Abstract
Human observers can identify the basic-level category of complex visual scenes with great speed and accuracy (Potter, 1975). Although one route to scene categorization might involve analysis of global scene properties such as spatial layout or image statistics (Davenport & Potter, 2004; Oliva & Torralba, 2001), a second route might involve analysis of within-scene objects (Quattoni & Torralba, 2009). To test the plausibility of an object-based recognition mechanism, 16 participants performed a four-alternative forced-choice task in which they identified color photographs of bathrooms, intersections, kitchens, and playgrounds, each of which contained two strongly diagnostic signature objects (e.g., stove and refrigerator for kitchen, traffic light and car for intersection). Each stimulus was presented for 50 ms followed by a pattern mask. Critically, the photographs could either be presented in their original form or with one or two signature objects obscured by a phase-scrambled mask (leaving the majority of the image intact). Relative to performance on unaltered photographs, obscuring a single signature object increased participants' reaction times by 37 ms and reduced accuracy on the scene classification task by 6%, while obscuring both signature objects increased reaction times by 124 ms and reduced accuracy by 19%. This effect is not simply a consequence of image degradation: even after controlling for the proportion of the scene that was masked, accuracy was significantly worse for scenes with both signature objects occluded than with just one. Although indoor and outdoor scenes were classified with comparable accuracy when one of the signature objects was removed, accuracy was significantly poorer for indoor scenes than for outdoor scenes when both signature objects were removed. Together, these results suggest that scene identification is facilitated by recognition of the objects within them. This process may be more important for the identification of indoor scenes, which often possess similar spatial layouts.
National Institute of Mental Health.