Drone-captured vehicle re-identification via perspective mask segmentation and hard sample learning

Unlike fixed cameras based Re-Identification (Re-ID), drones have greater maneuverability, but they also bring more difficulties to Re-ID, such as more viewing angles and complex backgrounds. Aiming at the problem of intra-class difference and inter-class similarity caused by viewing angles, we propose the Multi View Mask Feature Based Hard Sample Learning (MVMF-HSL) method, which can combined weighted features from different perspectives. Considering complex perspectives with the prominent top-down features, the boundary mask information extracted from different perspectives may be error. First, a mask segmentation network is designed to segment the vehicle image into four different regions, front, back, side and roof, and the mask features of these four different viewing angle regions are extracted respectively. Then, a Common Visible Attention Module (CVA-Module) is proposed to assign weights to the four mask features, which assigns the feature weights of different perspectives. Finally, a metric learning method focusing on hard samples, called Hard Samples Learning (HSL), is proposed to train the vehicle Re-ID network. Because there is rare masked vehicle dataset from drone shooting, We label a new dataset with mask labels from the drone shooting view. Experimental results show that the utilization of features from different perspectives and the learning of difficult samples effectively improve the Re-ID performance.