Aligned and Detail Guided Retrieval: Multi-scale Fine-Grained Features Enhancement for End-to-End Person Search

Accurate bounding boxes and discriminative feature representations are essential for Person Search. Although existing methods enhance small-scale person detection via feature pyramid networks, they still face two critical issues. Firstly, the continuous feature rescaling operations in the feature pyramid networks may result in misalignment between the bounding box and feature semantics. Secondly, insufficient learning of fine-grained features leads to limited expressiveness in person representation. Therefore, we propose the Align and Detail Guided Person Search Network (ADGPS), which extracts fine-grained features of the person while ensuring feature alignment. Specifically, it comprises two models. We designed the Attention Flow Alignment Model (AFAM) to alleviate the misalignment issue caused by the aggregation process of different scales. It aligns features by introducing semantic flow between features at different scales. We present the Person Fine-grained Feature Extraction Model (PFFE) to enhance detail learning via feature aggregation and expansion across dimensions to extract the people fine-grained features. In addition, we propose the novel ArcOIM loss, which introduces additional angular margins to guide the network in enhancing the distance between inter-class person features. Extensive experiments on two datasets demonstrate that our method outperforms state-of-the-art approaches on the CUHK-SYSU and PRW with 96.21% and 58.73% in mAP, respectively.