Prototype-referenced shape encoding revealed by high-level aftereffects.

We used high-level configural aftereffects induced by adaptation to realistic faces to investigate visual representations underlying complex pattern perception. We found that exposure to an individual face for a few seconds generated a significant and precise bias in the subsequent perception of face identity. In the context of a computationally derived 'face space,' adaptation specifically shifted perception along a trajectory passing through the adapting and average faces, selectively facilitating recognition of a test face lying on this trajectory and impairing recognition of other faces. The results suggest that the encoding of faces and other complex patterns draws upon contrastive neural mechanisms that reference the central tendency of the stimulus category.

Recognition and sex categorization of adults' and children's faces: examining performance in the absence of sex-stereotyped cues.

The ability of children and adults to classify the sex of children's and adults' faces using only the biologically based internal facial structure was investigated. Face images of 7- to 10-year-old children and of adults in their 20s were edited digitally to eliminate hairstyle and clothing cues to sex. Seven-year-olds, nine-year-olds, and adults classified a subset of these faces by sex and were asked, subsequently, to recognize the faces from among the entire set of faces. This recognition task was designed to assess the relationship between categorization and recognition accuracy. Participants categorized the adult faces by sex at levels of accuracy varying from just above chance (7-year-olds) to nearly perfect (adults). All participant groups performed less accurately for children's faces than for adults' faces. The 7-year-olds were unable to classify the children's faces by sex at levels above chance. Finally, the faces of children and adults were equally recognizable--a finding that has theoretical implications for understanding the relationship between categorizing and identifying faces.

The face typicality-recognizability relationship: encoding or retrieval locus?

Using a crossover recognition memory testing paradigm, we tested whether the effects on face recognition of the memorability component of face typicality (Vokey & Read, 1992, 1995) are due primarily to the encoding process occurring during study or to the retrieval process occurring at test. At study, faces were either veridical in form or at moderate (Experiment 1) or extreme (Experiment 2) levels of caricature. The variable of degree of facial caricature at study was crossed with the degree of caricature at test. The primary contribution of increased memorability to increased hit rate was through increased distinctiveness at study. Increased distinctiveness at test contributed to substantial reductions in the false alarm rate, too. Signal detection analyses confirmed that the mirror effects obtained were primarily stimulus/memory-based, rather than decision-based. Contrary to the conclusion of Vokey and Read (1992), we found that increments in face memorability produced increments in face recognition that were due at least as much to enhanced encoding of studied faces as they were to increased rejection of distractor faces.

On the other side of the mean: the perception of dissimilarity in human faces.

We created a 'face space' using a laser-scan representation of faces. In this space, a caricature can be made by moving a face away from the average face, along the line connecting the particular face to the average face. Here, we move the face along this line in the other direction, proceeding through the mean and 'out the other side'. This results in a face that is 'opposite', in a computational sense, to the original face. We morphed several faces into their anti-faces and sampled the morph trajectory in five discrete steps. We then collected similarity ratings from human participants for all possible pairs of morphed faces to determine how the distances in the 'physical face space' related to the distances in the 'psychological face space'. The data indicate that there is a perceptual discontinuity of face identity as the face crosses over to the 'other side of the mean'. We consider these results in the context of face-space models of human face processing.

Three-dimensional shape and two-dimensional surface reflectance contributions to face recognition: an application of three-dimensional morphing.

We measured the three-dimensional shape and two-dimensional surface reflectance contributions to human recognition of faces across viewpoint. We first divided laser scans of human heads into their two- and three-dimensional components. Next, we created shape-normalized faces by morphing the two-dimensional surface reflectance maps of each face onto the average three-dimensional head shape and reflectance-normalized faces by morphing the average two-dimensional surface reflectance map onto each three-dimensional head shape. Observers learned frontal images of the original, shape-normalized, or reflectance-normalized faces, and were asked to recognize the faces from viewpoint changes of 0, 30 and 60 degrees. Both the three-dimensional shape and two-dimensional surface reflectance information contributed substantially to human recognition performance, thus constraining theories of face representation to include both types of information.

Stimulus-specific effects in face recognition over changes in viewpoint.

