Final Report Summary - FACEVAR (Face Recognition: Understanding the role of within-person variability)
The aim of this project is to understand face recognition: how can we recognize the people we know with such ease, even across a very wide range of conditions? We focus on within-person variability – an aspect of face perception that has not been studied scientifically before. Our faces change all the time, over short periods while we speak or express emotions, and over longer periods as we age and experience changes in weight and health. Here we examine the ways in which a face becomes familiar to us, as we learn the different ways it can appear.
In one part of the project, we establish how face images themselves vary, by carrying out computational analyses of different photos of the same person. In doing so, we have discovered that different people vary in idiosyncratic ways. In short, the ways that one face varies are not the same as the ways another face varies. This fact seems to be used by the human visual system in order to develop robust representations of familiar people: i.e. methods that allow our brains to recognize the people we know over wide variations.
These studies reveal some interesting results. It seems that our brains do not somehow make judgments about the arrangements of features (e.g. distances between the eyes) as had previously been thought. Instead, the brain makes a more sophisticated statistical model of familiar faces, based on fundamental patterns of light and colour.
While people are very good at recognizing the faces they know, they are much poorer with unfamiliar faces. In this project we have extended this to working passport officers and police, across several countries. We have shown that experienced passport officers are rather poor at matching people to their photos – in fact, they are no better than members of the general public, despite extensive training. We have developed novel ways to improve this situation, and have carried out experiments in which passport photos are replaced by multiple photos, or composites, made from many different photos of the same person. We have also developed some novel techniques for personnel selection: Some people are naturally better than others at face matching, and selecting these people turns out to be a much more efficient way of improving overall performance than focusing on training.
Finally, we ask why society continues to rely on photo-ID, despite clear evidence that it is unreliable. In a series of studies, we have found that viewers do not know just how poor they are at matching unfamiliar faces. Their ability to recognize familiar faces is so good, that they tend to over-generalise this, and believe themselves (falsely) to be experts with all faces. This newly-discovered bias in perception may lead us, as a society, to over-rely on photographic ID.
In one part of the project, we establish how face images themselves vary, by carrying out computational analyses of different photos of the same person. In doing so, we have discovered that different people vary in idiosyncratic ways. In short, the ways that one face varies are not the same as the ways another face varies. This fact seems to be used by the human visual system in order to develop robust representations of familiar people: i.e. methods that allow our brains to recognize the people we know over wide variations.
These studies reveal some interesting results. It seems that our brains do not somehow make judgments about the arrangements of features (e.g. distances between the eyes) as had previously been thought. Instead, the brain makes a more sophisticated statistical model of familiar faces, based on fundamental patterns of light and colour.
While people are very good at recognizing the faces they know, they are much poorer with unfamiliar faces. In this project we have extended this to working passport officers and police, across several countries. We have shown that experienced passport officers are rather poor at matching people to their photos – in fact, they are no better than members of the general public, despite extensive training. We have developed novel ways to improve this situation, and have carried out experiments in which passport photos are replaced by multiple photos, or composites, made from many different photos of the same person. We have also developed some novel techniques for personnel selection: Some people are naturally better than others at face matching, and selecting these people turns out to be a much more efficient way of improving overall performance than focusing on training.
Finally, we ask why society continues to rely on photo-ID, despite clear evidence that it is unreliable. In a series of studies, we have found that viewers do not know just how poor they are at matching unfamiliar faces. Their ability to recognize familiar faces is so good, that they tend to over-generalise this, and believe themselves (falsely) to be experts with all faces. This newly-discovered bias in perception may lead us, as a society, to over-rely on photographic ID.