A context-based model of attention

Artificial visual systems need an attentional selection mechanism to constrain costly processing to relevant parts. An important design decision for such systems concerns the locus of selection. To guide the selection mechanism, traditional models of attention use either an early locus of selection based on low-level features (e.g., conspicuous edges) or a late locus of selection based on high-level features (e.g., object templates). An early locus is computationally cheap and fast but is an unreliable indicator of the objecthood. A late locus is computationally expensive and slow and requires the object to be selected to be known, rendering selection for identification useless. To combine the advantages of both loci, we propose the COBA (COntext BAsed) model of attention, that guides selection on the basis of the learned spatial context of objects. The feasibility of context-based attention is assessed by experiments in which the COBA model is applied to the detection of faces in natural images. The results of the experiments show that the COBA model is highly successful in reducing the number of false detections. From the results, we may conclude that context-based attentional selection is a feasible and efficient selection mechanism for artificial visual systems.