Abstract
Scenes of the outside world share a common characteristic: they have specific statistical properties distinguishing them from purely random images. In the frame of natural images, scenes exhibit a spatial structure and dominant orientation regularities evocative of their category membership (Guérin-Dugué, Bernard and Oliva, 1998, Perception, 27, 151 ; Oliva and Torralba, 2001, International Journal of Computer Vision, 42,145–175). These global properties lead to defining the concept of scene spatial signature.
The aim of the current study was to measure the influence of this specific global information of the scene on the localization of local targets. Three experiments investigated the relative weight of both global and local levels of information. A jumbling paradigm, borrowed from Biederman's study (1972, Science, 177, 77–80), allowed to control the local spatial organization around the target as well as the relative spatial arrangement of elements in the whole scene. Test images were filtered or unfiltered. Low-pass filtering of the images impaired the physical description of the target.
Results showed that, initially, when images were blurred, the global structure played a major role. However, after a learning period, subjects started using the invariant distinctive features of the target to localize it whatever the aspect of the scene.
In analogy to the scene spatial signature, these local information distinguishing the target from its close environment may define the object spatial signature. This perceptual learning process frees the subjects from having to rely on global information. The present study shows that perceptual strategies are not determined once and for all. The visual system is flexible and juggles with the various information available at both global and local levels to adapt to visual constraints.