Over the past two decades, computational models of face recognition have led to important insights about the properties of human memory for faces. Recent comparisons between humans and state-of-the-art face recognition algorithms have demonstrated that the best computer systems are now more accurate than humans on some challenging face recognition tasks. I will review human-machine comparison studies that consider both qualitative and quantitative aspects of face recognition. The results of these studies suggest that although computational models have made enormous progress solving face recognition tasks, the current "state-of-the-art" for machines is only at the level of human abilities for unfamiliar face recognition. The recognition of familiar faces is fast and robust against changes in illumination, viewpoint, and facial expressions. I will argue that real human expertise for faces lies, not in our ability to selectively remember so many individual faces, but rather in our impressive abilities to recognize fewer people, with great reliability and robustness. The key to this expertise may lie in the representations we make of the whole person, including the shape and motions of their body and face. For this reason, it is important to study not only face recognition, but to consider how person recognition unfolds in natural viewing environments.
Sunday, April 08, 2012
Free and open to the public