Jupiter Notebooks doesn't work within GitHub preview by unknown reasons, so please pull the project locally.
| Name | Description | Link |
|---|---|---|
| celebA_train_500 | The dataset is used to train network | https://disk.yandex.ru/d/S8f03spLIA1wrw |
| celebA_ir | The dataset is used to calculate TPR@FPR | https://disk.yandex.com/d/KN4EEkNKrF_ZXQ |
| Network | Test Accuracy | TPR@FPR (fpr=0.05) | TPR@FPR (fpr=0.1) | TPR@FPR (fpr=0.2) | TPR@FPR (fpr=0.5) |
|---|---|---|---|---|---|
| ResNet18 + Standard Cross-entropy Loss | 0.75 | thr = 0.69, tpr = 0.65 | thr = 0.67, tpr = 0.76 | thr = 0.63, tpr = 0.87 | thr = 0.57, tpr = 0.97 |
| ResNet18 + ArcFace + Cross-entropy Loss | 0.71 | thr = 0.41, tpr = 0.43 | thr = 0.29, tpr = 0.58 | thr = 0.19, tpr = 0.76 | thr = 0.07, tpr = 0.95 |
Networks weights are located in /trained directory
[1] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou ArcFace: Additive Angular Margin Loss for Deep Face Recognition link
[2] Jiaheng Liu1, Haoyu Qin2, Yichao Wu 2, Ding Liang AnchorFace: Boosting TAR@FAR for Practical Face Recognition link