[论文解读] Spherical-GOF: Geometry-Aware Panoramic Gaussian Opacity Fields for 3D Scene Reconstruction

Zhe Yang, Guoqiang Zhao+3 more|arXiv (Cornell University)|Mar 9, 2026

Advanced Vision and Imaging被引用 0

一句话总结

球面 GOF 在全向全景重建中引入球面射线空间高斯不透明度场框架，相较于基于投影的基线在几何一致性和深度质量方面有所提升，同时保持有竞争力的光度渲染。它使用基于球面的射线采样、保守边界和全景感知正则化。

ABSTRACT

Omnidirectional images are increasingly used in robotics and vision due to their wide field of view. However, extending 3D Gaussian Splatting (3DGS) to panoramic camera models remains challenging, as existing formulations are designed for perspective projections and naive adaptations often introduce distortion and geometric inconsistencies. We present Spherical-GOF, an omnidirectional Gaussian rendering framework built upon Gaussian Opacity Fields (GOF). Unlike projection-based rasterization, Spherical-GOF performs GOF ray sampling directly on the unit sphere in spherical ray space, enabling consistent ray-Gaussian interactions for panoramic rendering. To make the spherical ray casting efficient and robust, we derive a conservative spherical bounding rule for fast ray-Gaussian culling and introduce a spherical filtering scheme that adapts Gaussian footprints to distortion-varying panoramic pixel sampling. Extensive experiments on standard panoramic benchmarks (OmniBlender and OmniPhotos) demonstrate competitive photometric quality and substantially improved geometric consistency. Compared with the strongest baseline, Spherical-GOF reduces depth reprojection error by 57% and improves cycle inlier ratio by 21%. Qualitative results show cleaner depth and more coherent normal maps, with strong robustness to global panorama rotations. We further validate generalization on OmniRob, a real-world robotic omnidirectional dataset introduced in this work, featuring UAV and quadruped platforms. The source code and the OmniRob dataset will be released at https://github.com/1170632760/Spherical-GOF.

研究动机与目标

• Motivate accurate 3D reconstruction from omnidirectional (ERP) panoramas without projection-induced distortions.
• Extend Gaussian Opacity Fields (GOF) to spherical ray space for panorama rendering.
• Provide efficient, projection-consistent ray-Gaussian interactions with robust culling and distortion-aware filtering.
• Introduce geometry-focused regularizers to stabilize training and improve depth/normal coherence.
• Demonstrate generalization across public panoramic benchmarks and real-world omnidirectional robotic data.

提出的方法

• Render in spherical ray space by sampling GOFs directly on the unit sphere, avoiding projection-based distortions.
• Derive a conservative spherical bounding rule to enable fast ray–Gaussian culling.
• Apply a spherical filtering scheme to adapt Gaussian footprints to distortion across panoramic pixels.
• Inflate Gaussian scales with a panorama-aware isotropic filter radius to prevent sub-pixel footprints and maintain density consistency (with opacity adjustment).
• Augment photometric loss with panorama-aware geometric regularizers (depth-normal consistency and depth-jump terms) to stabilize geometry under ERP distortions.

实验结果

研究问题

• RQ1 How can GOF be adapted to spherical, omnidirectional panoramas without relying on planar projection approximations?
• RQ2 Does spherical ray-space GOF provide improved multi-view geometric consistency and depth accuracy on ERP panoramas compared to projection-based 3DGS methods?
• RQ3 Can panorama-aware regularizers mitigate high-frequency geometric artifacts while preserving photometric rendering quality?
• RQ4 Is the approach robust to global panorama rotations and transferable to different omnidirectional camera configurations (including annular/pseudo-annular setups)?
• RQ5 How well does the method generalize to real-world robotic omnidirectional data beyond public benchmarks?

主要发现

• Spherical-GOF achieves lower depth reprojection error (DRE) and higher cycle inlier ratio (CIR) across OmniBlender, OmniPhotos, and OmniRob datasets compared to strongest baselines.
• The method reduces DRE by substantial margins (e.g., up to 62.7% in certain OmniBlender settings) and increases CIR by about 22% relative to SPaGS on tested scenes.
• Qualitatively, depth maps are smoother with cleaner discontinuities and normals are more stable on planar regions, with reduced texture-induced artifacts.
• The approach shows strong rotation robustness, maintaining stability under large global panorama rotations where projection-based methods degrade.
• OmniRob evaluations indicate our method adapts to different omnidirectional camera configurations (UAV, annular, pseudo-annular) with favorable geometric consistency and reduced depth artifacts.