Learning to pool high-level features for face representation

The available face descriptors are always generated by a hand-designed pooling scheme or without a pooling process. We propose to learn a pooling scheme for high-level features. First, we obtain the local features on the densely sampled points on a face image. Then, a weighted-sum pooling is used to obtain the high-level feature of a block of this face image. By learning the pooling weights, the structure information of local features is integrated into the high-level feature of the block. At the same time, a linear transformation is learned to reduce the dimension of this high-level feature. Our main contribution is the method of learning the pooling scheme, which can capture the structure information between the local features in a block. This structure information includes the facial structures and contours. The experiments on multiple face datasets confirm the efficiency and effectiveness of our method.