What Does 3VR Face Recognition Korean Results Mean?

Recently, results from a facial recognition test for the Korean National Police were reported [link no longer available] widely by 3VR of obtaining: "85-92 percent accuracy in recognizing and matching faces in a few crowded, highly-trafficked public train stations in Seoul."

The manufacturer claims these results are the best ever in any public video facial recognition study (compared to only 60% historically). One of the security magazine editors suggested that this was reason to "change the standard analytics copy" 

A Korean National Police spokesman praised the results, expressing interest and confidence in deploying facial recognition widely. It is always impressive to see end users expressing confidence in new surveillance technologies.

The Importance of Test Results

In the security industry, test results are key benchmarks for buyers and integrators in evaluating technologies. Improvements in test results can be important drivers for expanding the use of technologies. As such, accuracy and clarity in test results are essential in ensuring successful deployments.

In facial recognition test results, it is common that detailed reports are released and in-depth quantifications are offered. These tests are often administered by government agencies. (Read a facial recognition overview and examine the 2006 US government study [link no longer available] for good examples and background.)

The Korean National Police Test

Three material differences should be considered when comparing the Korean National Police Test to other previous publicly released studies:

• While the test was sponsored by the Korean government, the tester, SK Networks, is a corporation that sells technology products, not an independent testing agency.
• The tester, SK Networks, is marketing 3VR products (see SK Networks' marketing website and a 20 page marketing proposal for 3VR products, both dated 2008 [Note: the website has since been taken down]). 3VR reports that they were not aware of these documents and that they have asked SK Networks to remove them from the Internet.
• No details were publicly provided on false error rates, system setup, or testing methodology. Small variances in any of these elements can result in significantly different accuracy statistics. 3VR reports that a 170 page test result document is available in Korean (NIA pub number NIA 111-RER-07113/2007.12) which they are in the process of translating but have not been given permission to share publicly.

Concerns about The Test

3VR's claim that this test result is the best ever raises concerns. First, since we cannot read the test results, it cannot be assessed how comparable the implementation and design of this test is to leading global scientific studies. Secondly, unlike other facial recognition tests conducted by independent testing agencies, SK Networks sells technology products and has a general interest in promoting the purchase and use of such products.

Setting appropriate expectations for advanced technology projects is crucial, especially when statistics are being cited. This is why the community depends on scientific, independent studies that can be peer reviewed by outside experts.

It is promising to hear end users interested in using facial recognition and pleased with their test results. However, comparable statistical claims require far higher scrutiny and clarity than provided yet for the Korean NPA test results.