Abstract
It is generally believed that surface-based representations bridge the perceptual gap between low-level image features and high-level scene categories. Here we propose a new method to investigate these representations.
First we show that the perception of complex 3D surface configurations can occur in a purely top-down manner, that is in the absence of structured local signals. Observers were asked to indicate the presence or absence of a such a configuration (a large ‘+’ in relief) in random-dot stereograms which, unbeknownst to them, contained only disparity white noise. Both halves of the random-dot stereograms subtended 2.470 × 2.470 deg of visual angle and were composed of a white background filled with 700 black texture elements each spanning 0.048 × 0.048 deg of visual angle (3 × 3 pixels). Each texture element was randomly positioned in the left eye and was shifted in the right eye equiprobably by either −0.963 (one pixel to the left) or 0.963 (one pixel to the right) arcmin. The number of texture elements (average black to white pixel ratio = 0.232) and their perceived depth (approximately 9 mm in front or behind the screen) were chosen such that a noisy cloud of dots was the dominant percept. Even though no signal was ever presented, observers detected, and reported that they could see, the target in an important proportion of the 20 000 stereograms to which they were exposed.
Second, using least-square multiple regression, we revealed the internal 3D surface representations, or more precisely the Wiener kernels, that best accounted for the observers' responses in the least square sense. The templates were shown to be spatially well defined and to be very stable across trials. Multiple regression has therefore allowed us to reveal pure (uncorrupted by low-level signal) 3D internal surface representations.
Supported by: NSERC PDF — 242082 — 2001 to BB; HFSP RG0109/1999-B to PM.