Non-Photorealistic Point Cloud Rendering via Per-Segment Image-Space Stylization and Transparency-Aware Compositing

Ole Wegen, Josafat-Mattias Burmeister, Max Reimann, Rico, Jürgen Döllner

Abstract: 3D point clouds are an integral part of spatial computational models across various application areas. However, visualizing them is challenging due to their inherent incompleteness, sparsity, and inaccuracies resulting from the acquisition process. Direct rendering of raw point clouds often results in cluttered and ambiguous images that fail to effectively communicate information. Non-photorealistic rendering (NPR) addresses these issues by using abstraction, simplification, and highlighting to reduce visual complexity and guide the viewer's focus. However, current NPR methods for point clouds typically apply a uniform visual style across the entire point cloud, causing a loss of details or saliency in certain areas. To address these issues, we previously introduced an approach for per-segment image-space stylization, which we extend in this paper, while also providing a more detailed description and evaluation. In our approach, the point cloud is first segmented, then each segment is rendered individually, and finally the per-segment renderings are composed into a unified image. By incorporating techniques from image-based artistic rendering, our approach offers significant artistic freedom in designing rendering pipelines. To ensure coherent compositing of the per-segment images into a consistent final image, we introduce a depth inpainting step to estimate depth values for pixels that received color during image-space stylization. Additionally, in this extended version, we propose a method for supporting transparency in per-segment point cloud NPR pipelines based on alpha-to-coverage, which balances performance and visual quality. Overall, our approach supports real-time NPR of point clouds, allowing users to explore various artistic styles and visualization options interactively.