Learning visual attention

Active vision is the skill that allows to control vision sensors and to filter visual signals in order to response to particular tasks and external stimuli in a fast and effective way. In biological and artificial systems it enables faster processing and reacting, providing meaningful information for subsequent activities such as navigation, manipulation, movements planning, interaction with the environment and with human partners, to mention only a few. The cognitive process at the very core of active vision is attention, i.e. the ability of selective information acquisition of salient spots and objects within cluttered and heterogeneous scenes. This is a crucial capability since it optimizes the deployment of limited cognitive and memory resources. As such processing constraints arise in robotic platforms as well, particularly in real time tasks, visual attention and oculomotor planning are nowadays among the most challenging issues in computer vision. Endowing an artificial system with attentive abilities means to simplify several cognitive and vision tasks, to speed up object detection and recognition so as routines in task learning frameworks. Moreover is a crucial feature in the field of cognitive robotics as it fosters movement, action and intention recognition, grounding reasoning, inference and internal representation of the world on perception of significant social cues and interaction with humans (it is generally accepted that fails in the development of visual attention are responsible for disorders of social development such as autism, [1]). Since the topic is directly inspired by neurophysiological mechanisms and cognitive behaviours it has benefited and in some cases repaid the contribution of Cognitive Sciences, providing interdisciplinary insights and useful collaborations and addressing important issues aimed at making robots possibly flexible and autonomous in a human-like way. Therefore roboticists nowadays agree in considering attention a fundamental component of every cognitive architecture.