Context-Aware Network for Person Search

The key to effective person search is aiming to localize the pedestrians and obtain the discriminative embeddings representation for person ReID from numerous surveillance scene images. And the existing one-step anchor-free methods can achieve a trade-off between speed and accuracy, but it can not fully exploit the contextual feature information of search context, resulting in undesirable localization. To alleviate this issue, we propose a Context-Aware Network for Person Search (CANPS) to delve into the high-level contextual information. In CANPS, firstly, context encoder is proposed to bridge the gap between the feature maps, achieved by distributing rich contextual information to prediction head layers. Second, we design the malleable center sampling strategy to reasonably expose sample region and focus on the centroid feature representations. What’s more, we design above components in a trainable bag-of-freebies manner, so that real-time person search can greatly improve the accuracy without increasing extra inference cost. Extensive experiments show that the approach we proposed can outperform current state-of-the-art methods in public CUHK-SYSU datasets.