Self-Driving Cars Aren't Good At Seeing Pedestrians With Dark Skin, According To New Study
The study found the cars had no problem "seeing" white pedestrians, however.
A new study has discovered self-driving cars may not recognize people with dark skin as easily as white people.
Like what you're reading?
Get more in your inbox.
Futurism reports Georgia Institute of Technology researchers recently published a study in arXiv that shows the artificial intelligence powered self-driving vehicles struggle with detecting pedestrians with dark skin.
The study, “Predictive Inequity in Object Detection,” tested the accuracy of object detection models, the code used by AI to "see" people and things in the surrounding environment, using the Fitzpatrick scale, a numerical classification for human skin color.
According to Vox, the study's authors took pictures of pedestrians of different races and ranked their skin tones using the Fitzpatrick scale. They then ran the photos through the object detection models to see how often the models detected the pedestrians with darker skin.
Blavitize your inbox! Join our daily newsletter for fresh stories and breaking news.
The models processed the photos of the three pedestrians with the darkest skin with “uniformly poorer performance” compared to the other images. Specifically, the models were 5 percent less accurate when tasked with detecting the three darkest skinned people, no matter the time of day or conditions on the road.
The study has not yet been peer-reviewed, so it technically has not been proven accurate by the science community. However, given most automotive companies do not publicly release their studies of the models, many in the AI community are advocating for the study's results to be taken into consideration.
The Georgia Tech team is also pushing for car manufacturers to take the results seriously and suggests the companies add more images of dark-skinned people to the models' training protocols.
Now, check these out: