Deep Domain Adaptation for Asian Face Recognition via Ada-IBN

Owing to the great power of convolution neural networks (CNNs) in extracting discriminative features, remarkable progress has been made in face recognition recently. However, there remains a significant gap between the performance of real-world scenario and ideal experimental condition. Due to the lack of training data in Asian face, a lot of well-known datasets, such as MS-Celeb-1M, VGG Face and CASIA-Web Face that mainly consist of Caucasians, are applied to verify in Asians directly, which results in poor performance. To address this issue from the perspective of domain adaptation, in this paper, we propose a conceptually simple but novel method named Adaptive Instance Batch Normalization (Ada-IBN). We first develop the method that transferring the knowledge from source to target domain to alleviate the discrepancy while increases the generalization capacity of CNNs. Then we design a dataset Cceleb-1K which captures identification photos and live pictures from one thousand different Asians. Most importantly, extensive experiments and analysis on our Cceleb-1k and LFW show that our Ada-IBN network achieves state-of-the-art performance in face verification task.