Performance comparison: OpenGL and CUDA implementations for a dataset of 271 million points and 1352 different viewpoints.
Abstract
Three-dimensional point clouds have conventionally been used along with several sources of information. This fusion can be performed by projecting the point cloud into the image plane and retrieving additional data for each point. Nevertheless, the raw projection omits the occlusion caused by foreground surfaces, thus assigning wrong information to 3D points. For large point clouds, testing the occlusion of each point from every viewpoint is a time-consuming task. Hence, we propose several algorithms implemented in GPU and based on the use of z-buffers. Given the size of nowadays point clouds, we also adapt our methodologies to commodity hardware by splitting the point cloud into several chunks. Finally, we compare their performance through the response time.