 |
 |
|
 |
||
                  |
Natural Scene Recognition from Global Properties: Seeing the Forest from "Camouflage+Enclosed" instead of the Trees Michelle Greene Computational Visual Cognition Lab, Brain & Cognitive Science, MIT Traditionally, behavioral and modeling research in scene recognition has concentrated on the objects that a scene contains. According to this approach, high-level scene recognition might proceed as follows: first by segmentation and parsing the image input into candidate objects. Next, one must recognizing those objects, and then understand the relationships between those objects. Finally, with all of this information, one might be able to recognize the scene before you. More recent work has raised trouble for this account by demonstrating conditions under which observers can detect a scene category without being able to recognize its composite objects. What if scene recognition doesn't require those laborious first stages? Here we provide the first evidence that rapid basic-level scene categorization can be mediated by global scene properties that describe the spatial layout and function of the natural environment (such as perspective, mean depth, degree of navigability, and camouflage), and are not anchored in any segmentation and object identification stages. We provide behavioral evidence that these properties are used by human observers during rapid scene categorization as well as formal evidence, by means of a model observer, that these properties can be sufficient for basic-level scene categorization. |