimmer and Jiří Bittner; NVIDIA GPU Gems 2.
• [3] vkGetQueryPoolResults — Khronos Group.