Video Person Re-Identification Using Attribute-Enhanced Features

In this work we propose to boost video-based person re-identification (Re-ID) by using attribute-enhanced feature presentation. To end, not only try use the ID-relevant attributes more effectively, but also for first time in literature harness ID-irrelevant help model training. The former mainly include gender, age, clothing characteristics, etc., which contain rich and supplementary information about pedestrian; latter viewpoint, action, are seldom used identification previously. particular, enhance significant areas of image with a novel Attribute Salient Region Enhance (ASRE) module that can attend accurately body pedestrian, so as better separate target from background. Furthermore, find many subject-relevant factors, like view angle movement have great impact on two-dimensional appearance pedestrian. We then exploit both via triplet loss called Viewpoint Action-Invariant (VAI) loss. Based above, design an Salience Assisted Network (ASA-Net) perform attribute recognition along identity recognition, enhancement hard sample mining. Extensive experiments MARS DukeMTMC-VideoReID datasets show our method outperforms state-of-the-arts. Also, visualizations learning results further prove effectiveness proposed method.