Face similarity space as perceived by humans and artificial systems

The performance of a local feature based system, using Gabor filters, and a global template matching based system, using a combination of PCA (principal component analysis) and LDA (linear discriminant analysis) was correlated with human performance on a recognition task involving 32 face images. Both systems showed qualitative similarities to human performance in that all but one of the calculated correlation coefficients were very or moderately high. The Gabor filter model seemed to capture human performance better than the PCA-LDA model since the coefficients for this model were higher for all examined conditions. These results indicate that the preservation of local feature based representation might be necessary to achieve recognition performance similar to that of humans