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

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.

Authors

Institutions

Publication Details

Journal
AI
Published
2026-11-18
DOI
https://doi.org/10.3390/ai7110490
Primary Topic
3D Shape Modeling and Analysis
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

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

Qinghai Jiang, Xinpeng Wang, Zhiyu Song, Jia Chen
2 links
AI
3D Shape Modeling and Analysis
article

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

Qinghai Jiang, Xinpeng Wang, Zhiyu Song, Jia Chen
article en

Abstract

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.

AIVol. 7(11)
Chinese Academy of Agricultural Sciences (CN), Ministry of Agriculture and Rural Affairs (CN), Nanjing Institute of Agricultural Mechanization (CN)
Chinese Academy of Agricultural Sciences, Agricultural Science and Technology Innovation Program
3D Shape Modeling and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon