Leveraging Billions of Faces to Overcome Performance Barriers in Unconstrained Face Recognition
چکیده
We employ the face recognition technology developed in house at face.com to a well accepted benchmark and show that without any tuning we are able to considerably surpass state of the art results. Much of the improvement is concentrated in the high-valued performance point of zero false positive matches, where the obtained recall rate almost doubles the best reported result to date. We discuss the various components and innovations of our system that enable this significant performance gap. These components include extensive utilization of an accurate 3D reconstructed shape model dealing with challenges arising from pose and illumination. In addition, discriminative models based on billions of faces are used in order to overcome aging and facial expression as well as low light and overexposure. Finally, we identify a challenging set of identification queries that might provide useful focus for future research. 1 Benchmark and results The LFW benchmark [6] has become the de-facto standard testbed for unconstrained face recognition with over 100 citations in the face recognition literature since its debut 3 years ago. Extensive work [15, 14, 13, 5, 7, 4, 10, 3, 8, 9, 11, 16] has been invested in improving the recognition score which has been considerably increased since the first non-trivial result of 72% accuracy. We employ face.com’s r2011b face recognition engine to the LFW benchmark without any dataset specific pre-tuning. The obtained mean accuracy is 91.3% ± 0.3, achieved on the test set (view 2) under the unrestricted LFW protocol. Figure 1 (a) presents the ROC curve obtained in comparison to previous results. Remarkably, much of the obtained improvement is achieved at the conservative performance range, i.e., at low False Acceptance Rates (FAR). 1face.com has a public API service [1] which currently employs a previous version of the engine. 1 ar X iv :1 10 8. 11 22 v1 [ cs .C V ] 4 A ug 2 01 1