TABLE 3.
Verification performance measures FDR, G‐mean, and I‐mean achieved by SphereFace on different experimental setups
| SphereFace | EER | FMR100 | FMR1000 | ZeroFMR | G‐mean | I‐mean | FDR | FTX |
|---|---|---|---|---|---|---|---|---|
| BLR‐BLP | 2.5237% | 3.5856% | 9.1958% | 43.1286% | 0.703659 | 0.132542 | 9.2316 | 0.00000 |
| BLR‐M12P | 15.6472% | 57.7897% | 79.1916% | 96.8344% | 0.280806 | 0.039197 | 1.8983 | 4.736% |
| M12R‐M12P | 25.5317% | 78.5115% | 91.3690% | 99.5655% | 0.533365 | 0.266994 | 0.9161 | 4.736% |
| BLR‐SMP | 15.4323% | 31.6846% | 46.3397% | 76.9503% | 0.415802 | 0.04813 | 2.6162 | 5.114% |
| SMR‐SMP | 12.7381% | 23.9176% | 39.1283% | 68.7882% | 0.55719 | 0.343760 | 2.332 | 5.371% |
Note: the performance degradation induced by the masked face probes.
Abbreviations: BLP, baseline probe; BLR, baseline reference; EER, equal error rate; FDR, Fisher discriminant ratio; FMR, false match rate; FTX, failure‐to‐extract rate; M12P, mask probe of the combined first and second capture scenarios; M1R/M2R/M12R, mask reference of the first, second, and combined capture scenarios, respectively; SMP, simulated mask probe; SMR, simulated mask reference.