Pi-GS: Sparse-View Gaussian Splatting with Dense 𝝅 3 \bm{\pi}^{3} Initialization
Manuel Hofer, Markus Steinberger, Thomas Köhler · 2026
Read Pi-GS: Sparse-View Gaussian Splatting with Dense 𝝅 3 \bm{\pi}^{3} Initialization by Manuel Hofer, Markus Steinberger, Thomas Köhler (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-03327.txt.
Description
Rendering new views from very few photos is made robust by Manuel Hofer, together with 2 colleagues, by seeding the scene with a dense learned point cloud. Three-dimensional Gaussian splatting renders in real time and trains quickly, but it depends on accurate camera poses and good initial point clouds. Traditional structure-from-motion pipelines often fail when views are sparse, and learned alternatives can be sensitive to pose or depth errors in difficult scenes.
Their method uses a reference-free point cloud estimation network for dense initialization. Regularization then mitigates geometric inaccuracies through uncertainty-guided depth supervision, normal consistency, and depth warping losses. Experiments across four standard datasets, including Tanks and Temples and MipNeRF360, reached state-of-the-art performance. The approach targets practical capture settings where photographers cannot circle a scene densely.
Catalog record
| CREATOR | Manuel Hofer, Markus Steinberger, Thomas Kƶhler |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-03327