Table 3.
The achieved verification performance of different experimental settings by ResNet-100, ResNet-50, and MobileFaceNet models along with EUM trained with triplet loss and EUM trained with SRT loss. The result is reported using synthetically generated masked faces of the LFW dataset. The FMR100_ThUMR-UMP are equal to 0.1736, 0.2052 and 0.2449 for ResNet-100, ResNet-50 and MobileFaceNet, respectively. The FMR1000_ThUMR-UMP are equal to 0.2451, 0.2617 and 0.3450 are for ResNet-100, ResNet-50 and MobileFaceNet, respectively. The lowest EER and the lowest average error of FMR100 and FMR1000 at the defined threshold are marked in bold. It can be noticed the significant improvement in the verification performance induced by our proposed approach (SRT) in most evaluation cases.
| FMR100_ThUMR-UMP |
FMR1000_ThUMR-UMP |
||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LFW | Setting | EER% | FMR100% | FMR1000% | FMR% | FNMR% | Avg.% | FMR% | FNMR% | Avg.% | G-mean | I-mean | FDR |
| UMR-UMP | 0.2660 | 0.2667 | 0.3333 | 1.0000 | 0.2667 | 0.6333 | 0.1000 | 0.3333 | 0.2167 | 0.7157 | 0.0026 | 33.0630 | |
| UMR-MP | 1.0000 | 0.9667 | 2.5667 | 1.0667 | 0.9667 | 1.0167 | 0.0667 | 2.9333 | 1.5000 | 0.5220 | 0.0019 | 13.1746 | |
| UMR-MP(T) | 1.7000 | 2.3667 | 4.4333 | 0.9667 | 2.5333 | 1.7500 | 0.0333 | 5.9667 | 3.0000 | 0.4115 | 0.0029 | 11.0452 | |
| UMR-MP(SRT) | 0.8667 | 0.8667 | 1.6000 | 1.2667 | 0.7667 | 1.0167 | 0.1000 | 1.7333 | 0.9167 | 0.5380 | 0.0024 | 15.0505 | |
| MR-MP | 0.9667 | 0.9667 | 2.4333 | 3.1000 | 0.7000 | 1.9000 | 0.5000 | 1.2000 | 0.8500 | 0.5996 | 0.0110 | 14.2278 | |
| MR-MP(T) | 1.7333 | 2.3333 | 10.1333 | 19.4667 | 0.3667 | 9.9167 | 6.7667 | 0.6667 | 3.7167 | 0.6290 | 0.0808 | 10.8161 | |
| ResNet-100 | MR-MP(SRT) | 0.9667 | 0.9667 | 2.0667 | 3.0000 | 0.6667 | 1.8333 | 0.4667 | 1.5333 | 1.0000 | 0.6035 | 0.0053 | 14.6018 |
| UMR-UMP | 0.3333 | 0.3000 | 0.4000 | 1.0000 | 0.3000 | 0.6500 | 0.1000 | 0.4000 | 0.2500 | 0.7023 | 0.0029 | 26.5107 | |
| UMR-MP | 1.4667 | 1.8333 | 3.3000 | 1.0000 | 1.8333 | 1.4167 | 0.1000 | 3.5667 | 1.8333 | 0.5117 | 0.0014 | 11.8522 | |
| UMR-MP(T) | 2.0000 | 2.7000 | 4.9667 | 0.6333 | 3.3333 | 1.9833 | 0.0667 | 6.6333 | 3.3500 | 0.4278 | 0.0020 | 10.5553 | |
| UMR-MP(SRT) | 1.1000 | 1.1333 | 2.4000 | 0.9667 | 1.1333 | 1.0500 | 0.2000 | 2.2000 | 1.2000 | 0.5427 | 0.0016 | 14.5079 | |
| MR-MP | 1.3667 | 1.7333 | 4.7333 | 3.0000 | 0.8333 | 1.9167 | 0.9000 | 1.9333 | 1.4167 | 0.5893 | 0.0158 | 12.2339 | |
| MR-MP(T) | 2.0333 | 2.9667 | 7.2000 | 10.8667 | 0.7667 | 5.8167 | 4.0333 | 1.5333 | 2.7833 | 0.6256 | 0.0525 | 10.2560 | |
| ResNet-50 | MR-MP(SRT) | 1.2333 | 1.4333 | 2.9667 | 2.2333 | 0.9333 | 1.5833 | 0.6333 | 1.5333 | 1.0833 | 0.6051 | 0.0053 | 13.4416 |
| UMR-UMP | 0.6333 | 0.6000 | 1.3000 | 1.0000 | 0.6000 | 0.8000 | 0.1000 | 1.3000 | 0.7000 | 0.6742 | 0.0051 | 18.2460 | |
| UMR-MP | 3.2333 | 5.9333 | 12.0333 | 0.7667 | 6.7333 | 3.7500 | 0.0000 | 18.2667 | 9.1333 | 0.4641 | -0.0011 | 7.5840 | |
| UMR-MP(T) | 3.6667 | 7.1333 | 17.6667 | 0.6000 | 8.7667 | 4.6833 | 0.0000 | 27.4333 | 13.7167 | 0.4023 | 0.0013 | 7.2341 | |
| UMR-MP(SRT) | 1.8667 | 2.4667 | 8.1333 | 0.8333 | 2.8667 | 1.8500 | 0.1000 | 9.3667 | 4.7333 | 0.5144 | 0.0006 | 10.2266 | |
| MR-MP | 3.3333 | 6.4667 | 17.9000 | 5.7667 | 2.6333 | 4.2000 | 0.8333 | 7.1333 | 3.9833 | 0.5688 | 0.0505 | 7.7096 | |
| MR-MP(T) | 3.0667 | 5.2000 | 13.6333 | 93.9000 | 0.0000 | 46.9500 | 72.1333 | 0.0667 | 36.1000 | 0.7495 | 0.3970 | 7.7594 | |
| MobileFaceNet | MR-MP(SRT) | 2.2667 | 3.5333 | 11.1000 | 2.3000 | 2.2333 | 2.2667 | 0.4667 | 5.9667 | 3.2167 | 0.5872 | 0.0091 | 9.6183 |