Facebook’s Artificial Intelligence team is using deep learning to create a software that will perform facial verification. DeepFace can identify whether unfamiliar faces in 2 different photos belong to the same person, scoring 97.25% on the test compared to humans who score 97.53% on average. The new software is still at the research stage, but Facebook could use its capabilities to suggest whom users should tag in specific photos.
The performance of DeepFace’s AI was tested using 3 face representation data sets: the Social Face Classification dataset, the Labelled Faces in the Wild database and the Youtube Faces dataset. For a single image, DeepFace runs the whole image verification process at 0.33 seconds on a single core Intel 2.2 GHz processor. DeepFace will be presented at the IEEE Conference on Computer Vision and Pattern in June to get feedback from the research community. The abstract to the DeepFace research paper and a dow
nload link for the pdf can be found here.