Abstract
Humans can quickly and reliably recognize natural scenes as members of particular categories (Potter 1975). Several features have been proposed to underlie this ability, e.g., the amplitude spectrum of spatial frequencies (Oliva and Torralba 2001); color (Oliva and Schyns 2000); integrity of the phase spectrum (Sadr and Sinha 2001); or scene structure represented by line drawings (Walther et al. 2011). How can we reconcile these different claims? Presumably, the brain uses all of these features when they are available. However, the timing of feature utilization likely depends on where in the visual processing hierarchy particular features are processed. Here we introduce a new method that allows us to measure the time order of features in their contribution to natural scene categorization. Participants were trained to categorize briefly presented, masked images of natural scenes from six different categories in a six-alternative forced-choice (6AFC) task. After staircasing stimulus onset asynchrony (SOA) to a performance level of 65% (chance: 16.7%), SOA was held constant for the remainder of the experiment. In test trials, the SOA was split in half, and during each half, a different version of the same natural scene was presented. The two versions consisted of two different feature representations, e.g., a line drawing and a color photograph. The order of the two versions was randomized and counter-balanced across categories. We expect accuracy in the 6AFC task to be higher when the versions are shown in the same order in which they are processed in the brain compared to presentation in the reverse order. In a validation of this method we found, for instance, that the color and texture information contained in color photographs were processed earlier than the structure information conveyed by line drawings (p<0.01), which is consistent with known neurophysiology of the human visual system.
Meeting abstract presented at VSS 2012