A little over 2 years after I started this blog, End Game is now available. I hope you will enjoy it and find it useful. If you’ve ever wanted to contribute to this blog, getting this book is a gre…
Commercial face recognition doesn’t work
The past years have seen a huge expansion in security and surveillance efforts, especially in airports. A lot of the surveillance revolves around correct identification of a particular person - for example, are you the person in your passport picture? Are you the person that is on that CCTV video?
Underpinning all of these systems is the notion that we can reliably establish the identity of an unknown person by comparing their appearance to an image (like a passport photo). But really, how reliable is this? I’m sure that some of you (myself included) have looked at their passport photo and thought - I look nothing like that now. My passport photo was taken nearly 9 years ago now - I would be lying if I said I still looked that way.
The photo above demonstrates the problem - photographs can be incredibly misleading. Humans find the task of matching photographs of unfamiliar people extremely difficult - yet security officers in airports are expected to do it hundreds of times a day. This task turns out to be highly error prone.
The researchers Kemp, Towell & Pike tested this. They gave supermarket check out staff a task: validate a photo credit card by deciding whether or not the person presenting the card matched the photo on it. Even though they were aware they were taking part in the study, they performed surprisingly badly, with about 50% of the cards being accepted when the wrong person presented them.
What about machine face recognition though? Is that better?
In short - no. There is an even higher error rate in commercial face recognition systems. Apparently a high false alarm rate in the automatic face recognition system at Manchester Airport in the UK was causing delays, and the response was to recalibrate the system so that it would accept a 30% likeness match in passport photos.
In real world automatic face recognition the performance levels are actually unusable, not just poor. 30% is rubbish - think of how little that actually is.
Why is this happening? Well, human faces actually all share the same basic template, so physical differences are pretty subtle. Eyes are usually in the same place, ears are, mouth etc. What does differ from picture to picture is the features within a face - the lighting, pose, age, hairstyle, weight, make up, expression, quality of the picture etc can all affect how your face looks, and so it becomes incredibly difficult to match two different pictures up correctly sometimes. Photos of different people can actually be more similar than photos of the same person taken at different times - like in the top picture.
So - what is the answer?
So, perhaps matching a photograph to an actual person isn’t a good idea. The suggestion is that perhaps we should be looking into alternative face representations that are more stable. There is no reason to assume that human beings should perform well when identifying a frozen snapshot - we didn’t evolve to recognise photographs, we evolved to recognise people.
Researchers (Rob Jenkins & David White) have suggested that because in real life our faces are dynamic, and constantly changing, that we should take this into account when creating ID photographs. Dr. Rob Jenkins, University of Glasgow, took 20 photographs, averaged them out, segmented them from the background and calculated landmarks of the face to create stabilised images of the face.
The photo above shows the result - even though these images look incredibly like the person in the photo - they are not real images. They are not single photographs that someone took - they are photographs created by researchers and yet they form a striking likeness.
When the researchers tested photos like these - they found that people were much better at matching unfamiliar faces to photos.
Of course, this is just one suggestion, and not the only way or best way proposed, but a very interesting approach nonetheless.
More than 20 years of research have established that neither humans nor computers can reliably match unfamiliar face photos. Perhaps it is time to abandon the idea that photographs provide a useful means of identifying unknown people.
My mind is like a racing engine, tearing itself to pieces because it is not connected up with the work for which it was built.
“
| — | Sherlock Holmes in His Last Bow, by Arthur Conan Doyle |
