Centre for Autonomous Systems,
KTH Royal Institute of Technology,
A human observer is constantly active, moving about in the world and shifting gaze while fixating different parts of the surrounding scene. Sometimes the movements are performed with an intent to acquire more information about some object or part of the scene that is of interest, but often there’s no direct relation between the motions and actions by the observer and what he/she is looking at. Nevertheless the motions and fixations performed always provide valuable information about the 3D structure of the world and not least about what possibly could constitute objects of interest.
These observations suggest that artificial systems – ‘seeing robots’ – should use similar approaches to perceive and act in their environment. In the talk I’ll discuss the role of monocular and binocular fixation in 3D perception by a robot vision system and how the fixation and body actions influence the ’understanding’ of the world. It will be shown that problems of e.g. figure-ground-segmentation that seem difficult in computer vision become easier when observer actions are taken into account. I’ll also show some examples of how we’ve implemented mechanisms of such types.