Discovering Underlying Person Structure Pattern with Relative Local Distance for Person Re-identification

원문 초록

Modeling the underlying person structure for person re-identification (re-ID) is difficult due to diverse deformable poses, changeable camera views and imperfect person detectors. How to exploit underlying person structure information without extra annotations to improve the performance of person re-ID remains largely unexplored. To address this problem, we propose a novel Relative Local Distance (RLD) method that integrates a relative local distance constraint into convolutional neural networks (CNNs) in an end-to-end way. It is the first time that the relative local constraint is proposed to guide the global feature representation learning. Specially, a relative local distance matrix is computed by using feature maps and then regarded as a regularizer to guide CNNs to learn a structure-aware feature representation. With the discovered underlying person structure, the RLD method builds a bridge between the global and local feature representation and thus improves the capacity of feature representation for person re-ID. Furthermore, RLD also significantly accelerates deep network training compared with conventional methods. The experimental results show the effectiveness of RLD on the CUHK03, Market-1501, and DukeMTMC-reID datasets. Code is available at \url{https://github.com/Wanggcong/RLD_codes}.