Abstract: High quality point rendering methods have been developed in the last years. A common drawback of these approaches is the lack of hardware support. We propose a novel point rendering technique that yields good image quality while fully making use of hardware acceleration.

Previous research revealed various advantages and drawbacks of point rendering over traditional rendering. Thus, a guideline in our algorithm design has been to allow both primitive types simultaneously and dynamically choose the best suited for rendering. An octree-based spatial representation, containing both triangles and sampled points, is used for level-of-detail and visibility calculations. Points in each block are stored in a generalized layered depth image. McMillan's algorithm is extended and hierarchically applied in the octree to warp overlapping Gaussian fuzzy splats in occlusion-compatible order and hence z-buffer tests are avoided. We show how to use off-the-shelf hardware to draw elliptical Gaussian splats oriented according to normals and to perform texture filtering. The result is a hybrid polygon-point system with increased efficiency compared to previous approaches.