Abstract
Identifying the relationships between objects in a scene is a fundamental goal in scene understanding. It is known that two interacting objects are perceptually grouped (Green & Hummel, 2006), and that interacting objects evoke greater activity in the lateral occipital complex (LOC) compared to noninteracting objects (Kim & Biederman, 2011). However, critical questions remain about the neural representations of interacting objects. Do object-sensitive regions such as LOC actually encode interactions between objects? How do scene-sensitive regions such as the parahippocampal place area (PPA) respond to simple two-object interactions?
In our fMRI experiment, subjects performed a one-back task while viewing three types of images: humans and objects in isolation, a human and object overlapping without a meaningful interaction, and a human interacting with an object in a familiar way. We then used MVPA methods to determine which regions are sensitive to the meaningful interactions.
We have obtained preliminary results from an experiment investigating two types of human-object interactions: riding a horse and playing a guitar. For each subject, we trained a classifier to discriminate between the responses to the object pairs person+horse and person+guitar, and then tested our classifier on a held-out run. We find qualitatively different results in object-sensitive areas and scene-sensitive areas. In LOC, both interacting and noninteracting object pairs can be decoded with similar accuracies. In PPA, however, decoding was better for interacting that noninteracting objects, despite the fact that both conditions contain the same human/object pairs. These results suggest that PPA’s sensitivity to scenes extends to a simple interaction of a single human and object, whereas LOC can represent object pairs regardless of whether the objects are interacting in a meaningful way or not.
Meeting abstract presented at VSS 2012