Two selection stages provide efficient object-based attentional control for dynamic vision

In this paper, we introduce semiattentive computations as the result of replacing the usual single selection stage of visual attention by two consecutive selection stages. They are motivated by shortcomings of conventional attention models and correlate well to findings in human attention. The first selection stage employs preattentive saliency computations for the complete available input, and selects a small number of discrete items. These are subject to the semiattentive processes of tracking and information accumulation. The second stage selects a single element from the result of the first selection stage for the conventional focus of attention. The implementation and efficiency of this scheme is demonstrated in this paper. Its main advantage is the efficient selection and inhibition of objects in dynamic scenes. It allows the serialized accumulation of information for a changing environment and provides an upto-date world model. The focus of this paper is on the quality of the computed world model and the object-related computations.