Face recognition in humans

Humans have a remarkable ability to interpret the visual appearance of a face. Consider that there are 7 and a half billion people on Earth. In many ways, their faces are all quite similar. They all have two eyes, two ears, a nose and a mouth and are very similar in size and shape and configuration. Yet we have no problem at all recognising people we know and never fail to distinguish between two faces we meet, with the possible exception of identical twins. In fact, we are often able to recognise acquaintances that we haven’t seen for decades, even when their appearance has changed dramatically. It is not surprising that humans evolved this extraordinary ability.

Being able to distinguish friend from foe and to ensure offspring were protected would have been essential for survival in early humans. Equally remarkable is our ability to recognise and interpret subtle facial expressions. For example, the tiny difference between genuine, sarcastic and angry smiles. It turns out that the human brain has a dedicated region specifically for visual face processing. It also turns out that our face processing ability exploits regularities in how faces tend to appear. In other words, we have some information about faces hard wired into our brains. For example, faces tend to occur the right way up. Our perception of them is disrupted when we see them inverted. This is well demonstrated by the Thatcher illusion.

Looking at these two upside down faces, it’s not obvious that either of them is unnatural. However, as soon as we invert them, we see that the face on the right has had the eyes and mouth inverted relative to the rest of the face. Once we see the image the right way around and our powerful face processing ability can be used, it is immediately obvious that something is wrong with the face. The observation is that when a face is upside down, the brain doesn’t see it as a face and processes it as it would a general object. Another example is the hollow face illusion.

Since we only ever see convex faces that stick out, when we look into the back of a mask, we can’t accept it is a concave face and flip our interpretation - even if that means we have to reverse the apparent rotation of the face.

Until only 4 or 5 years ago, machines and computer vision were nowhere near matching the face recognition abilities of humans. However, thanks to deep learning that has now been completely reversed such that machines can outperform humans in many scenarios.