EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Geometry-Aware Initialization with Footprint-Calibrated Opacity for 3D Gaussian Splatting

Xinpeng Wang, Qinghai Jiang, Jia Chen, Zhiyu Song

AI · 18.11.2026 · 7 (11) · 490

DOI: 10.3390/ai7110490 ↗

Kurz gefasst

Fachartikel in der Zeitschrift AI, online erschienen am 18.11.2026. Verfasst von Xinpeng Wang und 3 weiteren Autor:innen. Die Arbeit nennt 32 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Agricultural Science and Technology Innovation Program, Chinese Academy of Agricultural Sciences.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Reconstructing fine structures with three-dimensional Gaussian splatting (3DGS) requires adequate local geometric support and appropriate opacity accumulation during densification. Sparse or uneven structure-from-motion (SfM) point clouds can leave fine structures insufficiently represented at initialization, while directly inheriting the parent opacity during cloning and splitting can alter the combined projected alpha response of the resulting Gaussians. To address these two limitations, this study integrates geometry-aware safe initialization (GA-SafeInit) and footprint-based opacity calibration (FootprintOpacity) into the 3DGS framework. GA-SafeInit uses local point distributions to add geometric support in line-like, surface-like, and spatially distributed neighborhoods under constraints on the number, positions, and scales of the added Gaussians. FootprintOpacity initializes a shared opacity for each clone or split pair by minimizing the discrepancy between its combined projected alpha footprint and that of the parent primitive. The method was evaluated against 3DGS, AbsGS, and Instant Neural Graphics Primitives (Instant-NGP) on eleven scenes, comprising seven self-collected and four public scenes, using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS) on held-out views. Compared with 3DGS, the proposed method increased PSNR in all eleven scenes by 0.053–1.054 dB and increased SSIM in ten scenes, while LPIPS improvements remained scene-dependent. Across the eleven scenes, its mean SSIM was 0.7709, compared with 0.7654 for 3DGS, 0.7486 for AbsGS, and 0.6628 for Instant-NGP. Ablation experiments on Chicken, Cactus2, and Bonsai showed that the combined method achieved the highest PSNR in two scenes, while the metric-wise contributions varied across scenes. The method has potential applications in image-based 3D digitization of biological specimens, plants, and small objects with fine structures.

Förderung

  • Agricultural Science and Technology Innovation Program
  • Chinese Academy of Agricultural Sciences

Zitieren

APA

Wang, X., Jiang, Q., Chen, J., & Song, Z. (2026). Geometry-Aware Initialization with Footprint-Calibrated Opacity for 3D Gaussian Splatting. AI, 7(11), 490. https://doi.org/10.3390/ai7110490

Vancouver

Wang X, Jiang Q, Chen J, Song Z. Geometry-Aware Initialization with Footprint-Calibrated Opacity for 3D Gaussian Splatting. AI. 2026;7(11):490. doi:10.3390/ai7110490

BibTeX
@article{Wang2026,
  author = {Wang, Xinpeng and Jiang, Qinghai and Chen, Jia and Song, Zhiyu},
  title = {Geometry-Aware Initialization with Footprint-Calibrated Opacity for 3D Gaussian Splatting},
  journal = {AI},
  year = {2026},
  volume = {7},
  number = {11},
  pages = {490},
  doi = {10.3390/ai7110490}
}
Kostenlos starten →