Reproduce “360-GS: Layout-guided Panoramic Gaussian Splatting For Indoor Roaming” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, Yanwen Guo
Researchers at Nanjing University developed 360-GS, a layout-guided panoramic Gaussian splatting framework for real-time indoor roaming, which extends 3D Gaussian Splatting to handle sparse panoramic inputs. This method significantly improves PSNR by 5.04 dB over baseline 3D-GS for 4-view inputs while maintaining real-time 60 FPS rendering.
Key insights
- Directly applying existing 3D Gaussian Splatting to panoramic images faces challenges due to projection distortion and unreliable initialization from sparse data.
- Leveraging strong structural priors, such as room layouts and monocular depth, effectively guides 3D Gaussian initialization and optimization, particularly in data-scarce indoor environments.
- A decomposed projection strategy, involving projection to a tangent plane followed by mapping to spherical coordinates, enables efficient and accurate 360° Gaussian splatting.
- Integrating scene priors through both initialization and a regularization loss during optimization significantly enhances geometric fidelity and reduces artifacts in reconstructed indoor scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.