Individual faces vary considerably in both the quality and quantity of the information they contain for recognition and for viewpoint generalization. In the present study, we assessed the typicality, recognizability, and viewpoint generalizability of individual faces using data from both human observers and from a computational model of face recognition across viewpoint change. The two-stage computational model incorporated a viewpoint alignment operation and a recognition-by-interpolation operation. An interesting aspect of this particular model is that the effects of typicality it predicts at the alignment and recognition stages dissociate, such that face typicality is beneficial for the success of the alignment process, but is adverse for the success of the recognition process. We applied a factor analysis to the covariance data for the human- and model-derived face measures across the different viewpoints and found two axes that appeared consistently across all viewpoints. Projection scores for individual faces on these axes (i.e. the extent to which a face's 'performance profile' matched the pattern of human- and model-derived scores on that axis), correlated across viewpoint changes to a much higher degree than did the raw recognizability scores of the faces. These results suggest that the stimulus information captured in the model measures may underlie distinct and dissociable aspects of the recognizability of individual faces across viewpoint change.

The perception of face gender: the role of stimulus structure in recognition and classification.

The perception of face gender was examined in the context of extending "face space" models of human face representations to include the perceptual categories defined by male and female faces. We collected data on the recognizability, gender classifiability (reaction time to classify a face as male/female), attractiveness, and masculinity/femininity of individual male and female faces. Factor analyses applied separately to the data for male and female faces yielded the following results. First, for both male and female faces, the recognizability and gender classifiability of faces were independent--a result inconsistent with the hypothesis that both recognizability and gender classifiability depend on a face's "distance" from the subcategory gender prototype. Instead, caricatured aspects of gender (femininity/masculinity ratings) related to the gender classifiability of the faces. Second, facial attractiveness related inversely to face recognizability for male, but not for female, faces--a result that resolves inconsistencies in previous studies. Third, attractiveness and femininity for female faces were nearly equivalent, but attractiveness and masculinity for male faces were not equivalent. Finally, we applied principal component analysis to the pixel-coded face images with the aim of extracting measures related to the gender classifiability and recognizability of individual faces. We incorporated these model-derived measures into the factor analysis with the human rating and performance measures. This combined analysis indicated that face recognizability is related to the distinctiveness of a face with respect to its gender subcategory prototype. Additionally, the gender classifiability of faces related to at least one caricatured aspect of face gender.

A moving cast shadow diminishes the Pulfrich phenomenon.

It is demonstrated that the presence of a moving cast shadow diminishes the Pulfrich phenomenon. This complements previous work by Kersten, Knill, Mamassian, and Bülthoff [1996 Nature (London) 379 31] indicating that visible cast shadows can override monocularly based cues to the perceived trajectory of a moving object. The present finding with the Pulfrich phenomenon indicates the effectiveness of shadows for overriding binocularly based cues to the perceived trajectory of a moving object.

Facial aging, attractiveness, and distinctiveness.

A standard facial caricature algorithm has been applied to a three-dimensional (3-D) representation of human heads, those of Caucasian male and female young adults. Observers viewed unfamiliar faces at four levels of caricature--anticaricature, veridical, moderate caricature, and extreme caricature--and made ratings of attractiveness and distinctiveness (experiment 1) or learned to identify them (experiment 2). There were linear increases in perceived distinctiveness and linear decreases in perceived attractiveness as the degree of facial caricature (Euclidean distance from the average face in 3-D-grounded face space) increased. Observers learned to identify faces presented at either level of positive caricature more efficiently than they did with either uncaricatured or anticaricatured faces. Using the same faces, 3-D representation, and caricature levels, O'Toole, Vetter, Volz, and Salter (1997, Perception 26 719-732) had shown a linear increase in judgments of face age as a function of degree of caricature. Here it is concluded that older-appearing faces are less attractive, but more distinctive and memorable than younger-appearing faces, those closer to the average face.

On the preattentive accessibility of stereoscopic disparity: evidence from visual search.

We examined the generality of the claim that stereoscopic disparity is detectable in parallel across the visual field. Using a search paradigm with random-dot stereograms, we varied the relative disparity of target and distractor items. When both target and distractors had crossed disparities, both search functions (i.e., target in front of distractors and target behind distractors) were linear with positive slopes. When both target and distractors had uncrossed disparities, the pattern of results depended upon whether the target was in front of or behind the distractors-specifically, when the target was in front of the distractors, search functions were similar to those seen for "crossed" search, but when the target was behind the distractors, a nonlinear search function was found. Finally, when the target and distractors straddled the plane of fixation, a nonlinear search function was found when the target was in front of the distractors; however, when the target was behind the distractors, a linear search function with a large positive slope was found. We show that the nonlinear search functions are consistent with the effects of an intervening global surface percept. We also show that the size of the stimulus display may be a factor in some relative depth cases. Additionally, we replicate Steinman's (1987) finding that search is parallel when the distractors are located at the plane of fixation and the target disparity is crossed, eliminating monocular and spatial cues to target presence that may have been present in his original study. In a final control experiment, we showed that reaction times did not increase with set size when observers performed another kind of perceptual task on similar random-dot stereogram displays. This eliminates the possibility that some of the results obtained here can be explained by increases in the difficulty of perceiving/fusing the stimuli when the number of distractors is increased.

