VDRS-VQA: Viewport-Domain Residual Sampling for No-Reference 360$^\circ$ Video Quality Assessment

State-of-the-art no-reference video quality assessment (NR-VQA) methods for conventional 2D videos achieve remarkable performance by leveraging quality-sensitive local spatial fragments. However, directly applying these planar models to 360$^\circ$ videos suffers from severe performance degradation, as 360$^\circ$ content is typically represented in the equirectangular projection (ERP) format, where latitude-dependent geometric stretching violates the local spatial assumptions underlying fragment-based sampling. To address this issue, we propose VDRS-VQA, which adapts ReLaX-VQA, a leading residual-driven 2D NR-VQA framework, to 360$^\circ$ video quality assessment via a viewport-domain residual sampling strategy without modifying its original backbone. Specifically, local spatial (LS) and spatio-temporal residual (STR) branches perform sampling in a viewport aggregation domain constructed from gnomonic viewports, while the global semantic (GS) branch retains ERP input to preserve holistic scene semantics. Extensive experiments on ODV-VQA and BIT360 show significant gains over the baseline, confirming that geometry-consistent sampling is critical for adapting planar NR-VQA models to 360$^\circ$ videos.