Attentional alignment network

Face alignment has recently generated great popularity in computer vision due to its widespread applications. The cascaded regression model has dominated and achieved great progress in the last decade, which however suffers from innate shortcomings, e.g., reliance on initialization. In this work, we propose attentional alignment networks (AAN), a novel end-to-end convolutional architecture for direct face alignment without relying on cascaded regression. AAN incorporates the attention mechanism into a convolutional regression network, which generates multiple attention maps for different convolutional layers to capture distinctive features in different granularity; by introducing intermediate supervision to create top-down attention maps, AAN attends to regions around facial landmarks, which enables it to establish more informative and discriminative representation closely related to facial landmarks. Extensive experiments on four commonly-used benchmark datasets demonstrate that the proposed AAN consistently delivers high performance on all datasets, surpassing previous methods by large margins, which shows its great effectiveness for direct face alignment.