SpherIQ: A Spherical Image Quality Transformer—Geometry-Aware MUSIQ for Omnidirectional IQA
Year: 2026
Type: article
Abstract: The immersive quality of Virtual Reality (VR) depends critically on high-fidelity omnidirectional and stereoscopic visual content. Traditional Image Quality Assessment (IQA) methods are ill-equipped to handle the non-Euclidean geometry of VR media, struggling with the severe spatial distortions introduced by equirectangular projections. In this paper, we present SpherIQ, a geometry-aware no-reference IQA architecture for monoscopic and stereoscopic 360-degree images. Our method projects the spherical visual field into a cubemap representation to mitigate polar distortions, then restores global spatial context through a geometric embedding scheme: 3D unit-sphere coordinates encoded via Rotary Position Embedding (RoPE) and learnable topological face embeddings, injected directly into the Transformer’s self-attention mechanism. This allows the model to reason about the 3D structure of the visual sphere and learn quality-relevant dependencies across discontinuous cube faces. To address the scarcity of VR IQA training data, we further introduce a fused spherical rotation augmentation that composes the rotation with the cubemap projection in a single bilinear sampling step, eliminating the double-interpolation blur of naive two-step approaches. Evaluated on the LIVE 3D IQA dataset under a rigorous 5-fold scene-disjoint cross-validation protocol, SpherIQ achieves a mean SRCC of $0.860 \pm 0.032$ and mean PCC of $0.830 \pm 0.049$ , with cross-dataset transfer demonstrating strong generalization to CVIQ and IQA-ODI. Ablation experiments further confirm that both 3D spherical RoPE and topological face embeddings contribute complementarily to predictive performance.
Source: IEEE Access
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7203445784