Abstract
The ability to process the information quickly and reliably in a visual scene is fundamental for survival. It is widely recognized that the visual system must operate an early and drastic input data reduction, to efficiently process relevant visual information within limited available resources. In past work, we used the principle of constrained maximum entropy to define a small number of local features that are optimal carriers of information. Experimentally, in fast vision conditions, we found that images where only optimal features are kept (sketches), and all the others are dropped, are still recognizable as the original versions. Here we explore whether these optimal local features still play an important role for gaze orientation and scene discrimination in more natural settings, where all the existing features are kept, but the overall information is reduced by showing only a fraction of the image. We briefly presented (25ms) a few small fragments, randomly extracted from natural images. Participants were invited to look at one fragment to be able to perform later a discrimination task (2IFC) between the full version of the image shown and a distractor. To reduce the contribution of global information on discrimination, in a second session, the contrast of the images could be randomly inverted. We found that eye movements head towards fragments containing more optimal features, especially in the session where image contrast could be inverted, and thus subjects had to presumably rely on local rather than global information. In this session, image discrimination performance was also significantly higher after saccades toward optimal features. Our results suggest that, in fast vision, optimal features, predicted by the constrained maximum-entropy model, are the preferential target of eye movements because of their saliency. These local features are also used for image discrimination when global information is not fully available.