One of the major research topic in computer vision has been the extraction of the geometric layout from a set of images. We intend to revisit this problem and investigate it within the context of visual attention.

In particular we will concentrate on the problem of depth computation using stereo. Currently this visual modality is viewed as a bottom-up process that builds a 3-D depth map from two images. Nevertheless, the field of view of a camera pair with a fixed geometry is inherently limited. Therefore, such a static sensor cannot build a complete and reliable representation of the world. At the other extreme, if the attention of the observer must focus onto a specific region, the narrowness of the field becomes a desirable feature. Active vision plays a crucial role and there should be a strong link between high-level factors (such as attention, memory, or expectation) and sensory-motor control. We will integrate the task of depth estimation within an attention-sensory-motor loop. Within such a paradigm depth must be estimated extremely fast. The stereo sensor itself will be an active camera head with modern hardware components. This research objective is representative of a fine coupling between science and technology.