Fast Normal Approximation of Point Clouds in Screen Space

Abstract

Displaying large point clouds of mainly planar point distributions yet comes with large restrictions regarding the surface normal and surface reconstruction. Point data needs to be clustered or traversed to extract a local neighborhood which is necessary to retrieve surface information. We propose using the rendering pipeline to circumvent a pre-computation of the neighborhood in world space to perform a fast approximation of the surface in screen space. We present and compare three different methods for surface reconstruction within a post-process. These methods range from simple approximations to the definition of a tensor surface. All these methods are designed to run at interactive frame-rates. We also present a correction method to increase reconstruction quality, while preserving interactive frame-rates. Our results indicate, that the on-the-fly computation of surface normals is not a limiting factor on modern GPUs. As the surface information is generated during the post-process, only the target display size is the limiting factor. The performance is independent of the point cloud’s size.