Mixed Attention-Aware Network for Person Re-identification

In this paper, we propose a Mixed Attention-Aware Network (MAAN), which consists of a Partial Hard Attention (PHA) and an Attention-aware Feature Fusion Network (AFFN). PHA applies hard attention to the local feature map to eliminate irrelevant background and extract more fine-grained human body features under the guidance of pose estimation. AFFN first applies soft attention to the global feature map, and then combines the local and global features with different attention-aware, and finally forms a mixed attention-aware feature to solve the pedestrian pose variations and severe occlusion problems. We perform two experiments on two large open source benchmarks, including Market-1501, CUHK03-NP. These verify our method achieve advanced result.