HLOD Re nement Driven by Hardware Occlusion Queries

In order to achieve interactive rendering of complex models comprising several millions of polygons, the amount of processed data has to be substantially reduced. Level-ofdetail (LOD) methods allow the amount of data sent to the GPU to be aggressively reduced at the expense of sacri cing image quality. Hierarchical level-of-detail (HLOD) methods have proved particularly capable of interactive visualisation of huge data sets by precomputing levels-of-detail at di erent levels of a spatial hierarchy. HLODs support out-of-core algorithms in a straightforward way and allow an optimal balance between CPU and GPU load during rendering. Occlusion culling represents an orthogonal approach for reducing the amount of rendered primitives. Occlusion culling methods aim to quickly cull the invisible part of the model and render only its visible part. Most recent methods use hardware occlusion queries (HOQs) to achieve this task. The e ects of HLODs and occlusion culling can be successfully combined. Firstly, nodes which are completely invisible can be culled. Secondly, HOQ results can be used for visible nodes when re ning an HLOD model; according to the degree of visibility of a node and the visual masking perceptual phenomenon, then it could be determined that there would be no gain in the nal appearance of the image obtained if the node were further re ned. In the latter case, HOQs allow more aggressive culling of the HLOD hierarchy, further reducing the amount of rendered primitives. However, due to the latency between issuing an HOQ and the availability of its result, the direct use of HOQs for re nement criteria cause CPU stalls and GPU starvation. This thesis introduces a novel error metric, taking visibility information (gathered from HOQs) as an integral part of re ning an HLOD model, this being the rst approach within this context to the best of our knowledge. A novel traversal algorithm for HLOD re nement is also presented for taking full advantage of the introduced HOQ-based error metric. The algorithm minimises CPU stalls and GPU starvation by predicting HLOD re nement conditions using spatio-temporal coherence of visibility. Some properties of the combined approach presented here involve improved performance having the same visual quality (whilst our occlusion culling technique still remained conservative). Our error metric supports both polygon-based and point-based HLODs, ensuring full use of HOQ results (our error metrics take full advantage of the information gathered in HOQs). Our traversal algorithm makes full use of the spatial and temporal coherency inherent in hierarchical representations. Our approach can be straightforwardly implemented.