Multi-level attention model for person re-identification

Abstract

Person re-identification (re-ID) is an important research topic in computer vision. Due to its important applications in video surveillance, it has been receiving increasing attention. The key challenge of this task is how to capture appearance variations under different camera views. Current state-of-the-art methods employ deep CNN feature and part-level representations to generate robust representation for pedestrians. However, when human parts are not well located and aligned, discriminative information is difficult to be captured. To address this issue, we propose a feature attention block for person re-ID task. The proposed block learns part-level attention on different local regions, and the weighted part-level features are pooled into a global representation. The proposed attention block can be extended to multi-level situation and generates more robust representation. The proposed feature attention block can be seamlessly integrated into existing CNN structures (e.g., ResNet and DenseNet), and is trained only with identity loss. We conduct extensive experiments on three popular person re-ID benchmarks including Market-1501, DukeMTMC-reID, and CUHK03. The proposed framework achieves promising results compared with current state-of-the-arts.