Dbtcc: Unsupervised Pedestrian Re-Identification Combined Double-Branch Transformer and Clustering Contrastive Learning

26 Pages Posted: 17 Aug 2023

See all articles by Xuan Wang

Xuan Wang

Yantai University

Zhaojie Sun

Yantai University

Abdellah Chehri

Royal Military College of Canada

Yongchao Song

Yantai University

Abstract

Detecting pedestrians plays a crucial role in various computer vision applications such as surveillance, public security, intelligent transportation systems, improved safety for vulnerable road users, and driving assistance systems. Unsupervised pedestrian re-identification (re-ID) enables the full utilization of existing data resources by feature extraction and clustering pedestrian images. It promotes the development of sustainable computing system design. However, in unsupervised re-ID approaches, the progress of clustering updates is typically inconsistent because of the limits of convolutional neural networks in obtaining fine-grained cues from features. To address this issue, we proposed the double-branch transformer and clustering contrastive learning (DBTCC) method for unsupervised re-ID tasks. Specifically, we designed a double-branch structure (DBST) based on an improved vision transformer to obtain multi-granularity local and global features by reshaping and averaging operations on the extracted feature. Furthermore, we constructed a cluster-level memory dictionary so that a unique cluster representation describes each cluster. In addition, we applied cluster contrastive learning for updating features, which addresses the consistency issue in cluster updating. The multi-granularity features retrieved by DBST improved the clustering outcomes, and the consistency of clustering can also be maintained effectively throughout the pipeline. Extensive experiments have demonstrated that DBTCC shows significant improvements compared to pure unsupervised methods on the Market-1501, DukeMTMC-reID, and MSMT17 datasets.

Keywords: Pedestrian Re-identification, Sustainable Computing System, Vision Transformer, Double-Branch Structure, Contrastive learning

Suggested Citation

Wang, Xuan and Sun, Zhaojie and Chehri, Abdellah and Song, Yongchao, Dbtcc: Unsupervised Pedestrian Re-Identification Combined Double-Branch Transformer and Clustering Contrastive Learning. Available at SSRN: https://ssrn.com/abstract=4544150 or http://dx.doi.org/10.2139/ssrn.4544150

Xuan Wang

Yantai University ( email )

32, Qingquan RD
Laishan District
Yantai, 264005
China

Zhaojie Sun

Yantai University ( email )

32, Qingquan RD
Laishan District
Yantai, 264005
China

Abdellah Chehri (Contact Author)

Royal Military College of Canada ( email )

Kingston
Canada

Yongchao Song

Yantai University ( email )

32, Qingquan RD
Laishan District
Yantai, 264005
China

Do you have negative results from your research you’d like to share?

Paper statistics

Downloads
35
Abstract Views
113
PlumX Metrics