Graph-based Attribute-aware Unsupervised Person Re-identification with Contrastive learning
Ge Cao, Qing Tang, Kanghyun Jo
Aug 17, 2022
Citations
0
Citations
Journal
2022 International Workshop on Intelligent Systems (IWIS)
Full text
Semantic Scholar
Key Takeaway Key Takeaway: The proposed Attribute Enhancement Contrastive Learning (AECL) framework effectively performs unsupervised person re-identification on large-scale datasets, outperforming the baseline and state-of-the-art methods.
Abstract
This paper is employed on the unsupervised per-son re-identification (Re-ID) task which does not leverage any annotation provided by the target dataset and other datasets. Previous works have investigated the effectiveness of applying self-supervised contrastive learning, which adopts the cluster-based method to generate the pseudo label and split each cluster into multiple proxies by camera ID. This paper applies the Attribute Enhancement Module (AEM), which utilizes Graph Convolutional Network to integrate the correlations between attributes, human body parts features, and the extracted dis-criminative feature. And the experiments are implemented to demonstrate the great performance of the proposed Attribute Enhancement Contrastive Learning (AECL) in camera-agnostic version and camera-aware version on two large-scale datasets, including Market-1501 and DukeMTMC-ReID. Compared with the baseline and the state-of-the-art, the proposed framework achieves competitive results.