Multilevel Collaborative Attention Network for Person Search

Abstract

Person search aims to apply pedestrian detection and person re-identification simultaneously to search persons in images, which inevitably introduces pedestrian box misalignment during the procedure. And the detected boxes usually have a large variety of scales on a single image. Together with cluttered background and occlusion, all these distracting factors make it difficult to extract discriminative pedestrian representations. However, these problems are usually ignored by current person search systems. In this work, we propose a novel Multilevel Collaborative Attention Network (MCAN) to fulfill person search task efficiently. A multilevel selective learning is introduced to extract scale-aware features in different levels, and a collaborative attention module consisting of hard regional attention and soft pixel-wise attention is designed to deal with misalignment, background noise and occlusion. MCAN achieves 60.1% top-1 accuracy and 29.1% mAP on PRW benchmark, demonstrating its superiority over current state-of-the-art methods.