Abstract
One class of object recognition models posit that the brain extracts the global information of an object (e.g., shape) to generate multiple hypotheses, or "candidates," about the identity of an object. Additionally, context constrains and refines the initial candidates. If this is the case, is it then possible to modulate the interpretation of an object with contexts related to the different candidates? In a series of experiments, we presented participants with low pass filtered images of isolated objects in the periphery for 70 ms (followed by a mask). Participants were asked to identify the object in an open-ended response. We found that without context, many objects were often misinterpreted as another object sharing similar global shapes. These different interpretations were normed and categorized by their frequency of response. We then repeated the experiment with a new set of participants. This time, the images were preceded by either a fixation point or a scene. The object was congruent to the scene when the correct interpretation (e.g., football) matched the scene (e.g., football field), "falsely congruent" when the scene (cosmetic shop) matched the most frequent misinterpretation (lips), or incongruous when neither interpretation were related to the scene (farm). Interestingly, the data suggests that once a stimuli was interpreted as a certain object, the interpretation could not be refuted with an inappropriate context. However, the interpretation could be reaffirmed with an interpretation-congruent context as evident by a 20% increase in accuracy for the truly congruent condition and a 16% increase in false congruent condition (compared to presentation with no scenes). The data follows a pattern similar to that of a confirmation bias. Additionally, the current data suggests a more minimal role of context in the early stages of object recognition in which context does not inhibit competition.
Meeting abstract presented at VSS 2014