Learning Discriminative Representation with Attention and Diversity for Large-Scale Face Recognition

Recent improvements of face recognition mainly come from investigations on metric or loss functions rather than the structure of CNNs and interpretation of face representation. By analyzing the application of face representation (e.g. age estimation), we find some principles (utilizing low-level features and preserving enough localization information) of pixel-level prediction approaches would help to enhance the discriminative ability of face representation. Therefore, we propose a new plug-and-play attention module to integrate low-level features and propose two types of diversity regularizers to maintain localization information while reduce redundant correlation. Moreover, our diversity regularizers can achieve decorrelation without eigenvalue decomposition or the approximation process. Visualization results illustrate that models with our attention module and diversity regularizers capture more critical localization information. And competitive performance on large-scale face recognition benchmark verifies the effectiveness of our approaches.