TriGaze360: A Three-Branch Viewport-Aware Network for Stereoscopic Omnidirectional Image Quality Assessment

Stereoscopic omnidirectional images (SOIs) play a key role in immersive VR, but are highly vulnerable to distortions introduced during capture, compression, and rendering. Existing SOIQA methods often fail to fully account for binocular inconsistencies, content-guided viewport selection, and temporally coherent viewing behavior. To address these issues, we present TriGaze360, a three-branch quality assessment framework that jointly models monocular degradations, binocular disparities, and sequential viewport interactions. The framework integrates a content-guided dynamic viewport sampling strategy, a temporal scoring module for inter-viewport dependency modeling, and multi-scale fusion with perceptual attention for enhanced distortion sensitivity. Extensive experiments on public SOIQA datasets show that TriGaze360 achieves state-of-the-art performance, providing a perceptually aligned and temporally aware solution for stereoscopic immersive content.