End-to-End Person Search Based on Content Awareness

Within the domain of computer vision, person search stands out as a crucial and intricate undertaking. With the growth of urban population and the popularity of surveillance cameras, person search demonstrates its versatility in a multitude of applications, particularly within the realm of video surveillance, public safety and intelligent transportation. The person search task is comprised of two subtasks, namely person detection and person re-identification (ReID),the task of person detection concentrates on identifying similarities among diverse identities. In contrast, person ReID task centres on discovering dissimilarities among distinct identities. In order to reduce the influence of variations in tasks during the model training phase, we segment the ResNet50 network according to the subtasks and propose an end-to-end content-aware person search framework. Where the detection branch is built on top of shallow features and the person ReID branch is built on top of deep features. Meanwhile, we introduce gated deformable convolution module(GDCM)to sense the location of the person and further distinguish the background from the person. Finally, the effectiveness of the method is demonstrated on two authoritative datasets(CUHK-SYSU、PRW) .