TY - GEN
T1 - Person re-identification with neural architecture search
AU - Zhang, Shizhou
AU - Cao, Rui
AU - Wei, Xing
AU - Wang, Peng
AU - Zhang, Yanning
N1 - Publisher Copyright:
© Springer Nature Switzerland AG 2019.
PY - 2019
Y1 - 2019
N2 - Most of the existing person re-identification (ReID) methods use a classification network pre-trained on external data as the backbone and then fine-tune it, which results in a network architecture that is fixed and dependent on pre-training of external data. There are also some methods that are specifically designed by human experts for ReID, but manual network design becomes more difficult as network requirements increase and often fails to achieve optimal settings. In this paper, we consider using emerging neural architecture search (NAS) technology as a tool to solve above problems. However, most of NAS methods deal with classification tasks, which causes NAS to not be directly extended to ReID. In order to coordinate the inconsistency between the two optimization goals, we propose to establish an objective function with the assistant of the triplet loss to guide the direction of architecture search. Finally, it is no longer dependent on external data to automatically generate a ReID network with excellent performance using NAS directly on the target dataset. The experimental results on three public datasets validate that our method can automatically and efficiently find the network architecture suitable for ReID.
AB - Most of the existing person re-identification (ReID) methods use a classification network pre-trained on external data as the backbone and then fine-tune it, which results in a network architecture that is fixed and dependent on pre-training of external data. There are also some methods that are specifically designed by human experts for ReID, but manual network design becomes more difficult as network requirements increase and often fails to achieve optimal settings. In this paper, we consider using emerging neural architecture search (NAS) technology as a tool to solve above problems. However, most of NAS methods deal with classification tasks, which causes NAS to not be directly extended to ReID. In order to coordinate the inconsistency between the two optimization goals, we propose to establish an objective function with the assistant of the triplet loss to guide the direction of architecture search. Finally, it is no longer dependent on external data to automatically generate a ReID network with excellent performance using NAS directly on the target dataset. The experimental results on three public datasets validate that our method can automatically and efficiently find the network architecture suitable for ReID.
KW - Neural architecture search
KW - Person re-identification
UR - http://www.scopus.com/inward/record.url?scp=85086144295&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-31654-9_46
DO - 10.1007/978-3-030-31654-9_46
M3 - 会议稿件
AN - SCOPUS:85086144295
SN - 9783030316532
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 540
EP - 551
BT - Pattern Recognition and Computer Vision- 2nd Chinese Conference, PRCV 2019, Proceedings, Part I
A2 - Lin, Zhouchen
A2 - Wang, Liang
A2 - Tan, Tieniu
A2 - Yang, Jian
A2 - Shi, Guangming
A2 - Zheng, Nanning
A2 - Chen, Xilin
A2 - Zhang, Yanning
PB - Springer
T2 - 2nd Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2019
Y2 - 8 November 2019 through 11 November 2019
ER -