.png){kind=icon}
.png){kind=icon}
Return
MPCCL: Multiview predictive coding with contrastive learning for person re-identification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2022.108710.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In this paper, we investigate a new representation learning approach, termed as Multiview Predictive Coding with Contrastive Learning (MPCCL), for person re-identification (re-ID). Different from the conventional re-ID approaches that focus on learning representations from semantic label, our approach learns the identification of invariant information via representation reconstruction, which explores more finegrained semantic information in representation space. Specifically, given a chosen identity, the learned representation of its single view can be reconstructed by those of other views. Therefore, kernel density estimation (KDE) is firstly introduced for the adaptive reconstruction of the representation. Then, contrastive learning is adopted to increase the distance between the representations of the same views with different identities. Finally, representation reconstruction and contrastive learning jointly supervise the representation learning process, thus obtaining fine-grained semantic information and appearance-free representations. Extensive experiments on several re-ID datasets demonstrate that the proposed approach yields state-of-the-art results. (c) 2022 Elsevier Ltd. All rights reserved.
Keywords:
Person re-identification
Kernel density estimation
Representation construction
Contrastive learning
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W