Evidence for perceptual convergence of object- and layout-based scene representations

Abstract

Behavioral data suggest that human scene recognition draws heavily upon analysis of global scene properties, such as three-dimensional layout (Greene and Oliva, 2009). At the same time, the identities of objects within scenes tend to correlate strongly with scene identity, making them potentially useful information sources during scene recognition. Consistent with this, recent work has shown that scene recognition is hampered when objects are removed, and that object-evoked response patterns in object-selective occipitotemporal cortical areas can predict patterns evoked by scenes associated with those objects (MacEvoy and Epstein, 2011). These results suggest the existence of an alternate, object-based route to scene recognition. In the present study, we asked whether these layout- and object-based routes converge at the perceptual level by examining the impact of objects upon perceived spatial layout of scenes. We used a web-based survey to obtain subjective spaciousness ratings of 500 exemplars of bathrooms, a scene category characterized by highly diagnostic objects and a wide range of spaciousness among exemplars. Similar to the design of Greene and Oliva (2010), subjects in the main experiment were adapted to scene examplars drawn from either top or bottom quintile of spaciousness ratings and then asked to provide ratings of the spaciousness of exemplars drawn from the middle three quintiles. Adapting scenes were shown either with diagnostic objects masked or visible. Spaciousness ratings of middle-range exemplars were significantly elevated or lowered after adaptation to bottom- or top-quintile exemplars, respectively, relative to unadapted control ratings. However, the magnitudes of adaptation effects in both directions were significantly smaller when subjects were adapted to scenes with visible diagnostic objects. The presence of diagnostic objects appears to have tempered excursions in the perceived spaciousness of adapting scenes, indicating that layout- and object-based scene processing channels converge at a perceptual level. Meeting abstract presented at VSS 2012