Open AccessMutual reinforced part‐aligned network for person re‐identification
Author(s) -
Liu Yongwen,
An Gaoyun,
Zheng Zhenxing
Publication year - 2019
Publication title -
electronics letters
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.375
H-Index - 146
ISSN - 1350-911X
DOI - 10.1049/el.2019.0363
Subject(s) - computer science , identification (biology) , task (project management) , representation (politics) , stacking , artificial intelligence , pattern recognition (psychology) , computer vision , engineering , botany , physics , systems engineering , nuclear magnetic resonance , politics , political science , law , biology
Part misalignment of the human body caused by complex variations in viewpoint and pose makes person re‐identification be a fundamental challenging task. In this Letter, the authors propose a new Part‐Aligned Network (PAN) stacked by Part‐Aligned Module (PAM) to learn the effective representation for person re‐identification. Specifically, PAN is constructed by stacking several PAMs which is simultaneously comprised of an appearance branch and a part branch. The appearance branch performs spatial modelling while the part branch captures human part information simultaneously. Appearance features and part features are then combined to explore the implicit complementary advantages and these two branches are learned in a mutually reinforcing way. Extensive experiments are conducted on the three most popular datasets: Market‐1501, DukeMTMC‐reID, and CUHK03. The experimental results demonstrate that the proposed PAN achieves superior performance to the existing state‐of‐the‐art methods.