A classification-based rendering method for point models

We present a classification-based high quality rendering method for large scenes with point-based models. Each model is represented by a uniformly sized point hierarchy. All the points at the same resolution in the hierarchy share the same splat radius, and the splat radius ratio between any two neighboring resolutions are the same. We use this data structure to minimize the number of rendering points with no compromise in the image quality. Compared to the standard OpenGL point primitive, it is quite expensive to rasterize the perspectively correct splat projection even when using the advanced features of modern GPUs. We propose a classification-based rendering algorithm to reduce the number of points rendered by the complicated splatting. The potentially visible points are collected and classified into two types: a pixel point, whose image projection size is within one pixel; and a multi-pixel splat, whose image projection size is larger than one pixel. Only the multi-pixel splats are rendered by a perspectively correct splatting algorithm. The point classification is integrated with the hardware accelerated rendering. By adopting the uniformly sized point hierarchy and the classification-based rendering, we can achieve effective rendering of large scenes.