Open AccessAutomatic Attribute Learning for Person Re-Identification
Author(s) -
Yang Li,
Huahu Xu,
Minjie Bian,
Junsheng Xiao
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1576/1/012007
Subject(s) - computer science , artificial intelligence , identification (biology) , convolutional neural network , similarity (geometry) , pattern recognition (psychology) , scalability , redundancy (engineering) , machine learning , representation (politics) , annotation , image (mathematics) , botany , database , politics , political science , law , biology , operating system
Attribute based person re-identification is more robust to image appearance changes than low-level visual feature based methods. However, manual person attribute annotation has low efficiency, poor scalability and poor discrimination. Therefore, a scalable automatic attribute learning method for person re-identification is proposed. Firstly, the mapping matrix of person attributes is designed and generated automatically according to the principles of discernibility, sharing and low redundancy. Then the visual features are extracted using convolutional neural network and attribute classifiers are trained to detect person attributes, and the attribute based representation of person is generated. Finally, the person re-identification is carried out by comparing the similarity of the attribute based representations. Extensive experiments have been carried out on two common datasets Market-1501 and Duke MTMC-reID with the comparison to the Sate-of-the-Arts methods, the method achieved the best performance, which proves the superiority and effectiveness of the method.