EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

GaussianFusion_AI: Urban 3-D Reconstruction and Label-Efficient Segmentation With Gaussian-Based Splatting

Thiago J. M. Baldivieso, Fabio Suim Chagas, Luiz C. P. R. Velho, Paulo F. F. Rosa

IEEE Geoscience and Remote Sensing Letters · 2026 · 23 · 8001605-8001605

DOI: 10.1109/lgrs.2026.3696540 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Geoscience and Remote Sensing Letters, 2026. Verfasst von Thiago J. M. Baldivieso und 3 weiteren Autor:innen. Die Arbeit nennt 19 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) Code 001 on Project.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This letter presents GaussianFusion_AI, a modular pipeline that integrates autonomous aerial vehicles (AAVs) RGB, thermal, and LiDAR data for 3-D urban reconstruction, latent-space segmentation, and continuous rendering via Gaussian-based splatting. In order to achieve dense multimodal point clouds and continuous views, the architecture combines multiview photogrammetry, consistent multisensor registration, Gaussian-weighted radiometric interpolation, and self-supervised latent encoding with anisotropic Gaussian-based splatting as its continuous rendering and representation primitive. The pipeline is evaluated on urban datasets from the city of Rio de Janeiro (Outeiro da Gloria and Quinta da Boa Vista park) and on the public USGS Darby benchmark. Quantitative results indicate stable label-efficient segmentations and average geometric errors of less than 0.6 m, assessed using silhouette coefficients, Davies-Bouldin index, and visual diagnostics. The proposed method, compared to a PointBERT baseline, produces coherent latent manifolds, demonstrates better compatibility with density-based clustering, and achieves competitive execution time. Furthermore, the Gaussian representations obtained showed preservation of spatial, spectral, and thermal gradients. This allows for differentiation in interpreting urban structures and suggests the operational viability of 3-D multimodal analysis in dense urban environments.

Förderung

  • the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) Code 001 on Project

Zitieren

APA

Baldivieso, T. J. M., Chagas, F. S., Velho, L. C. P. R., & Rosa, P. F. F. (2026). GaussianFusion_AI: Urban 3-D Reconstruction and Label-Efficient Segmentation With Gaussian-Based Splatting. IEEE Geoscience and Remote Sensing Letters, 23, 8001605-8001605. https://doi.org/10.1109/lgrs.2026.3696540

Vancouver

Baldivieso TJM, Chagas FS, Velho LCPR, Rosa PFF. GaussianFusion_AI: Urban 3-D Reconstruction and Label-Efficient Segmentation With Gaussian-Based Splatting. IEEE Geoscience and Remote Sensing Letters. 2026;23:8001605-8001605. doi:10.1109/lgrs.2026.3696540

BibTeX
@article{Baldivieso2026,
  author = {Baldivieso, Thiago J. M. and Chagas, Fabio Suim and Velho, Luiz C. P. R. and Rosa, Paulo F. F.},
  title = {GaussianFusion_AI: Urban 3-D Reconstruction and Label-Efficient Segmentation With Gaussian-Based Splatting},
  journal = {IEEE Geoscience and Remote Sensing Letters},
  year = {2026},
  volume = {23},
  pages = {8001605--8001605},
  doi = {10.1109/lgrs.2026.3696540}
}
Kostenlos starten →