ABSTRACT
Person re-identification (Re-ID) can determine whether a pedestrian target can be matched across diverse regions or cameras, thereby alleviating the problem between massive surveillance data and inefficient manual retrieval. Inspired by attribute-person recognition (APR) network, this paper proposes an improved Re-ID method based on attribute learning, which uses an attribute-guided attention mechanism module and an attribute dependency learning module to learn fine-grained attribute features and rich dependencies among them. After that, a joint model with the integration of attribute recognition and person identity recognition is built for end-to-end training. Experimental results show that the proposed method can effectively improve Re-ID accuracy and achieve a competitive recognition performance.
Acknowledgments
We thank Zheng et al. (2015), Ristani et al. (2016), and Lin et al. (2019) for providing person re-ID datasets and pedestrian attribute sets, and hyk1996 for providing the source code of attribute learning baseline model (https://github.com/hyk1996/Person-Attribute-Recognition-MarketDuke) on the GitHub website. They have significantly helped in preparing this paper.
Disclosure of potential conflicts of interest
No potential conflict of interest was reported by the author(s).