Training face recognition models requires large and diverse training datasets. Synthetic data can reduce dependence on real identities, but must preserve a useful class structure in addition to visual plausibility. This thesis studies synthetic dataset generation through identity-conditioned latent face swapping. A donor image provides pose, expression, illumination, and background, while the target identity is defined by an ArcFace embedding. The proposed pipeline uses an ID-SiT generative model in the latent space of a pretrained autoencoder, face masks to constrain changes to the facial region, and tangent-Gaussian sampling of target identities. Utility is evaluated indirectly by training a separate face recognition model and testing it on LFW, CFP-FP, CPLFW, AgeDB-30, and CALFW. The main filtered dataset contains 494.409 images and 10.000 synthetic identities and achieves an average verification accuracy of 0.870; conservative identity-diagnostic filtering improves its utility. In the semi-synthetic augmentation scenario, retaining five real images and adding 45 generated images per identity to obtain 50 images in total increases average accuracy from 0.688 to 0.890 compared with the corresponding real-only dataset. In the bias analysis on BFW, the main synthetic dataset exceeds HSFace10K in accuracy (0.721 vs. 0.697), but the accuracy range does not indicate a reduction in bias. The results show that the geometry of the complete training dataset, not only individual image quality, is crucial for synthetic face recognition data.
|
