The paper introduces a new bottom-up visual saliency model called Graph-Based Visual Saliency (GBVS), which consists of two main steps: forming activation maps on feature channels and normalizing them to highlight conspicuity. The model is designed to be simple and biologically plausible, leveraging the parallel nature of graph algorithms. GBVS is evaluated on a dataset of 749 variations of 108 natural images, achieving 98% of the ROC area of a human-based control, outperforming classical algorithms like Itti & Koch's by 14%. The model's performance is attributed to its ability to robustly highlight salient regions, even when they are distant from object borders, and its center bias, which aligns with human visual experience. The paper also discusses the potential extension of GBVS to multiresolution representations to further improve performance.The paper introduces a new bottom-up visual saliency model called Graph-Based Visual Saliency (GBVS), which consists of two main steps: forming activation maps on feature channels and normalizing them to highlight conspicuity. The model is designed to be simple and biologically plausible, leveraging the parallel nature of graph algorithms. GBVS is evaluated on a dataset of 749 variations of 108 natural images, achieving 98% of the ROC area of a human-based control, outperforming classical algorithms like Itti & Koch's by 14%. The model's performance is attributed to its ability to robustly highlight salient regions, even when they are distant from object borders, and its center bias, which aligns with human visual experience. The paper also discusses the potential extension of GBVS to multiresolution representations to further improve performance.