Research map / Image generation / Super-resolution
Spherical Pseudo-Cylindrical Representation for Omnidirectional Image Super-resolution
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 a latitude-adaptive pseudo-cylindrical representation for ERP images (per-row upsampling optimized by search), a viewport-based loss computed on the 3D sphere, and a recursive CNN-Transformer SR backbone for omnidirectional image super-resolution.
Problem
Super-resolving equirectangular projection (ERP) omnidirectional images is hard for 2D SR methods because ERP sampling density is inhomogeneous and distorted across latitude.
Approach
Replace ERP with a per-latitude adaptive pseudo-cylindrical representation whose row widths are optimized by a search using a pre-trained 2D SR model, and train with a loss computed on rendered viewports of the 3D sphere instead of the ERP plane, so the representation is model-agnostic to off-the-shelf SR networks.
Key results
- On ODI-SR at x8, the method reaches 24.72 WS-PSNR vs 24.37 for the second-best (Swin2SR/SphereSR), a 0.32 dB gain.
- At x16 on ODI-SR, WS-PSNR improves by 0.34 dB over the second-best method (22.90 vs 22.56).
- Using LAPR improves SwinIR, ELAN and Swin2SR at x8 on ODI-SR (e.g., SwinIR 24.34 to 24.48).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed LAPR instead of original ERP (ablation) | Ours · ODI-SR ×8 | WS-PSNR: 24.48 -> 24.72 | Helped |
| Use the proposed LAPR instead of original ERP (ablation) | Ours · ODI-SR ×8 | WS-SSIM: 0.6688 -> 0.6886 | Helped |
| Use the proposed LAPR instead of sinusoidal projection (ablation) | Ours · ODI-SR ×8 | WS-PSNR: 24.56 -> 24.72 | Helped |
| Use the proposed LAPR instead of sinusoidal projection (ablation) | Ours · ODI-SR ×8 | WS-SSIM: 0.6791 -> 0.6886 | Helped |
| Apply LAPR (ablation) | SwinIR · ODI-SR ×8 | WS-PSNR: 24.34 -> 24.48 | Helped |
| Apply LAPR (ablation) | ELAN · ODI-SR ×8 | WS-PSNR: 24.35 -> 24.50 | Helped |
| Apply LAPR (ablation) | Swin2SR · ODI-SR ×8 | WS-PSNR: 24.37 -> 24.52 | Helped |
| Use viewport-based loss instead of LMAE (ablation) | Ours · ODI-SR ×16 | WS-PSNR: 22.80 -> 22.90 | Helped |
| Use viewport-based loss instead of LMAE (ablation) | Ours · ODI-SR ×16 | WS-SSIM: 0.6310 -> 0.6480 | Helped |
| Replace RSTB backbone with 32 residual blocks (ablation) | Ours · ODI-SR ×8 | WS-PSNR: 24.72 -> 24.62 | Hurt |
| Replace RSTB backbone with 32 residual blocks (ablation) | Ours · ODI-SR ×8 | WS-SSIM: 0.6886 -> 0.6779 | Hurt |
| Use the proposed method instead of Swin2SR | Ours · ODI-SR ×8 | WS-PSNR: 24.37 -> 24.72 | Helped |
Limitations
- Latitude search depends on a pre-trained 2D SR model and a quality measure, adding optimization cost
- Viewport loss and representation evaluated only on ODI-SR and SUN 360 datasets with bicubic degradation
Experimental setup
Datasets: ODI-SR
Benchmarks: ODI-SR SUN 360 Panorama
Metrics: WS-PSNR WS-SSIM
Baselines: EDSR DRN HAN NLSN SwinIR ELAN Swin2SR 360-SS LAU-Net SphereSR
Builds on: SwinIR LAU-Net EDSR sinusoidal projection
Closest papers by meaning
- Omnidirectional Image Super-resolution via Bi-projection Fusion (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 0.90)
- ODGS: 3D Scene Reconstruction from Omnidirectional Images with 3D Gaussian Splattings (NeurIPS 2024 · similarity 0.87)
- SC-OmniGS: Self-Calibrating Omnidirectional Gaussian Splatting (ICLR 2025 · similarity 0.86)
- Elite360D: Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion (CVPR 2024 · similarity 0.84)
- SphereUFormer: A U-Shaped Transformer for Spherical 360 Perception (CVPR 2025 · similarity 0.84)
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.