Accurate Facial Recognition Requires Accurate Data
We have all heard the phrase “garbage in, garbage out”. The first use can be traced to 1957 when an article in a syndicated newspaper described the work of US Army mathematicians on early computers. The inference over 60 years ago was that "sloppily programmed" inputs inevitably lead to incorrect outputs.
Today, as we struggle to manage the information deluge in the era of Big Data, too many tech solutions are based on the belief that the size of the database is the determining factor for delivering value. But in fact, especially in the case of facial recognition, it is quality not quantity that enables today’s algorithms to really provide benefit. And too often, the data that is used to train the algorithms is not complete or accurate.
Recent research from Joy Buolamani at the MIT Media Lab’s Civic Media group delivered some disturbing results. She tested three face recognition systems using a data set with 1,270 faces that included lawmakers from three African nations and three Nordic countries and included a high percentage of women. She found that a person would be accurately recognized 99% of the time - if the image was of a white man. But the darker the skin, the more inaccurate the system was – with up to a nearly 35% error rate for images of darker skinned women.
I first encountered this issue many years ago when deploying a solution in an urban hospital, Northwest Memorial Physicians Group. Many of the physicians and nurses were African-American and most of the rooms used less than ideal lightning. This presented two major challenges to accurate facial recognition: the light was not optimal, and the users had darker skin tones which made it more difficult for the cameras to accurately resolve the required facial features.
We refined our algorithms to leverage a database of worldwide users that was statistically significant and included people from not just Western countries, but Latin America, India and Asia. The solution needed to be skin color-neutral and focus on capturing facial features and their relationships. Another critical challenge was to make our solution work in most lighting situations – whether the user was outside in the brightest daylight or even in complete darkness.
Today, we are able to deliver a mobile 3D facial recognition solution that is consistently accurate regardless of a user’s skin tone or lighting conditions. We have been able to provide accurate data going in and value coming out – rather than the “garbage” scenario.
(image: Randallbritten CC BY-SA 3.0 https://creativecommons.org/licenses/by-sa/3.0/ )