Level of detail based occlusion culling for dynamic scenes

This paper presents a non-conservative occlusion culling technique for dynamic scenes with animated or user-manipulated objects. We use a multi-pass algorithm, which decides the visibility based on low level of detail representations of the geometric models. Our approach makes efficient use of hardware support for occlusion queries and avoids stalling the graphics pipeline. We have tested our approach for large real-world models from different areas. Our results show that the algorithm performs well for medium complex scenes with 5 to 20 million triangles, with a very low number of hardly noticeable pixel errors, typically in the range of 0.02 percent of the total number of visible pixels.