OGSR: point-centric online Gaussian surface refinement for monocular 2D Gaussian SLAM
Applied Optics · 20.11.2026 · 65 (33) · 11767
DOI: 10.1364/ao.616617 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Optics, online erschienen am 12.11.2026. Verfasst von Jianpeng Xu und 2 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben.
Abstract (Originalsprache)
Monocular Gaussian SLAM unifies camera tracking, explicit mapping, and high-quality rendering, yet its persistent map remains sensitive to geometric errors introduced by monocular depth estimation. Once biased geometry is repeatedly written into Gaussian primitives, photometric optimization may preserve plausible appearance without restoring displaced centers to the underlying physical surface, leading to surface thickening, duplicated geometry, and boundary smearing. We present online Gaussian surface refinement (OGSR), a point-centric surface-refinement network that directly corrects the currently visible Gaussian centers rather than predicting another dense depth map. OGSR combines pixel-aligned image evidence—comprising the current RGB image, a map-rendered depth hypothesis, a causal monocular inverse-depth observation, and their scale-aligned disagreement residual—with local 3D geometry encoded from each Gaussian center and its neighboring primitives. The fused representation predicts a bounded three-dimensional displacement and a reliability score for each visible Gaussian. Repeated observations of the same persistent Gaussian are transformed into a common map frame and integrated through reliability-weighted identity fusion for online center refinement. On replica and held-out ScanNet++ scenes, OGSR reduces the mean surface error from 0.136 to 0.027 m and from 0.294 to 0.068 m, respectively, while preserving more than 98% of initially accurate Gaussians. On replica, OGSR also improves PSNR from 25.16 to 26.43 dB with approximately 8% additional scene-level runtime. The results show that point-centric refinement improves persistent Gaussian geometry while preserving rendering quality with moderate additional computation.
Zitieren
APA
Xu, J., Wang, Y., & Bu, F. (2026). OGSR: point-centric online Gaussian surface refinement for monocular 2D Gaussian SLAM. Applied Optics, 65(33), 11767. https://doi.org/10.1364/ao.616617
Vancouver
Xu J, Wang Y, Bu F. OGSR: point-centric online Gaussian surface refinement for monocular 2D Gaussian SLAM. Applied Optics. 2026;65(33):11767. doi:10.1364/ao.616617
BibTeX
@article{Xu2026,
author = {Xu, Jianpeng and Wang, Yifan and Bu, Fanliang},
title = {OGSR: point-centric online Gaussian surface refinement for monocular 2D Gaussian SLAM},
journal = {Applied Optics},
year = {2026},
volume = {65},
number = {33},
pages = {11767},
doi = {10.1364/ao.616617}
}
EUVIMED