Communications of the ACM
How Well Do Facial Recognition Algorithms Cope With a Million Strangers?
The University of Washington MegaFace Challenge was the first competition aimed at evaluating and improving the performance of facial recognition algorithms at the million-image scale.
Credit: University of Washington
Assessing the performance of face-recognition algorithms at the million-person scale is the purpose of the MegaFace Challenge hosted by University of Washington (UW) researchers.
"We need to test facial recognition on a planetary scale to enable practical applications--testing on a larger scale lets you discover the flaws and successes of recognition algorithms," says UW professor Ira Kemelmacher-Shlizerman. "We can't just test it on a very small scale and say it works perfectly."
The researchers initially compiled a dataset of 1 million publicly available Flickr images, representing 690,572 unique individuals. Facial-recognition teams then were challenged to download the database and see how their algorithms fared in differentiating between 1 million possible matches. The algorithms were evaluated on verifying whether two photos were of the same person, and identifying matches to the photo of one individual to a different picture of the same person among 1 million "distractors."
Google's FaceNet exhibited the strongest performance on one test, slipping from near-perfect accuracy when dealing with a smaller number of images to 75% accuracy on the million-person test; other algorithms that did well at a small scale declined by much larger percentages when performing the tougher task, to as low as 33%t accuracy.
From UW Today
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA
No entries found