Attention-Aware Compositional Network for Person Re-identification

  • 2018-05-09 01:43:10
  • Jing Xu, Rui Zhao, Feng Zhu, Huaming Wang, Wanli Ouyang
  • 20


Person re-identification (ReID) is to identify pedestrians observed fromdifferent camera views based on visual appearance. It is a challenging task dueto large pose variations, complex background clutters and severe occlusions.Recently, human pose estimation by predicting joint locations was largelyimproved in accuracy. It is reasonable to use pose estimation results forhandling pose variations and background clutters, and such attempts haveobtained great improvement in ReID performance. However, we argue that the poseinformation was not well utilized and hasn't yet been fully exploited forperson ReID. In this work, we introduce a novel framework called Attention-AwareCompositional Network (AACN) for person ReID. AACN consists of two maincomponents: Pose-guided Part Attention (PPA) and Attention-aware FeatureComposition (AFC). PPA is learned and applied to mask out undesirablebackground features in pedestrian feature maps. Furthermore, pose-guidedvisibility scores are estimated for body parts to deal with part occlusion inthe proposed AFC module. Extensive experiments with ablation analysis show theeffectiveness of our method, and state-of-the-art results are achieved onseveral public datasets, including Market-1501, CUHK03, CUHK01, SenseReID,CUHK03-NP and DukeMTMC-reID.


Introduction (beta)



Conclusion (beta)