HAT: Homography-Based Alternate Training for Pose-Invariant Face Recognition

Over the past decade, deep learning has greatly improved the performance of face recognition. Unfortunately, pose-invariant face recognition still remains a challenge. The discriminative power of learned features can be degraded dramatically due to pose variations. Pose information is inevitably encapsulated by the network during training. This identity-irrelevant information has negative contribution to the recognition accuracy. In this paper, we propose a neat, lightweight and yet powerful method for learning pose-free features, namely Homography-based Alternate Training (HAT). Our method leverages homography for representing arbitrary poses, and optimizes the network in an alternate manner so that pose information cannot be captured in the network during training. Unlike previous methods, our method does not introduce any additional computation at inference stage, and can well handle a wide range of poses including yaw and pitch. Experimental results show that our method consistently outperforms the state-of-the-art methods in pose-invariant face recognition, especially under extreme poses. The code and trained model will be made publicly available.