Research map / 3D vision / Gaussian splatting
Splatter-360: Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images
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 Splatter-360, an end-to-end generalizable 3D Gaussian splatting framework for wide-baseline panoramic images that uses a spherical cost volume, a bi-projection (ERP + cubemap) encoder with monocular depth features, and cross-view attention.
Problem
Real-time novel view synthesis from sparse, wide-baseline 360° panoramic images, where existing 3DGS and NeRF methods overfit training views and perspective-based pipelines suffer from distortion and cube-map seam issues.
Approach
Performs multi-view matching directly in the spherical (equirectangular) domain via a spherical sweep to build a spherical cost volume, and fuses ERP and cubemap features with DepthAnything V2 monocular depth features in a 3D-aware bi-projection encoder to predict pixel-aligned Gaussians.
Key results
- On HM3D, Splatter-360 reaches 28.293 PSNR, 0.875 SSIM, 0.155 LPIPS, 2.662 dB above PanoGRF.
- On Replica, Splatter-360 reaches 29.888 PSNR, 0.924 SSIM, 0.097 LPIPS, 1.968 dB above PanoGRF.
- Removing the spherical cost volume drops PSNR by 5.271 dB on Replica and 2.263 dB on HM3D.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Splatter-360 instead of MVSplat† | Panoramic novel-view synthesis · HM3D | PSNR: 27.179 -> 28.293 | Helped |
| Use Splatter-360 instead of MVSplat† | Panoramic novel-view synthesis · Replica | PSNR: 28.399 -> 29.888 | Helped |
| Add spherical cost volume (ablation) | Splatter-360 · Replica | PSNR: 23.850 -> 29.121 | Helped |
| Add spherical cost volume (ablation) | Splatter-360 · HM3D | PSNR: 25.224 -> 27.487 | Helped |
| Add cross-view attention (ablation) | Splatter-360 · Replica | PSNR: 28.217 -> 29.121 | Helped |
| Add cross-view attention (ablation) | Splatter-360 · HM3D | PSNR: 26.918 -> 27.487 | Helped |
| Add ERP encoder (ablation) | Splatter-360 · Replica | PSNR: 26.985 -> 29.121 | Helped |
| Add ERP encoder (ablation) | Splatter-360 · HM3D | PSNR: 25.905 -> 27.487 | Helped |
| Add CP encoder (ablation) | Splatter-360 · Replica | PSNR: 28.673 -> 29.121 | Helped |
| Add CP encoder (ablation) | Splatter-360 · HM3D | PSNR: 27.277 -> 27.487 | Helped |
| Add monocular depth features (ablation) | Splatter-360 · Replica | PSNR: 28.654 -> 29.121 | Helped |
| Add monocular depth features (ablation) | Splatter-360 · HM3D | PSNR: 27.380 -> 27.487 | Helped |
Surprising findings
- Perspective-based baselines (HiSplat, DepthSplat, MVSplat) pretrained on perspective data generalize poorly to cube-map panoramic inputs, with PSNR around 17-19 dB.
Limitations
- Evaluated only on two synthetic indoor datasets (HM3D and Replica).
- Perspective baselines were tested with official Re10K checkpoints due to GPU memory limits, so comparisons are not like-for-like.
Experimental setup
Datasets: HM3D Replica
Benchmarks: HM3D Replica
Metrics: PSNR SSIM LPIPS Abs Diff Abs Rel RMSE delta<1.25
Baselines: PanoGRF MVSplat DepthSplat HiSplat PixelSplat
Builds on: MVSplat DepthSplat UniMatch PanoGRF Depth Anything V2
Closest papers by meaning
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.90)
- PanSplat: 4K Panorama Synthesis with Feed-Forward Gaussian Splatting (CVPR 2025 · similarity 0.90)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.89)
- MS-GS: Multi-Appearance Sparse-View 3D Gaussian Splatting in the Wild (NeurIPS 2025 · similarity 0.89)
- Unified Panoramic Geometry Estimation via Multi-View Foundation Models (NeurIPS 2026 · similarity 0.89)
- MonoSplat: Generalizable 3D Gaussian Splatting from Monocular Depth Foundation Models (CVPR 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.