Position-squeeze and excitation module for facial attribute analysis

In this paper, we focus on multiple facial attribute recognition in a single Convolutional Neural Network (CNN). We propose a Position-Squeeze and Excitation (PSE) module, which incorporates the spatial information of different attributes into CNN training. By adding a lateral branch which computes a weight mask for each attribute, the PSE module can help the network learn features from where attributes naturally appear. Moreover, the module can be added as a branch to any classical convolutional neural network to perform end-to-end multi-attribute classification. Experiments show that, our solution has achieved high accuracy on both the CelebA dataset and the LFWA dataset.