Research map / Image generation / Super-resolution
Fast Omni-Directional Image Super-Resolution: Adapting the Implicit Image Function with Pixel and Semantic-Wise Spherical Geometric Priors
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 FAOR, a fast arbitrary-scale omnidirectional image super-resolution model that adapts LIIF-style implicit image functions to ERP images using affine feature modulation with pixel-wise and semantic distortion priors plus a geodesic-based latent resampling.
Problem
Omni-directional image super-resolution must handle non-uniform oversampling from ERP projection, but existing spherical convolution and reprojection methods are complex and slow at inference.
Approach
Incorporates spherical geometry into an implicit image function without complex operators: affine transformations of ERP latent features guided by a pixel-wise stretching-ratio map and an instance segmentation map (ATFM), and a geodesic-based spherical linear interpolation over latent vectors (resampling-then-representation) so the implicit function runs once per HR pixel.
Key results
- On ODI-SR at x8, FAOR reaches 24.58 WS-PSNR vs 24.24 for OSRT (averaging ~0.39 dB gain over OSRT across scales on ODI-SR).
- On SUN360 Panorama, FAOR improves over OSRT by 0.59 dB average WS-PSNR and 0.017 WS-SSIM.
- Ablation: removing the geodesic-based resampling drops WS-PSNR by 0.12 dB at x2 scale... (0.17 dB at x2 vs 30.12 to 29.95).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FAOR instead of OSRT | ODI-SR and SUN360 Panorama · ODI-SR and SUN360 Panorama ×2, ×4, ×8, and ×16 | WS-PSNR: +0.39 dB on ODI-SR; +0.59 dB on SUN360 | Helped |
| Use FAOR instead of OSRT | ODI-SR and SUN360 Panorama · ODI-SR and SUN360 Panorama ×2, ×4, ×8, and ×16 | WS-SSIM: +0.006 on ODI-SR; 0.017 on SUN 360 | Helped |
| Add pixel-wise stretching ratio map Md (ablation) | FAOR SAFE · ODI-SR ×8 | WS-PSNR: 24.47 -> 24.58 | Helped |
| Add pixel-wise stretching ratio map Md (ablation) | FAOR SAFE · ODI-SR ×8 | WS-SSIM: 0.6575 -> 0.6581 | Helped |
| Add pixel-wise stretching ratio map Md and instance segmentation map Ms (ablation) | FAOR SAFE · ODI-SR ×8 | WS-PSNR: 24.35 -> 24.58 | Helped |
| Add pixel-wise stretching ratio map Md and instance segmentation map Ms (ablation) | FAOR SAFE · ODI-SR ×8 | WS-SSIM: 0.6502 -> 0.6581 | Helped |
| Use geodesic-based spherical resampling instead of LIIF normalized interpolation (ablation) | FAOR SGIF · ODI-SR ×8 | WS-PSNR: 24.42 -> 24.58 | Helped |
| Use geodesic-based spherical resampling instead of LIIF normalized interpolation (ablation) | FAOR SGIF · ODI-SR ×8 | WS-SSIM: 0.6533 -> 0.6581 | Helped |
Surprising findings
- Geodesic resampling gains are small at x16 (22.78 with and without sphere resampling), and at x8 WS-SSIM of OSRT/LAU-Net is higher than FAOR.
Limitations
- Relies on an offline instance segmentation map from an external model, adding preprocessing dependence.
- Gains over OSRT are not uniform across metrics and scales (e.g., WS-SSIM at x8 and x16 is lower than some baselines).
Experimental setup
Datasets: ODI-SR SUN360
Benchmarks: ODI-SR SUN360 Panorama
Metrics: WS-PSNR WS-SSIM
Baselines: Cubic LIIF OPE-SR 360-SS LAU-Net OSRT
Builds on: LIIF OSRT LAU-Net OPE-SR
Closest papers by meaning
- Spherical Pseudo-Cylindrical Representation for Omnidirectional Image Super-resolution (AAAI 2024 · similarity 0.90)
- Omnidirectional Image Super-resolution via Bi-projection Fusion (AAAI 2024 · similarity 0.86)
- Arbitrary-Scale Image Generation and Upsampling using Latent Diffusion Model and Implicit Neural Decoder (CVPR 2024 · similarity 0.85)
- Edge-SD-SR: Low Latency and Parameter Efficient On-device Super-Resolution with Stable Diffusion via Bidirectional Conditioning (CVPR 2025 · similarity 0.84)
- Occlusion-Embedded Hybrid Transformer for Light Field Super-Resolution (AAAI 2025 · similarity 0.84)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.83)
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.