AAGNet: Attribute-Aware Graph-Based Network for Occluded Pedestrian Re-Identification

In large consumer sites, pedestrian re-identification (Re-ID) has the potential to enhance identify loyal consumers and create a more enjoyable shopping experience. Current Re-ID models always rely on some certain pedestrian feature descriptors, including body parts matching and pose key points, to extract part-level features. However, occlusion always causes a tremendous amount of noise and affects the feature representation, thereby significantly degrading the performance of those models. To address this problem, we propose an attribute-aware graph-based network (AAGNet) for Occluded Re-ID. Specifically, we develop a part-attribute feature extractor that maps the manually labeled pedestrian features into word vectors, and combines them with specific body part to obtain both attribute features and part features. The weight information of body parts and attributes are learned through graph convolution networks. Moreover, we introduce an occluded Re-ID dataset called Occluded-Market that can support the subsequent studies of occluded Re-ID. Comparative experimental results evidently demonstrate that the AAGNet shows superior performance in terms of accuracy, efficiency, and robustness on two open-source data sets. Our study can provide data and methodological support for further research on the occluded Re-ID and technological baseline for Re-ID-based commercial analytic applications in large consumer sites. The dataset is available at: github.com/Occluded_Market.