TY - GEN
T1 - Multiple Granularities with Gradual Transition Network for Person Re-identification
AU - Lu, Jialin
AU - Zhao, Qingjie
AU - Wang, Lei
N1 - Publisher Copyright:
© 2022, Springer Nature Singapore Pte Ltd.
PY - 2022
Y1 - 2022
N2 - Person re-identification (Re-ID) is a challenging task in computer vision, which aims at retrieving a target pedestrian from a gallery of person images captured from various cameras. Recent part-based methods, which employ horizontal splitting to integrate global and local information as final person representation, are not efficient enough in cases where the discriminative information near the splitting boundary is missing or incomplete due to partition. To address this issue, we proposed a novel method called Multiple Granularities with Gradual Transition Network (MGGTN) to fully mine fine-grained features at each part level and make the person representation more discriminative and robust. Our model introduces multi-branch network architecture to extract features with multiple granularities and uses a gradual transition strategy to obtain partial regions instead of easily partitioning the feature map into several stripes. Experimental results demonstrate the effectiveness of our method for Re-ID task. Especially, we achieve the new state-of-the-art results on both DukeMTMC-ReID and CUHK03 datasets and obtain the top rank1 result on Market1501 dataset.
AB - Person re-identification (Re-ID) is a challenging task in computer vision, which aims at retrieving a target pedestrian from a gallery of person images captured from various cameras. Recent part-based methods, which employ horizontal splitting to integrate global and local information as final person representation, are not efficient enough in cases where the discriminative information near the splitting boundary is missing or incomplete due to partition. To address this issue, we proposed a novel method called Multiple Granularities with Gradual Transition Network (MGGTN) to fully mine fine-grained features at each part level and make the person representation more discriminative and robust. Our model introduces multi-branch network architecture to extract features with multiple granularities and uses a gradual transition strategy to obtain partial regions instead of easily partitioning the feature map into several stripes. Experimental results demonstrate the effectiveness of our method for Re-ID task. Especially, we achieve the new state-of-the-art results on both DukeMTMC-ReID and CUHK03 datasets and obtain the top rank1 result on Market1501 dataset.
KW - Different granularities feature learning
KW - Multi-branch network
KW - Person re-identification
UR - http://www.scopus.com/inward/record.url?scp=85123584937&partnerID=8YFLogxK
U2 - 10.1007/978-981-16-9247-5_26
DO - 10.1007/978-981-16-9247-5_26
M3 - Conference contribution
AN - SCOPUS:85123584937
SN - 9789811692468
T3 - Communications in Computer and Information Science
SP - 328
EP - 342
BT - Cognitive Systems and Information Processing - 6th International Conference, ICCSIP 2021, Revised Selected Papers
A2 - Sun, Fuchun
A2 - Hu, Dewen
A2 - Wermter, Stefan
A2 - Yang, Lei
A2 - Liu, Huaping
A2 - Fang, Bin
PB - Springer Science and Business Media Deutschland GmbH
T2 - 6th International Conference on Cognitive Systems and Signal Processing, ICCSIP 2021
Y2 - 20 November 2021 through 21 November 2021
ER -