Fusion-Attention Network for person search with free-form natural language

In the task of searching persons from surveillance videos or large scale image dataset, it is more challenging to utilize free-form natural language to retrieve persons than using images and attributes. Thus, to deal with the challenges brought from the complexity of free-from natural language and visual-description mapping, we propose to strengthen the role of textual descriptions by means of fusion and attention mechanisms to make the discriminative words visually sensitive. Specifically, we develop an end-to-end fusion-attention structure, called Description-Strengthened Fusion-Attention Network (DSFA-Net) to tackle the challenging task. Specifically, DSFA-Net has a fusion sub-network and an attention sub-network, where three attention mechanisms are applied. Extensive experiments are performed on the large-scale CUHK-PEDES, which demonstrate the superiority of DSFA-Net. (C) 2018 Elsevier B.V. All rights reserved.