DSOSR: Degradation-Separated Real-World Omnidirectional Image Super-Resolution Via Projection Fusion Representation

ICLR 2026 Conference Submission357 Authors

Keywords: omnidirectional image, super-resolution, real-world, projection representation

Abstract: With the growing demand for immersive visual experience in virtual and augmented reality, high-resolution (HR) and high-quality (HQ) omnidirectional images (ODIs) are becoming increasingly essential. However, the limited capabilities of capturing device and transmission bandwidth constrain ODI resolution, hindering the rendering of fine 360$^{\circ}$ details. This challenge is further compounded by unknown real-world degradations and geometric distortions, which severely degrade ODI visual quality. Although real-world super-resolution (Real-SR) has been widely studied, existing degradation simulations fail to accurately characterize the complex imaging pipeline of ODIs. In practice, ODIs are usually collected by fisheye cameras and projected from the sphere to a plane through Equirectangular Projection (ERP), which introduces aliasing and domain-specific distortions. To bridge this gap, we propose a Degradation-Separated real-world Omnidirectional image Super-Resolution (DSOSR) framework that explicitly models the combined degradations from fisheye imaging and ERP projection. DSOSR is built upon two key insights: (1) projection degradations with strong priors significantly affect the distribution of random degradations, and (2) human attention in immersive scenarios typically focuses on local attractive viewpoints. Motivated by these observations, we develop a Perspective Projection Representation (PPR) to extract viewpoint features in parallel with the ERP branch, thereby isolating aliased degradations across domains. A Degradation-Specific Module (DSM) is then incorporated to separately modulate ERP-induced intrinsic geometric distortions and PPR-induced random real-world degradations. Furthermore, a Projection Fusion Attention Module (PFAM) is introduced to exploit inter-dependencies between ERP and PPR features, enabling more effective fusion of complementary representations. Extensive experiments demonstrate that the proposed DSOSR achieves state-of-the-art performance, delivering visually compelling and high-fidelity omnidirectional Real-SR results for practical applications.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 357