Usefulness of Synthetic Data in Biometric Recognition

Ocular recognition plays a crucial role in scenarios where masks or partial faces limit the use of face recognition. However, privacy regulations and restrictions have hindered access to large-scale public databases necessary for training and evaluating recognition models. This thesis investigates whether synthetic face data can substitute real images while retaining the detail required for ocular recognition. Ocular crops were extracted from two real datasets, VGGFace2 and AgeDB, and from two synthetic datasets, GanDiffFace and DCFace. We evaluated five ResNet50 models trained under different conditions: VGGFace2 (face), VGGFace2 (ocular), GanDiffFace (ocular), Glint360K (face baseline), and ImageNet (baseline). Identification performance was measured using Rank-1 Accuracy, and verification was measured using Equal-Error Rate (EER). The model trained only on synthetic ocular data achieved 99% Rank-1 Accuracy and 4.9% EER on synthetic test data, but performance dropped to below 15% Rank-1 Accuracy and approximately 32% EER on real ocular images. In contrast, the model trained on real ocular images reached 68% Rank-1 Accuracy and 11.6% EER on real data, and 84% Rank-1 Accuracy and 9.6% EER on synthetic data. These results suggest that while the ocular region of current synthetic faces is suitable for benchmarking, they are not yet a suitable replacement for real data in training and evaluating models intended for real-world deployment.