Modeling visual attention on scenes
Brice Follet, Olivier Le Meur, Thierry Baccino · HAL (Le Centre pour la Communication Scientifique Directe) · 2010
ABSTRACT. Research in the computational modelling of the visual attention has mushroomed in recent years. First generation of computational models, called bottom-up models, allows to calculate a saliency map indicating the degree of interest of each area of a picture. These models are purely based on the low-level visual features. However, it is now well known that the visual perception is not a purely bottom-up process. To improve in a significant manner the quality of the prediction, top-down information (prior knowledge, expectations, contextual guidance) have to be taken into account. We propose in this article to describe some bottom-up models and the metrics used to assess their performances. New generation of models based both on low-level and high-level information is also briefly described. To go one step further in the understanding of the cognitive processes, new computational tools have been recently proposed and are listed in the final section. RÉSUMÉ. La modélisation computationelle de l’attention visuelle connaît actuellement un essor considérable. Les premières modèles, purement basés sur l’attention dite exogène, per-mettent de calculer une carte de saillance indiquant les zones d’intérêt visuel d’une image. Cependant, afin d’améliorer cette prédiction, il s’avère nécessaire de prendre en compte des