Three-dimensional caricatures of human heads: distinctiveness and the perception of facial age.

A standard facial-caricaturing algorithm was applied to a three-dimensional representation of human heads. This algorithm sometimes produced heads that appeared 'caricatured'. More commonly, however, exaggerating the distinctive three-dimensional information in a face seemed to produce an increase in the apparent age of the face--both at a local level, by exaggerating small facial creases into wrinkles, and at a more global level via changes that seemed to make the underlying structure of the skull more evident. Concomitantly, de-emphasis of the distinctive three-dimensional information in a face made it appear relatively younger than the veridical and caricatured faces. More formally, face-age judgments made by human observers were ordered according to the level of caricature, with anticaricatures judged younger than veridical faces, and veridical faces judged younger than caricatured faces. These results are discussed in terms of the importance of the nature of the features made more distinct by a caricaturing algorithm and the nature of human representation(s) of faces.

Sex classification is better with three-dimensional head structure than with image intensity information.

The sex of a face is perhaps its most salient features. A principal components analysis (PCA) was applied separately to the three-dimensional (3-D) structure and graylevel image (GLI) data from laser-scanned human heads. Individual components from both analyses captured information related to the sex of the face. Notably, single projection coefficients characterized complex differences between the 3-D structure of male and female heads and between male and female GLI maps. In a series of simulations, the quality of the information available in the 3-D head versus GLI data for predicting the sex of the face has been compared. The results indicated that the 3-D head data supported more accurate sex classification than the GLI data, across a range of PCA-compressed (dimensionality-reduced) representations of the heads. This kind of dual face representation can give insight into the nature of the information available to humans for categorizing and remembering faces.

An 'other-race effect' for categorizing faces by sex.

It is well-known that people recognize faces of their own race more accurately than faces of other races-a phenomenon often referred to as the 'other-race effect'. Using brief presentations of faces, we show a similar effect for the task of discriminating the sex of a face. Specifically, Caucasian observers discriminated male and female Caucasian faces more accurately/efficiently than did Oriental observers, and Oriental observers discriminated male and female Japanese faces more accurately/efficiently than did Caucasian observers. This result indicates that, under suboptimal viewing conditions, the identification of even the most salient of facial characteristics-face sex-is impaired for other-race faces. This finding suggests, also, that the nature and diversity of our experience with faces may affect not only the quality of the face representation for later access by recognition processes, but also the efficiency of a perceptual discrimination process. Intriguingly, too, we found that female observers, for both races tested, were considerably more accurate at the sex classification task than were male observers.

More about the difference between men and women: evidence from linear neural networks and the principal-component approach.

The ability of a statistical/neural network to classify faces by sex by means of a pixel-based representation has not been fully investigated. Simulations with pixel-based codes have provided sex-classification results that are less impressive than those reported for measurement-based codes. In no case, however, have the reported pixel-based simulations been optimized for the task of classifying faces by sex. A series of simulations is described in which four network models were applied to the same pixel-based face code. These simulations involved either a radial basis function network or a perceptron as a classifier, preceded or not by a preprocessing step of eigendecomposition. It is shown that performance comparable to that of the measurement-based models can be achieved with pixel-based input (90%) when the data are preprocessed. The effect of the eigendecomposition preprocessing of the faces is then compared with spatial-frequency analysis of face images and analyzed in terms of the perceptual information it captures. It is shown that such an examination may offer insight into the facial aspects important to the sex-classification process. Finally, the contribution of hair information to the performance of the model is evaluated. It is shown that, although the hair contributes to the sex-classification process, it is not the only important contributor.

Structural aspects of face recognition and the other-race effect.

