Research map / Image generation / Super-resolution
Omnidirectional Image Super-resolution via Bi-projection Fusion
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 BPOSR, a dual-branch network that fuses equirectangular (ERP) and cubemap (CMP) features with horizontal-striped attention, perspective-shift attention, and block attention fusion for omnidirectional image super-resolution.
Problem
Omnidirectional image (ODI) super-resolution is suboptimal when conventional 2D SR is applied to ERP, which suffers distortion, while CMP reduces distortion but has limited field of view and discontinuities.
Approach
Jointly process ERP and CMP branches with tailored attention per projection (horizontal striped windows for ERP, perspective-shift rolling on CMP) and fuse them with a block attention fusion module across depths and projections.
Key results
- On ODI-SR at x4, BPOSR reaches 26.95 WS-PSNR, 0.71 dB above the best SISR baseline DRN (per paper text).
- At x8 on ODI-SR, BPOSR achieves 24.61 WS-PSNR, 0.29 dB over DRN and above OSRT (24.53).
- On SUN360 at x4, BPOSR gains 0.12 dB WS-PSNR over OSRT (27.59 vs 27.47).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BPOSR instead of OSRT | ODI-SR super-resolution · ODI-SR ×4 | WS-PSNR: 26.89 -> 26.95 | Helped |
| use BPOSR instead of OSRT | ODI-SR super-resolution · ODI-SR ×8 | WS-PSNR: 24.53 -> 24.61 | Helped |
| use BPOSR instead of OSRT | ODI-SR super-resolution · ODI-SR ×16 | WS-PSNR: 22.69 -> 22.72 | Helped |
| use BPOSR instead of OSRT | SUN360 super-resolution · SUN360 ×4 | WS-PSNR: 27.47 -> 27.59 | Helped |
| use BPOSR instead of OSRT | SUN360 super-resolution · SUN360 ×8 | WS-PSNR: 24.38 -> 24.47 | Helped |
| use BPOSR instead of OSRT | SUN360 super-resolution · SUN360 ×16 | WS-PSNR: 22.13 -> 22.16 | Helped |
| use bi-projection (ERP and CMP branches) instead of CMP in both branches (ablation) | BPOSR · ODI-SR ×8 | WS-PSNR: 24.30 -> 24.61 | Helped |
| use bi-projection (ERP and CMP branches) instead of ERP in both branches (ablation) | BPOSR · ODI-SR ×8 | WS-PSNR: 24.47 -> 24.61 | Helped |
| use 4 × 16 horizontal attention windows instead of 8 × 8 (ablation) | BPOSR ERP branch · ODI-SR ×8 | WS-PSNR: 24.48 -> 24.61 | Helped |
| add perspective shift with rotation ratio 3 instead of removing it (ablation) | BPOSR CMP branch · ODI-SR ×8 | WS-PSNR: 24.49 -> 24.61 | Helped |
| use BAFM instead of mean fusion (ablation) | BPOSR fusion module · ODI-SR ×8 | WS-PSNR: 24.53 -> 24.61 | Helped |
| use BAFM instead of 1×1 convolution fusion (ablation) | BPOSR fusion module · ODI-SR ×8 | WS-PSNR: 24.55 -> 24.61 | Helped |
Limitations
- Evaluated only on ODI-SR and SUN360 datasets with synthetic downsampling.
- Compute hardware and training cost not reported in main text.
Experimental setup
Datasets: ODI-SR SUN360
Benchmarks: ODI-SR SUN360
Metrics: WS-PSNR WS-SSIM
Baselines: SRCNN VDSR LapSRN MemNet MSRN EDSR D-DBPN RCAN DRN 360-SS LAU-Net SphereSR
Builds on: SwinIR ELAN LAM
Closest papers by meaning
- Spherical Pseudo-Cylindrical Representation for Omnidirectional Image Super-resolution (AAAI 2024 · similarity 0.91)
- Fast Omni-Directional Image Super-Resolution: Adapting the Implicit Image Function with Pixel and Semantic-Wise Spherical Geometric Priors (AAAI 2025 · similarity 0.86)
- Elite360D: Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion (CVPR 2024 · similarity 0.85)
- ODGS: 3D Scene Reconstruction from Omnidirectional Images with 3D Gaussian Splattings (NeurIPS 2024 · similarity 0.83)
- Occlusion-Embedded Hybrid Transformer for Light Field Super-Resolution (AAAI 2025 · similarity 0.82)
- SC-OmniGS: Self-Calibrating Omnidirectional Gaussian Splatting (ICLR 2025 · similarity 0.82)
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.