Abstract
*contributed equally What makes two faces appear similar to each other? We examined how the dissimilarity of pairs of faces depend upon the faces’ features. We used the Basel Face Space (BFS), a high dimensional coordinate system where axes control facial features and each point is an individual face, to synthesize photorealistic faces. We generated pairs of faces that varied systematically in the angle separating them in the space (θ), and in their distances from the average face (r1 and r2). We used a novel method to collect dissimilarity judgments, in which subjects (n=26) placed each pair of faces vertically on a large touch screen to indicate how dissimilar the two faces appeared. The vertical axis was anchored by reference points denoting identicality and total dissimilarity: pairs of identical face at the bottom of the screen, and pairs of BFS antifaces at the top. We asked to what extent different functions of the BFS coordinates of two faces can predict face dissimilarity judgements. Candidate functions for predicting perceived dissimilarity included (a) linear and (b) sigmoid functions of (1) the BFS Euclidean distance between two faces and (2) linear combinations of the polar angle between faces (θ), and the two eccentricities (r1, r2) of the faces. We fitted models with half the data and compared their predictive performance on the other half. A sigmoid function of Euclidean distance in BFS provided a reasonable approximation to dissimilarity judgments. However, preliminary evidence suggests that similarity does not fall off exactly isotropically in face space as we move away from a reference face in different directions. We relate the results to predictions of alternative models of neuronal coding, including norm-based and exemplar-tuning models. Our results show that BFS provides a useful quantitative model for investigating the representation of faces and predicting perceptual judgments.
Meeting abstract presented at VSS 2015