Schema-Driven Influences in Recovering 3-D Shape from Motion in Human and Computer Vision

Abstract

To investigate the role of prior knowledge in perceiving the hollow-face illusion, we set out to correlate human performance with that of a computational model that tracks facial features. Toward this goal, we prepared a video of a thin human mask, realistically painted on both the convex and concave sides, that rotated by 360 degrees. It is well known that when humans view the rotating hollow part of the mask, they perceive it as a convex face that rotates in the opposite direction. We submitted this video as input to DeCarlo & Metaxas’s (Int. J. Comput. Vis. 38(2):99–127, 2000) face tracking algorithm that uses a combination of optical flow and feature alignment to track moving faces; importantly, the algorithm uses a schema of a 3-D convex face. The algorithm was fooled, as humans are, into misperceiving a concave face as a convex face rotating in the opposite direction. Notably, this is a significant case where a computer vision algorithm “experiences” an illusion. However, when the convex-face schema was replaced with a schema that allowed both convex and concave faces, the algorithm tracked the mask accurately. The similarity in the responses of the computer vision algorithm and human observers reinforces the hypothesis that a built-in convex 3D face schema imposes a depth inversion for concave faces that in turn forces the false interpretation of motion signals.