Leveraging Intra and Inter-Dataset Variations for Robust Face Alignment

Face alignment is a critical topic in the computer vision community. Numerous efforts have been made and various benchmark datasets have been released in recent decades. However, two significant issues remain in recent datasets, e.g., Intra-Dataset Variation and Inter-Dataset Variation. Inter-Dataset Variation refers to bias on expression, head pose, etc. inside one certain dataset, while Intra-Dataset Variation refers to different bias across different datasets. In this study, we show that model robustness can be significantly improved by leveraging rich variations within and between different datasets. This is non-trivial because of inconsistent landmark definitions between different datasets and the serious data bias within one certain dataset. To address the mentioned problems, we proposed a novel Deep Variation Leveraging Network (DVLN), which consists of two strong coupling sub-networks, e.g., Dataset-Across Network (DA-Net) and Candidate-Decision Network (CD-Net). In particular, DA-Net takes advantage of different characteristics and distributions across different datasets, while CD-Net makes a final decision on candidate hypotheses given by DA-Net to leverage variations within one certain dataset. Extensive evaluations show that our approach demonstrates real-time performance and dramatically outperforms state-of-the-art methods on the challenging 300-W dataset.