End-to-End Detection and Re-identification Integrated Net for Person Search

Abstract

This paper proposes a pedestrian detection and re-identification (re-id) integrated net (I-Net) in an end-to-end learning framework. The I-Net is used in real-world video surveillance scenarios, where the target person needs to be searched in the whole scene videos, and the annotations of pedestrian bounding boxes are unavailable. Comparing to the successful OIM method [31] for joint detection and re-id, we have three distinct contributions. First, we implement a Siamese architecture instead of one stream for an end-to-end training strategy. Second, a novel on-line pairing loss (OLP) with a feature dictionary restricts the positive pairs. Third, hard example priority softmax loss (HEP) with little computation cast is proposed to deal with the online hard example mining. We show our results on CUHK-SYSU and PRW datasets. Our method narrows the gap between detection and re-identification, and achieves a superior performance.