Abstract
Humans instantaneously and effortlessly obtain semantic information from visual scenes, yet it remains unclear which scene features drive this identification. This study examines the possibility that a subset of objects are central to scene categorization. Many models of object usage in scenes assume that all objects contribute equally to scene category. However, some objects are more diagnostic of their environments. For example, while “chairs” can be found in many types of scenes, “refrigerators” are evocative of kitchens, and are therefore diagnostic of the scene category. Using a labeled scene database, we defined diagnostic objects as those found nearly exclusively in a single scene category (p(category|object) > 0.9). We obscured all diagnostic objects in each image using localized phase scrambling (diag- condition) and compared these images with both unmodified images, and images with a similar amount of non-diagnostic information obscured (rand- condition). Observers (N=14) viewed 996 images for 250 ms each and performed 2AFC categorization, while 64-channel EEG was recorded. Observers were most accurate (0.91) for the intact images, but were also significantly more accurate for the rand- (0.90) compared with the diag- (0.86, p< 0.001, d=0.87). Observers were also 69 ms slower to classify the diag- images compared with the rand- (p< 0.0001, d=0.37). EEG waveforms were submitted to a linear classifier trained on unmodified images in order to determine the time-resolved decoding accuracy of diag- and rand- conditions. We found decodable information starting around 60 ms post-image onset for both experimental conditions. Decoding accuracy was significantly lower for diag- images starting 143 ms after image onset, suggesting that while the diag- images did not disrupt initial visual processing, diagnostic objects contributed to later semantic processing.
Acknowledgement: National Science Foundation (1736274) grant to Michelle R. Greene