Abstract
The human face usually provides reliable information for identifying someone. When illumination and expression variability make it impossible to identify a person from the face, people use the body, but are unaware of doing so (Rice et al., 2011). The goal here was to examine humans’ use of the face versus body for identification when the quality of identity information in the face varied. Participants matched identity in pairs of same-identity and different-identity images. Same-identity images were selected using state-of-the-art face recognition algorithms from an international competition, based on the algorithm-computed similarity of the faces in the pair. Image pairs were stratified into good, moderate, and poor groups, ranking from the highest (highly similar images of the same person) to lowest (highly dissimilar images of the same person) similarity scores. We compared identification accuracy for image pairs (n=60) digitally edited to include only the face or the body. Humans judged whether the pairs showed the "same person" or "different people". We found that people performed more accurately with the face than with the body (F(1, 42)= 58.55, p <.0001). Human performance mirrored the algorithm-generated face-quality stratifications, with match accuracy decreasing monotonically from the good, to the moderate, to the poor condition (F(2, 84)= 98.86. p <.0001). We also found an interaction between the information available (face versus body) and the image pair groups (good, moderate, and poor) (F(2, 84)= 8.28, p <.0006). This interaction indicated that as the quality of the identity information in the face decreased, the gap between identification accuracy for the face versus body likewise decreased. In summary, both the face and the body contribute to identification accuracy. The relative importance of the face versus body, however, can shift with the quality of information in the face.
Meeting abstract presented at VSS 2013