Object perception through visual attention

Dynamic and robust visual perception is made possible through active vision, which offers an alternative to the conventional feedforward architectures in computer vision, these are typically static and passive, relying heavily on large datasets and substantial computational resources. When learning to recognise new objects, infants actively explore their environment through coordinated sensorimotor interactions, guided by feedback that shapes perception. The integration of Sensorimotor Contingencies Theory (SMCT) with active vision is effectively demonstrated in the work of Maye et al. (2011), where sensorimotor sequences are used to learn object representations within a visual scene. However, most vision datasets either (i) rely on static, 2D frame-based input that fails to capture the embodied, sensorimotor nature of real-world object interaction, while also demanding high computational resources and large amounts of data, or (ii) consist of 3D datasets that, although they represent more realistic geometries, typically lack mechanisms for identifying attentional regions of interest (ROIs) or simulating fixational eye movements. Furthermore, existing pipelines seldom integrate all components essential for autonomous robotic perception, including continuous gaze shifts, event-based sensing, and saliency-driven object exploration. To address these gaps, we develop a bioinspired, event-based attentional architecture, building upon the work of D'Angelo et al. (2025), that emulates visual exploration by shifting the focus of attention through object motion sensitivity. This mechanism mimics overt selective attention and enables efficient filtering of visual input, allowing the system to focus computational resources on the most informative regions of the scene. Expanding on this foundation, we incorporate sensorimotor sequences for object learning and propose a four-stage process: (i) simulation of micro-saccadic fixations on synthetic 3D-modeled objects; (ii) conversion of rendered video into event streams; (iii) computation of dynamic saliency maps to extract attentional regions of interest (ROIs); and (iv) generation of spatial saccadic trajectories.