EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

LumeGS: Robust Low‐Light 3D Gaussian Splatting via Logarithmic Structural Optimization

Tenghui Yao, Ying Zhao, Ming Li, Boyao Jia

IET Computer Vision · 2026-01 · 20 (1)

DOI: 10.1049/cvi2.70097 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IET Computer Vision, online erschienen am 05.11.2026. Verfasst von Tenghui Yao und 3 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

3D Gaussian Splatting (3DGS) achieves real‐time novel view synthesis, but its optimization becomes unstable when the input views are captured under extreme low‐light conditions. Dark images provide weak and noisy structural supervision, which may lead to insufficient Gaussian densification, dead‐black renderings, or noise‐induced floaters. We propose LumeGS, a lightweight training‐time optimization framework for low‐light 3DGS. Instead of introducing an auxiliary neural image signal processor or an illumination network, LumeGS modifies the optimization objective using logarithmic radiance supervision and gamma‐space edge‐guided structural regularisation. The logarithmic term increases the relative contribution of dark‐region structural discrepancies, while the edge‐guided term suppresses noise‐driven densification by encouraging coherent boundary reconstruction. At inference time, LumeGS preserves the standard 3DGS rasterisation pipeline and introduces no additional learnt inference module; only an optional pixel‐wise gamma mapping is used for display and is included in our runtime analysis. We evaluate LumeGS on the Gaussian‐DK benchmark and the LOM dataset under different capture protocols. Experiments compare LumeGS with vanilla 3DGS, simple gamma/brightness baselines, enhancement‐preprocessed 3DGS, ablated variants, and reported low‐light/challenging‐light view synthesis methods where available. The results show that LumeGS improves perceptual quality and structural fidelity under the evaluated low‐light settings while retaining real‐time rendering efficiency.

Zitieren

APA

Yao, T., Zhao, Y., Li, M., & Jia, B. (2026). LumeGS: Robust Low‐Light 3D Gaussian Splatting via Logarithmic Structural Optimization. IET Computer Vision, 20(1). https://doi.org/10.1049/cvi2.70097

Vancouver

Yao T, Zhao Y, Li M, Jia B. LumeGS: Robust Low‐Light 3D Gaussian Splatting via Logarithmic Structural Optimization. IET Computer Vision. 2026;20(1). doi:10.1049/cvi2.70097

BibTeX
@article{Yao2026,
  author = {Yao, Tenghui and Zhao, Ying and Li, Ming and Jia, Boyao},
  title = {LumeGS: Robust Low‐Light 3D Gaussian Splatting via Logarithmic Structural Optimization},
  journal = {IET Computer Vision},
  year = {2026},
  volume = {20},
  number = {1},
  doi = {10.1049/cvi2.70097}
}
Kostenlos starten →