Performance Bounds for Computational Models of Visual Saliency in 360 Videos

Estimating performance bounds can improve the comparative analysis of saliency models by exposing their strengths and weaknesses. This paper evaluates two approaches for estimating bounds on immersive videos: Equator Bias and Saliency Sum. For validation, we compare them with outputs from two attention models - Spherical U-Net and DAVE - using three metrics: AUC-Judd, NSS, and CC. Experiments were conducted on 6 videos from the PAVS10K dataset, which includes eye-tracking data from ~20 observers. Saliency Sum achieved the best scores across all metrics, while Equator Bias scored the lowest, indicating that both approaches have significant potential for representing upper and lower performance bounds, respectively.