StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering

StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering

July 2024 | LUKAS RADL* and MICHAEL STEINER*, Graz University of Technology, Austria MATHIAS PARGER, Huawei Technologies, Austria ALEXANDER WEINRAUCH, Graz University of Technology, Austria BERNHARD KERBL, TU Wien, Austria MARKUS STEINBERGER, Graz University of Technology, Austria and Huawei Technologies, Austria
The paper "StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering" addresses the issue of popping artifacts and view inconsistencies in 3D Gaussian Splatting (3DGS) during view rotation. The authors propose a novel hierarchical rasterization approach that effectively eliminates these artifacts by implementing a per-pixel sorting strategy. This method ensures accurate depth computation and reduces the number of Gaussians, leading to improved rendering performance and view consistency. The hierarchical approach includes per-tile depth calculation, local resorting, and a hierarchical rendering pipeline with load balancing and culling techniques. The results show that the proposed method achieves comparable image quality to 3DGS while reducing popping artifacts and improving view consistency. The authors also provide a user study and quantitative evaluations to validate the effectiveness of their approach. The renderer and optimizations are publicly available on GitHub.The paper "StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering" addresses the issue of popping artifacts and view inconsistencies in 3D Gaussian Splatting (3DGS) during view rotation. The authors propose a novel hierarchical rasterization approach that effectively eliminates these artifacts by implementing a per-pixel sorting strategy. This method ensures accurate depth computation and reduces the number of Gaussians, leading to improved rendering performance and view consistency. The hierarchical approach includes per-tile depth calculation, local resorting, and a hierarchical rendering pipeline with load balancing and culling techniques. The results show that the proposed method achieves comparable image quality to 3DGS while reducing popping artifacts and improving view consistency. The authors also provide a user study and quantitative evaluations to validate the effectiveness of their approach. The renderer and optimizations are publicly available on GitHub.
Reach us at info@study.space