Visual coding of natural contours leads to poor discrimination of object-shape around canonical views.

Although progress has been made in understanding how we detect visual contours we know much less about how we encode their shape. Here we describe a psychophysical paradigm that does this by quantifying the perceptual similarity of complex contours: observers decided which of three outline contours (strings of Gabors derived from silhouettes of natural objects) was the "odd-man-out" (where one was derived from a subtly different 3D view of the same object). We estimated the minimum perceptible contour change (i.e. rotation-in-depth) for different starting views of a 3D hand-object. We report poorest discrimination of contours around canonical views ("characteristic" or "typical" object views); small rotations in non-canonical ("unusual") views tend to have more readily perceptible consequences leading to better performance. This finding extends to other objects, and is robust to random-scaling of contours, to randomization of local orientation structure, and even to replacement of oriented elements with non-oriented Gaussian blobs. We compared our results to predictions from a simple model of shape similarity using cross-correlation of silhouette-images. Meeting abstract presented at VSS 2015.