Model Performance Metrics
True Positive Rate at specified False Positive Rate: IJB-C Dataset
False Positive Rate | LITE_V3 | TFV5 | TFV5_2 | TFV6 | TFV7 |
1 in 1,000 | 0.97924 | 0.97827 | 0.98021 | 0.97893 | 0.98359 |
1 in 10,000 | 0.96942 | 0.96733 | 0.97019 | 0.96993 | 0.977704 |
1 in 100,000 | 0.95286 | 0.94657 | 0.95332 | 0.95245 | 0.96661 |
Receiver Operating Characteristic (ROC) Tables
Use the following ROC curves and tables to choose the optimal operating threshold for your use case. For guidance on how to interpret an ROC curve, consult this blog post.
Below we provide operating similarity-score thresholds based on evaluations run on various datasets. These metrics must be considered representative or illustrative, not controlling, as a Licensee's input data will vary from the tested datasets described herein. All Licensees are encouraged to evaluate performance metrics and thresholds using their environment data.
IJB-C
You can learn about the characteristics of the IJB-C dataset here: https://noblis.org/wp-content/uploads/2018/03/icb2018.pdf
Similarity Score Threshold at False Positive Rates
False Positive Rate | LITE_V3 | TFV5 | TFV5_2 | TFV6 | TFV7 |
1 in 1,000 | 0.28803 | 0.26996 | 0.26620 | 0.26210 | 0.22404 |
1 in 10,000 | 0.36326 | 0.34768 | 0.34423 | 0.33785 | 0.29976 |
1 in 100,000 | 0.44098 | 0.44167 | 0.43454 | 0.41902 | 0.38314 |
Private Dataset [deprecated]
The following ROC curves were generating using a private dataset. The dataset primarily consists of male faces
LITE V2
TAR at FPR 10^-5: 0.9032
LITE
TAR at FPR 10^-5: 0.8487