We have tested the briefly described algorithms to establish visibility relations on several different models of architectural interiors. Table 1 depicts times spend by the visibility preprocessing as well as the asset of visibility computations. The percentage of cells classified as visible shows, that the visibility computations significantly reduce the amount of data necessary to process during rendering. However, efficiency of the visibility preprocessing tends to decrease rapidly with lower density of occluders in the environment (scene named 6supr). In the case of sparsely occluded environments the size of visibility trees grows, because of the combinatorial explosion of the passable portal sequences. Now, let us provide a short discussion as a subject of the following paragraphs.
Table 1: Results of the spatial subdivision and visibility
computations. The table shows the time spend by the preprocessing as
well as the average percentage of the portion of the model classified
as visible. Measured on K5/100MHz, 32MB RAM, Linux OS.
We have showed that rendering can be significantly speeded-up using an algorithm to determine the visible portion of the model in advance (i.e. before actually rendering it). However, a question remains how much work should be done as a preprocessing and what should be left for real-time calculations. Obviously, more precise preprocessing yields higher storage complexity for the precomputed information, while on the other hand reduces the complexity of on-line calculations and possibly increases their efficiency. Another question, which has not been answered yet, is when the usage of the portal based methods starts to be inefficient, leaving a place for the obscuration culling to be applied.
As already mentioned, the PVS based approach is efficient for processing densely occluded environments with cells and portals inherited in their structure. For such types of MVEs this method gives very good results and the preprocessed information is very close to the actually visible portion of the model. Moreover the dynamic queries can be performed very simply and quickly. However, if the model is not feasible enough preprocessing based on the PVS approach leads to enormous storage requirements. In such cases even more time can be spend during the dynamic queries than it would be needed using some most naive algorithm.
Obscuration culling methods are a general scheme, since practically no assumptions about the scene structure are required. This arises from the fact, that no search for the topology of the model has to be performed, namely no portals and cells need to be found. Obscuration culling methods use the ``already defined'' primitives such as polygons, to cull away the invisible portions of the model in contrast to the determination of the potentially visible sets induced by the transparent(!) portions of the model. Their potential weakness is the lack of knowledge about connectivity of primitives causing an occlusion, namely for densely occluded models (note the contrast to cell-portal approach).
It seems that most of todays research focuses on the improvement of the obscuration culling methods, since they are not so sensitive to the type of model being visualized and provide more freedom in terms of balance between preprocessing and real-time computations. Moreover as shown in [CT96], when more calculations are left for the real-time processing, the use of spatial and temporal coherence can significantly reduce their complexity.