IPVMU Certified | 08/19/14 05:24pm
Even without details, this claim raises red flags to me:
"While humans recognize faces at an accuracy rating of 97.53% on Labeled Faces in the Wild, the recognition system developed by CUHK was tested using thousands of picture sets, and it can recognize faces at an accuracy of 99.15%, regardless of changes in lighting, make-up and camera angles.
First: "99.15%" of how many known "labeled" faces? The release says 'thousands', but is careful not to directly associate it with the false detect rate of 'unlabeled' faces or give an exact number. Matching smaller datasets is much easier than large populations. There's more data to compare and closer variations among potentially matched faces.
Second: There's lots of fluff talk about 'advanced algorithms', but what expectations are given to the camera front-end? No elaboration on PPF, lightning, or camera positioning requirements are offered - especially in the non-optimal way that surveillance cameras are deployed.
Third: Were faces 'detected' one at a time, in a single-file queue line stopping in front of a camera, or was it a mass of people marching together in a crowd? A lab test often doesn't correlate to production use.
To me, reading this comes off as marketing fluff for any other countless face rec platforms. We are given an impressive success percentage, but in explaining this, the release claims nothing more than better 'secret sauce' when so many physical technology barriers must be addressed to be fundamentally different.
The second part of the claim sounds completely irresponsible:
"It can recognize faces at an accuracy of 99.15%, regardless of changes in lighting, make-up and camera angles"
It's physically impossible for it not to be impacted by 'lighting and camera angles'.
Money picture from CUHK:
Makes you wanna get one of those masks...
Though back in the real world even the NIST admits facial recognition keeps improving...
The National Institute of Standards and Technology (NIST) reports that results from its 2013 test of facial recognition algorithms show that accuracy has improved up to 30 percent since 2010...
In both years the study used a database of 1.6 million faces. In 2010, the images were frontal "mugshot" images from law enforcement agencies that closely comply with the ANSI/NIST ITL 1-2011 Type 10 standard. In 2013, researchers added a small database of images taken for visa applications that meet an ISO/IEC (International Organization for Standardization/International Electrotechnical Commission) standard and 140,000 webcam images taken in poorly controlled environments that do not comply with any standard...
The study also shows that rates of missing facial matches increase as the database size increases as expected, but that it does so only slowly. When the number of facial images increased by a factor of 10—from 160,000 to 1.6 million—the error rate only increased by about 1.2 times. This slower-then-expected growth in error rates occurs in many natural phenomenon, and "is largely responsible for the operational utility of face identification algorithms," explains Grother
P.S. I have the identified face in top right box above: Ray "Scarecrow" Bolger