Efficient person search via learning-to-normalize deep representation

Ning Lv, Xuezhi Xiang, Xinyao Wang, Jie Yang, Rokia Abdein

Abstract: Highlights•An effective one-step method with learning-to-normalize deep representations.•Focal loss is adopted to focus on hard samples that are not easily handled by re-ID.•Gradient centralization is introduced to improve the generalization performance.