Humanoid Early Visual Processing Using Attention Mechanisms

The basic, biologically inspired idea of the authors (and some more in the scientific community) is to apply an attention-based type of filter on the great amount of visual input data and only perform further analysis on the tiny rest. This residual is what we call the regions of interest (ROIs). There have been many approaches to computation of salient features in a static image, e.g. [1] shows that high contrast regions seem to attract attention or [2] reports that salient regions can be computed using multiscale images. Others on the other hand argue that local complexity can be a measure of saliency [3]. Also, a learning approach for visual saliency models has been proposed recently [4]. Following these ideas, one foundation of our approach is the claim, that fundamental attention attractors originating from sensory input can be either static salient features in a single frame or dynamics in the input data sequence (considering temporal properties). Inspired by the idea in [5], which is claimed to be biologically plausible, we extend the saliency attention approach with the idea, that vision is a process of active and sometimes even volitional exploration of the environment. Thus, considering the second before mentioned assumption, i.e. based on the theory of inhibition of return as shown to be plausible in human visual psychophysics [6], we implement top-down cognitive feedback in the proposed system. Moreover, we will go one step further and integrate a possiblity not only for attention inhibition, but also for directed attention guidance. This reinforcement is triggered by cognitive processes reasoning about relevant additional information to gain from a specific region (see Section III). Although we do not claim to implement the entire framework of [5] e.g. the inattentional or change blindness, we in deed show that a system utilizing the basic ideas performs considerably better than without. Not contradicting, but complementing the work of other authors [7], [8], we do not want to focus solely on building a biologically plausible visual systems, but our primary target is to apply the underlying ideas of such frameworks to a real-world robotic setup. We therefore avoid complex neural, connectionist or machine learning techniques where possible, giving preference to discrete algorithms. These fast and efficient algorithms allow for realtime performance and high accuracy for manipulation tasks on standard hardware. The vision system presented in this paper is part of the JAST1 human-robot dialog system.