TY - GEN
T1 - Triplet Multi-task Learning Strategy for Person Re-identification Using Deep Learning
AU - Bilakeri, Shavantrevva
AU - Karunakar, A. K.
N1 - Publisher Copyright:
© 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.
PY - 2023
Y1 - 2023
N2 - The majority of existing person re-identification methods are based on human part partitioning, semantic segmentation of the human body, or metric learning. However, they failed to recognize their joint benefits and mutual complementary effect. In this paper, we employ a Triplet Multi-task Learning (TML) strategy that combines three tasks simultaneously, including person re-identification, semantic segmentation, and triplet prediction. The usefulness of local and global features created by the Region Aligned Pooling (RAP) module is highlighted, as it makes the framework robust to posture variation, backdrop clutter, and occlusion. The part segmentation module is considered with the goal of handling spatial misalignment. In addition, the triplet prediction module is added to decrease the intraclass separability and increase the inter-class distance. Extensive tests show that our method outperforms prior-art techniques and consistently achieves excellent results on popular benchmark datasets such as CUHK03, Market-1501, and DukeMTMC-reID.
AB - The majority of existing person re-identification methods are based on human part partitioning, semantic segmentation of the human body, or metric learning. However, they failed to recognize their joint benefits and mutual complementary effect. In this paper, we employ a Triplet Multi-task Learning (TML) strategy that combines three tasks simultaneously, including person re-identification, semantic segmentation, and triplet prediction. The usefulness of local and global features created by the Region Aligned Pooling (RAP) module is highlighted, as it makes the framework robust to posture variation, backdrop clutter, and occlusion. The part segmentation module is considered with the goal of handling spatial misalignment. In addition, the triplet prediction module is added to decrease the intraclass separability and increase the inter-class distance. Extensive tests show that our method outperforms prior-art techniques and consistently achieves excellent results on popular benchmark datasets such as CUHK03, Market-1501, and DukeMTMC-reID.
UR - https://www.scopus.com/pages/publications/85149931429
UR - https://www.scopus.com/pages/publications/85149931429#tab=citedBy
U2 - 10.1007/978-981-19-6634-7_31
DO - 10.1007/978-981-19-6634-7_31
M3 - Conference contribution
AN - SCOPUS:85149931429
SN - 9789811966330
T3 - Lecture Notes in Networks and Systems
SP - 447
EP - 461
BT - Proceedings of International Conference on Data Science and Applications - ICDSA 2022
A2 - Saraswat, Mukesh
A2 - Chowdhury, Chandreyee
A2 - Kumar Mandal, Chintan
A2 - Gandomi, Amir H.
PB - Springer Science and Business Media Deutschland GmbH
T2 - 3rd International Conference on Data Science and Applications, ICDSA 2022
Y2 - 26 March 2022 through 27 March 2022
ER -