The other-race effect was examined in a series of experiments and simulations that looked at the relationships among observer ratings of typicality, familiarity, attractiveness, memorability, and the performance variables of d' and criterion. Experiment 1 replicated the other-race effect with our Caucasian and Japanese stimuli for both Caucasian and Asian observers. In Experiment 2, we collected ratings from Caucasian observers on the faces used in the recognition task. A Varimax-rotated principal components analysis on the rating and performance data for the Caucasian faces replicated Vokey and Read's (1992) finding that typicality is composed of two orthogonal components, dissociable via their independent relationships to: (1) attractiveness and familiarity ratings and (2) memorability ratings. For Japanese faces, however, we found that typicality was related only to memorability. Where performance measures were concerned, two additional principal components dominated by criterion and by d' emerged for Caucasian faces. For the Japanese faces, however, the performance measures of d' and criterion merged into a single component that represented a second component of typicality, one orthogonal to the memorability-dominated component. A measure of face representation quality extracted from an autoassociative neural network trained with a majority of Caucasian faces and a minority of Japanese faces was incorporated into the principal components analysis. For both Caucasian and Japanese faces, the neural network measure related both to memorability ratings and to human accuracy measures. Combined, the human data and simulation results indicate that the memorability component of typicality may be related to small, local, distinctive features, whereas the attractiveness/familiarity component may be more related to the global, shape-based properties of the face.

Learning to see random-dot stereograms.

In the present study some specific properties of the learning effects reported for random-dot stereograms are examined. In experiment 1 the retinal position-specific learning effect was reproduced and in a follow-up experiment it was shown that the position specificity of learning can be accounted for by selective visual attention. In experiments 2 and 3 evidence was obtained that suggests that observers can learn, to a certain degree, monocular random-dot patterns and that this learning facilitates the depth percept. This result indicates that the traditional belief that random-dot stereograms are devoid of monocularly recognizable or useful forms should be reconsidered. In the second set of experiments the learning of two binocular surface properties of random-dot stereograms, depth edges and internal depth regions, was investigated. It was shown in experiment 4 that the depth edges of random-dot stereograms are not learned, whereas the results of experiment 5 indicate that the internal depth regions are learned. Finally, in experiment 6 it was shown that depth edges are learned when the internal depth regions of the stereogram are ambiguous. The results are discussed in terms of the importance of the particular type of stimulus used in the learning process and in terms of perceptual learning and attention.

Structure from stereo by associative learning of the constraints.

A computational model of structure from stereo that develops smoothness constraints naturally by associative learning of a large number of example mappings from disparity data to surface depth data is proposed. Banks of disparity-selective graded response units at all spatial locations in the visual field were the input data. These cells responded to matches of luminance change at convergent, divergent, or zero offsets in the left and right 'retina' samples. Surfaces were created by means of a pseudo-Markov process. From these surfaces, shaded marked and ummarked surfaces were created, along with random-dot versions of the same surfaces. Learning of these example shaded and shaded marked surfaces allowed the system to solve stereo mappings both for the surfaces it had learned and for surfaces it had not learned but which had been created by the same pseudo-Markov process. Further, the model was able to solve some random-dot versions of the surfaces when the surfaces had been learned as shaded marked surfaces.

Associative learning of scene parameters from images.

An important problem for both biological and machine vision is the construction of scene representations from 2-D image data that are useful for recognition. One problem is that there can be more than one world out there giving rise to the image data at hand. Additional constraints regarding the nature of the environment have to be used to narrow the range of solutions. Although effort has gone into understanding these constraints, relatively little has been done to understand how neurallike learning networks may be used to solve scene-from-image problems. A paradigm is proposed in which stochastic models of scene properties are used to provide samples of image and scene representations. Distributed associative networks are taught, by example, the statistical constraints relating the image to the representation of the scene. This technique is applied to problems in optic flow, shape-from-shading, and stereo.

The role of frequency versus informational cues in uncertain frequency detection.

The role of frequency and informational cues in the detection of tones of uncertain frequency was investigated using the probe-signal method [G.Z. Greenberg and W.D. Larkin, J. Acoust. Soc. Am. 44, 1513-1523 (1968)] with auditory and visual test patterns. Patterns consisted of a sequence of events, either tones or visual stimuli, preceding the tone to be detected. Both frequency and informational cues were available in the auditory patterns, whereas only informational cues were available in the visual patterns. Results indicated that observers in the auditory condition displayed trial-by-trial selective attention to one or another frequency band as a function of the cues provided by earlier pattern components. In contrast, listeners in the visual condition displayed simultaneous attention to the two separate frequency bands that could possibly contain the tone, regardless of the information provided by the cues. Results are discussed in terms of single- and multiple-band models of attention.