Hierarchical Attention Networks

We propose a novel attention network, which accurately attends to target objects of various scales and shapes in images through multiple stages. The proposed network enables multiple layers to estimate attention in a convolutional neural network (CNN). The hierarchical attention model gradually suppresses irrelevant regions in an input image using a progressive attentive process over multiple CNN layers. The attentive process in each layer determines whether to pass or suppress feature maps for use in the next convolution. We employ local contexts to estimate attention probability at each location since it is difficult to infer accurate attention by observing a feature vector from a single location only. The experiments on synthetic and real datasets show that the proposed attention network outperforms traditional attention methods in various attribute prediction tasks.