Tampa Police Give Up On Face Recognition Cameras

Saint Aardvark writes "The city of Tampa has given up on their face-recognition system attached to street surveillance cameras."

Comments

Common problem with recognition systems.

[AlecC]

This system seems to have tripped across a common problem with all id recognition systems - face, retina, voice, fingerprint, whatever. That is that they are used in two completely different modes.

One mode is the verification mode: this person claims to be Mr XYZ: is he? For this purpose, you only have one identity to match. If the answer comes out "maybe" instead of "yes" or "no", you can take another photo/scan/whatever. You can use extremely number intensive checking techniques because you are only trying to match ONE face/eye/... to ONE record. And the people being checked have at least some incentive to help their system (remove glasses, get a rescan when they have hair cut or grow beard). Systems can be made to do this very reliably in this mode - call it mode 1.

You can scale this up a little bit, while maintaining reliability. A car, for example, might recognise the voices of four registered drivers and adjust itself to suit, or a secure area form a few tens of people. Call this mode 1A.

The second mode is when you are trying to detect any one of a large list of possible people in a huge crowd, when they may have changed their characteristics significantly, either intentionalyy or unintentionally. Call this mode 2.

The trouble is that a lot of people assume that, if you can scale from 1 to 1A, the scaling from 1A to 2 will be linear. Which it won't. As well as the linear scaling of vastly more records to match (a linear scaleing), there is the the no-rescan, chjanged face, uncooperative facto, which acts quadratically with the fist. This means the problem explodes uncontrollably very soon.

Some of the people making this assumption should know better.