Inconsistent Face Recognition Performance From VMS
Face recognition is harder than it looks...
We are having trouble getting consistent performance from a not very well-known VMS-based facial recognition technology a customer has purchased – what are we doing wrong?
We have ordinary turret cameras of reasonable quality with a fixed 4mm focal length. We find sometimes we get a read and sometimes not – it’s too inconsistent to be of much help. What would you recommend? Should we use different cameras, focal lengths, or is the system at fault?
A: The system may not be at fault, but some VMS-based face recognition may not be as discerning or highly developed as a dedicated face recognition solution – you’ll need to experiment. When it comes to face recognition, we’d be looking for longer focal lengths – around 9mm for a 1/3-inch sensor. Longer focal lengths will give you flatter images with less distortion and more accurate measurements – vital for face recognition.
Any wide-angle lens is going to be stretching and bending the scene, especially around the edges – the wider, the more pronounced the effect will be. There’s no chance people distorted into cone heads are going to be recognisable to face recognition software. A 4mm lens is not too wide but it will still mess with people’s heads.
We are going to be looking at some face recognition software at SecTech Camera Shootout – if you’ve not seen it by then, you should come along and check that out, as well as any other face recognition software being shown on the stands at SecTech Roadshow.