Spherical-GOF: Geometry-Aware Panoramic Gaussian Opacity Fields for 3D Scene Reconstruction
New method slashes depth error by 57% and improves geometric consistency by 21% over top baselines.
A research team led by Zhe Yang and Kailun Yang has introduced Spherical-GOF, a novel framework that solves a key challenge in 3D computer vision: accurately reconstructing 3D scenes from omnidirectional, 360-degree images. Traditional 3D Gaussian Splatting (3DGS) methods are designed for standard perspective cameras and perform poorly on panoramic inputs, introducing severe distortion and geometric errors. Spherical-GOF rethinks the rendering process by performing Gaussian Opacity Fields (GOF) ray sampling directly in spherical ray space on a unit sphere. This fundamental shift ensures consistent interactions between light rays and the 3D Gaussians representing the scene, regardless of the panoramic distortion.
To make this spherical approach practical, the team developed two key innovations: a conservative spherical bounding rule for rapidly culling irrelevant Gaussians during ray casting, and a spherical filtering scheme that dynamically adapts Gaussian footprints to match the varying pixel density in a panoramic image. These technical advances make the rendering process both efficient and robust. The system was rigorously tested on established panoramic benchmarks like OmniBlender and OmniPhotos, where it demonstrated superior photometric quality and a major leap in geometric accuracy. Quantitative results show a 57% reduction in depth reprojection error and a 21% improvement in cycle inlier ratio compared to the strongest prior methods.
The team further validated Spherical-GOF's real-world applicability by introducing and testing on a new dataset called OmniRob, captured from UAV and quadruped robotic platforms. This proves the method's robustness for applications in autonomous navigation and robotics, where accurate 3D understanding from omnidirectional cameras is critical. The release of the source code and the OmniRob dataset will provide the community with essential tools to advance research in panoramic 3D reconstruction, bridging a significant gap between cutting-edge neural rendering and practical robotic vision systems.
- Reduces depth reprojection error by 57% compared to top baselines on OmniBlender/OmniPhotos benchmarks.
- Introduces spherical ray sampling and bounding for consistent 3D Gaussian rendering on 360-degree images.
- Validated on new OmniRob dataset from UAV/quadruped robots, proving real-world robotic vision utility.
Why It Matters
Enables robots and VR systems to build accurate 3D maps from standard 360 cameras, critical for navigation and simulation.