Visual attention and target detection in cluttered natural scenes

Abstract

Rather than attempting to fully interpret visual scenes in a
\nparallel fashion, biological systems appear to employ a serial strategy by
\nwhich an attentional spotlight rapidly selects circumscribed regions in the
\nscene for further analysis. The spatiotemporal deployment of attention
\nhas been shown to be controlled by both bottom-up (image-based) and
\ntop-down (volitional) cues. We describe a detailed neuromimetic computer
\nimplementation of a bottom-up scheme for the control of visual
\nattention, focusing on the problem of combining information across modalities
\n(orientation, intensity, and color information) in a purely stimulusdriven
\nmanner. We have applied this model to a wide range of target
\ndetection tasks, using synthetic and natural stimuli. Performance has,
\nhowever, remained difficult to objectively evaluate on natural scenes,
\nbecause no objective reference was available for comparison. We
\npresent predicted search times for our model on the Search–2 database
\nof rural scenes containing a military vehicle. Overall, we found a poor
\ncorrelation between human and model search times. Further analysis,
\nhowever, revealed that in 75% of the images, the model appeared to
\ndetect the target faster than humans (for comparison, we calibrated the
\nmodel’s arbitrary internal time frame such that 2 to 4 image locations
\nwere visited per second). It seems that this model, which had originally
\nbeen designed not to find small, hidden military vehicles, but rather to
\nfind the few most obviously conspicuous objects in an image, performed
\nas an efficient target detector on the Search–2 dataset. Further developments
\nof the model are finally explored, in particular through a more
\nformal treatment of the difficult problem of extracting suitable low-level
\nfeatures to be fed into the saliency map.