Face Recognition Performance
Sighthound is a leader in deep learning, specializing in building intelligent convolutional neural networks that can run real time in resource-constrained environments, including on embedded systems. Sighthound's facial recognition system leads the world, with measured accuracy of 99.2% against PubFig200, some ten percentage points clear of other commercial or academic methods.
Our model uses 3.8 million images, or about 1.5% of the number of images Google lists for training FaceNet. Our pipeline uses only a single crop of the face. Crops are partial areas of training images used to enhance results. However each additional crop slows down performance in recognition tasks, meaning the models are optimized for test-taking, not for real world use. Sighthound is not specifically tailored to verification tasks, and the same model can be used with industry-leading performance on identification tasks.
The result is that Sighthound’s software is highly accurate, robust to a variety of real world use cases and runs in real time.
PubFig200 Face Recognition Accuracy
Sighthound Inc. All PubFig200 images removed from model training
Deep Reconstruction Models for Image Set Classification, Munawar Hayat, Mohammed Bennamoun and Senjian An, PAMI15
P. Zhu, L. Zhang, W. Zuo, and D. Zhang, “From Point to Set: Extend the Learning of Distance Metrics”, ICCV13
M. Yang, P. Zhu, L. V. Gool, and L. Zhang, “Face Recognition Based on Regularized Nearest Points Between Image Sets”, FG 2013
E. Ortiz, A. Wright, and M. Shah, “Face Recognition in Movie Trailers via Mean Sequence Sparse Representation-Based Classification", CVPR13
Y. Hu, A. S. Mian, and R. Owens, “Face Recognition Using Sparse Approximated Nearest Points Between Image Sets”, PAMI12
PubFig 200 is a data set of 58,797 images of 200 public figures. All academic papers cited with the presentation dates. All commercial systems tested in March 2016. Microsoft declined to make its system available for testing.