Analyzing and Reducing the Damage of Dataset Bias to Face Recognition With Synthetic Data

It is well known that deep learning approaches to facerecognition suffer from various biases in the available train-ing data. In this work, we demonstrate the large potentialof synthetic data for analyzing and reducing the negativeeffects of dataset bias on deep face recognition systems. Inparticular we explore two complementary application areasfor synthetic face images: 1) Using fully annotated syntheticface images we can study the face recognition rate as afunction of interpretable parameters such as face pose. Thisenables us to systematically analyze the effect of differenttypes of dataset biases on the generalization ability of neu-ral network architectures. Our analysis reveals that deeperneural network architectures can generalize better to un-seen face poses. Furthermore, our study shows that currentneural network architectures cannot disentangle face poseand facial identity, which limits their generalization ability.2) We pre-train neural networks with large-scale syntheticdata that is highly variable in face pose and the number offacial identities. After a subsequent fine-tuning with real-world data, we observe that the damage of dataset bias inthe real-world data is largely reduced. Furthermore, wedemonstrate that the size of real-world datasets can be re-duced by 75% while maintaining competitive face recogni-tion performance. The data and software used in this workare publicly available.