HDNet: Human-like discrimination with visual key for few-shot cross-domain object detection

Mao Zhen Liu, Xiaoguang Di, Wenzhuang Wang

Abstract: Highlights•We propose a novel method, HDNet, for object detection tasks.•We explored the commonalities of the same object across multiple forms and domains.•We improved weak classification ability of the detector for new classes.•We introduce instance-level attention to eliminate background noise bias.•We redesigned an ARM to determine coefficient of multiple loss functions.