Panoramic Video Saliency Prediction with Spherical-Aware Two-Stream Modeling

We present Spher-ATSal, a sphere-native twostream framework for 360° video saliency under equirectangular projection (ERP). The spatial stream operates directly in ERP with geometry-aware attention, mitigating polar distortion and seam artifacts without cube-map conversion. The temporal stream employs a spherical ConvLSTM to capture long-range dynamics and produces a single-channel temporal map at the same resolution. The two streams are fused by parameterfree per-pixel multiplication, preserving ATSal’s simplicity and avoiding cross-representation alignment. Training adopts sphereconsistent objectives. On the VR-EyeTracking benchmark, SpherATSal consistently outperforms the original ATSal in CC, NSS, and AUC-J with only modest inference overhead, making it a plug-and-play, engineering-friendly solution for VR/AR rendering and viewport prediction.