Abstract
Many theories of object recognition assume that the representation of an object specifies its axis structure (e.g., Marr, 1982). Can LO (an area critical for shape recognition) distinguish between highly similar objects, all with the same shaped parts, that differ only in the relative positions of their parts, i.e., in their axis structures? We tested the issue using fMRI multi-voxel pattern analysis. Our stimuli consisted of nine images, generated from three views (rotations in depth and in the plane) of each of three different novel objects, all composed of the same three geons, but differing in the arrangement of those parts. Unlike several prior studies, which used diverse sets of colored photos of familiar objects that differed greatly in many attributes, the images were all highly similar line drawings with no shading or familiar interpretation, and thus represent a theoretically clear test of shape selectivity per se. While viewing single presentations of the nine images, subjects identified each object by button press (1, 2, or 3), ignoring the object's orientation. A support vector machine classifier was trained and tested on independent splits of the data in different regions of interest. In V1, the classifier performed more accurately at separating groups of images of similar global orientation, and more poorly at separating groups of images based on the identity of the objects. In LO, this effect was reversed: greater accuracy was achieved separating objects (that is, different axis structures) than different global orientations. We interpret this double dissociation between V1 and LO as a fundamental shift in the shape similarity space, and conclude that LO is more sensitive to the relative positions of an object's component parts–i.e., its axis structure–than to the global orientation of the object.
NSF BCS 04-20794, 05-31177, 06-17699 to IB.