PRINT BOOKMARK

Can machines out-perform humans at facial recognition?

Nick Whitehead | 06 Feb 2015 | Comments

Recent controversy about the retention of facial images by Police, where the subjects have no criminal record, has also led to discussion of the performance of facial recognition technology. An officer with the Met Police has been quoted as saying that people, referred to as “super recognisers”, dramatically outperformed automated facial recognition systems when looking at CCTV video imagery. And even with high quality imagery captured at the airport, the officer states that the automated recognition only gets it right 90% of the time (source: BBC).

Given the claims of accuracy from manufacturers of facial recognition engines these are disappointing statements. So, what is the reality here?

Accuracy claims

Facial recognition vendors submit their technology to independent testing to substantiate the claims for accuracy. However, translation of this bench test performance into real systems is much more challenging. Facial recognition engines are tuned to perform well on the test databases but when faced with imagery from other sources the accuracy is degraded.

It is worth noting, however, that the accuracy of the “super recognisers” mentioned above is assumed to be perfect. This isn’t necessarily the case. There are a number of high profile incidents of poor recognition with staff that have been trained to confirm a passenger’s identity when compared to their passport (source). We have examined the application of facial recognition systems to prevent just this kind of issue and conclude that they have an important role to play (source).

“Horses for courses”

In my opinion, facial recognition when correctly applied, can out-perform humans. It is certainly more consistent in the quality of its comparison in that it doesn’t get tired, distracted or bored. Our approach is to use automation to replace these mundane activities but tuned to a level which means that, when certainty cannot be offered, a human is available to manage the exception process. Humans are really good at recognition in this context.

If you’re interested in learning more about facial recognition and biometric technology, you might be interested in the following Angles article.