Neurocomputational bases of object and face recognition

Abstract

A number of behavioural phenomena distinguish the recognition of faces and objects, even when members of a set of objects are highly similar. Because faces have the same parts in approximately the same relations, individuation of faces typically requires speci¢cation of the metric variation in a holistic and integral representation of the facial surface. The direct mapping of a hypercolumn-like pattern of activation onto a representation layer that preserves relative spatial ¢lter values in a two-dimensional (2D) coordinate space, as proposed by C. von der Malsburg and his associates, may account for many of the phenomena associated with face recognition. An additional re¢nement, in which each column of ¢lters (termed a`jet') is centred on a particular facial feature (or ¢ducial point), allows selectivity of the input into the holistic representation to avoid incorporation of occluding or nearby surfaces. The initial hypercolumn representation also characterizes the ¢rst stage of object perception, but the image variation for objects at a given location in a 2D coordinate space may be too great to yield su¤cient predictability directly from the output of spatial kernels. Consequently, objects can be represented by a structural description specifying qualitative (typically, nonaccidental) characterizations of an object's parts, the attributes of the parts, and the relations among the parts, largely based on orientation and depth discontinuities (as shown by Hummel & Biederman). A series of experiments on the name priming or physical matching of complementary images (in the Fourier domain) of objects and faces documents that whereas face recognition is strongly dependent on the original spatial ¢lter values, evidence from object recognition indicates strong invariance to these values, even when distinguishing among objects that are as similar as faces.

Face Recognition Is Affected by Similarity in Spatial Frequency Range to a Greater Degree Than Within-Category Object Recognition.

“…The visual system could exploit this characteristic of sharp edges in matching images with different spatial frequency content by using nonlinear analyses of spatial frequency information. Our finding that objects are less sensitive to SFO variations than faces is compatible with Biederman and Kalocsai's (1997) results showing that spatially complementary images of objects are more easily recognized than similarly filtered face images. However, our study differed from theirs in a number of important ways.…”

“…Although Biederman and Kalocsai's (1997) results are intriguing, there are some limitations to their study. For instance, they examined spatial complementarity in terms of orientation and frequency at the same time, making it impossible to know what the effect of either factor alone might have been.…”

“…This is somewhat similar to a condition of zero SFO, except that Biederman and Kalocsai's manipulation affected both spatial frequency and orientation, whereas SFO manipulations affect spatial frequency in isolation. Biederman and Kalocsai's (1997) results showed that participants found it quite easy to match or recognize these spatially complementary images if they represented chairs but that they had difficulty doing the same task with face images. As an explanation of these findings, they proposed that the two types of images are stored in different ways.…”

Organization of face and object recognition in modular neural network models

“…More recent versions (Wiskott et al, 1997) place the jets over particular facial features (fiducial points) for greater accuracy. Biederman and Kalocsai (1997) show how Wiskott et al's representation can account for psychological phenomena in face recognition, and the system was recently the top performer in the U.S. Army's FERET Phase III face recognition competition (Okada, et al, 1998). Thus the Gabor wavelet jet is a good representation for face recognition.…”

“…Farah, Wilson, Drain and Tanaka (1998) review the literature on this topic and provide new evidence for holistic perceptual representations. Biederman and Kalocsai (1997) propose a computational basis for such representations. They show that the outputs of an array of overlapping local spatial filters similar to some of the receptive fields in visual cortex, as used in Wiskott, Fellous, Kruger & von der Malsburg's (1997) face recognition system, can account for human performance in experiments using face stimuli but cannot account for human performance in other object recognition tasks.…”