LOL - Look How Low Megvii Face++ Put Their Cameras - A Lesson In Face Recognition Weakness

Business Insider has a nice profile on Megvii Face++, one of the 2 Chinese startups that have received massive investment for their facial recognition. The profile has many pictures, one thing they did not call out, that is worth doing so is this image I've marked up with the red box for the cameras and the arrow showing the height relative to the people's faces being scanned.

This is no accident. Indeed, from speaking with Megvii's representatives at ISC West in Vegas and IFSEC in London, they have firm height restrictions, i.e., the cameras cannot be mounted too high (they've said 2 meters) or else performance will degrade significantly.

On the one hand, the higher you mount a camera, the less of their face it sees (and the more of the top of their head). From our testing:

As such, it is not surprising that facial recognition performance would suffer.

On the other hand, think about the real world challenges. In your own office, with your own employees, you can mount an array of cameras 4' high and expect employees to look at them. Not so in public. Cameras tend to be mounted 3, 4, 5 meters or more (think about off light pole or traffic poles, etc.).

Relevant here is the NY Times excellent investigation, Inside China’s Dystopian Dreams: A.I., Shame and Lots of Cameras, in which they observe:

Chinese authorities regularly state, and overstate, their capabilities. In China, even the perception of surveillance can keep the public in line.

But if you are real user wanting to actually benefit from facial recognition, keep this issue in mind.

Login to read this IPVM discussion.
Why do I need to log in?
IPVM conducts unique testing and research funded by member's payments enabling us to offer the most independent, accurate and in-depth information.

*** ********* ************ ************ (*************), **** ***** ** **** *****. For ****** ******* ** ***** ********** *********** (****** **, ********, welfare, ***.) * ***** ***** *** *********** ** *** ******* would ******** **** ***********.

****, ************* *** ******, **** **** ****** ******** ** ******** to ** ************ ****** ************, ****** ******* ** ***** :(