Disambiguating 2D-3D Correspondences in Gaussian Splatting-based Feature Fields for Visual Localization
Miso Lee Sangeek Hyun Yerim Jeon Jae-Pil Heo · 2026
Read Disambiguating 2D-3D Correspondences in Gaussian Splatting-based Feature Fields for Visual Localization by Miso Lee Sangeek Hyun Yerim Jeon Jae-Pil Heo (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-2605-07351.txt.
Description
Miso Lee and three co-authors fix a matching problem that limits camera localization from Gaussian splatting feature fields. Fields tuned for looks create many-to-one links between image pixels and 3D points, destabilizing pose estimation, and spawn extra Gaussians lacking consistency across views. Their SplitGS-Loc framework attacks both issues using Gaussian attributes directly.
Each Gaussian is split into smaller ones, turning ambiguous mappings into precise one-to-one correspondences. Composition weights then select Gaussians that contribute strongly and consistently across views, aggregating discriminative features. The resulting compact fields stabilize pose solving and reach leading results on localization benchmarks, without per-scene training or iterative pose refinement.
Catalog record
| CREATOR | Miso Lee Sangeek Hyun Yerim Jeon Jae-Pil Heo |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.07351