Research map / 3D vision / Gaussian splatting
ODGS: 3D Scene Reconstruction from Omnidirectional Images with 3D Gaussian Splattings
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ODGS, a CUDA rasterizer that projects 3D Gaussians onto per-Gaussian tangent planes of the unit sphere and maps them to equirectangular images, enabling fast 3DGS reconstruction from omnidirectional images.
Problem
Standard perspective 3DGS rasterization distorts omnidirectional (360-degree) images, and existing NeRF-based 360 reconstruction methods are slow to train and render.
Approach
For each Gaussian, define a tangent plane to the unit sphere perpendicular to the ray through its center, rasterize the Gaussian with a perspective local affine approximation there, then apply the equirectangular distortion (sec-theta stretch) and scaling to map it into the omnidirectional image; add an elevation-dependent densification gradient threshold.
Key results
- At 10 min on OmniBlender, ODGS reaches 32.76 PSNR / 0.9234 SSIM / 0.0469 LPIPS vs 28.29 PSNR for EgoNeRF and 29.36 for 3DGS(P).
- ODGS renders an omnidirectional image in about 0.028 s versus roughly 23-24 s for EgoNeRF and about 62 s for NeRF(P) per the reported timings.
- On 360VO at 100 min, ODGS achieves 26.68 PSNR and 0.1264 LPIPS, outperforming all baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ODGS instead of the strongest PSNR baseline at 10 minutes | 3D scene reconstruction · OmniBlender, Ricoh360, OmniPhotos, 360Roam, OmniScenes, 360VO | PSNR: 32.76 vs 29.36; 24.94 vs 25.12; 26.24 vs 25.61; 21.08 vs 20.45; 24.42 vs 23.61; 24.63 vs 22.87 | Mixed |
| use ODGS instead of the strongest PSNR baseline at 100 minutes | 3D scene reconstruction · OmniBlender, Ricoh360, OmniPhotos, 360Roam, OmniScenes, 360VO | PSNR: 33.05 vs 30.89; 26.27 vs 25.49; 27.04 vs 26.90; 20.85 vs 21.18; 24.51 vs 24.76; 26.68 vs 23.78 | Mixed |
| increase cubemap perspective training views from 6 to 18 (ablation) | 3DGS (P) · OmniBlender, Ricoh360, OmniPhotos, 360Roam, OmniScenes, 360VO | PSNR: 29.36 -> 27.85; 25.12 -> 24.76; 25.61 -> 24.93; 20.17 -> 20.88; 23.61 -> 24.00; 22.87 -> 23.22 (10 min); 21.19 -> 24.56; | Mixed |
| use dynamic elevation-dependent densification instead of static densification (ablation) | ODGS · qualitative road-scene reconstruction | reconstruction quality | Helped |
Surprising findings
- Perspective 3DGS on cubemap faces overfits quickly and degrades after 100 minutes (e.g., OmniBlender 3DGS(P) falls from 29.36 to 21.19 PSNR).
- Ricoh360 at 10 min: 3DGS(P) achieves higher PSNR (25.12) than ODGS (24.94).
Limitations
- Local affine (first-order) approximation of the equirectangular projection introduces errors, since straight lines become curves in ERP.
- Assumes the projected covariance is small enough to ignore the tangent-plane-to-sphere gap; relies on the split rule to keep Gaussians small.
- Rasterizer is a manually implemented CUDA kernel; the authors note the 3DGS-style pipeline is approximate rather than exact.
Experimental setup
Datasets: OmniBlender Ricoh360 OmniPhotos 360Roam OmniScenes 360VO
Benchmarks: OmniBlender Ricoh360 OmniPhotos 360Roam OmniScenes 360VO
Metrics: PSNR SSIM LPIPS rendering time
Baselines: NeRF(P) 3DGS(P) TensoRF EgoNeRF
Builds on: 3D Gaussian splatting EWA splatting NeRF EgoNeRF
Closest papers by meaning
- OmniSplat: Taming Feed-Forward 3D Gaussian Splatting for Omnidirectional Images with Editable Capabilities (CVPR 2025 · similarity 0.91)
- SC-OmniGS: Self-Calibrating Omnidirectional Gaussian Splatting (ICLR 2025 · similarity 0.90)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.89)
- 6DGS: Enhanced Direction-Aware Gaussian Splatting for Volumetric Rendering (ICLR 2025 · similarity 0.88)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.