Threshold-Based Hierarchical Clustering for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Unsupervised domain adaptation presents significant challenges in person re-identification (re-ID).
• Existing cluster-based re-ID methods often underutilize outlier information during clustering and fail to leverage all source and target data during training.

Purpose of the Study:

• To address limitations in existing person re-ID methods for unsupervised domain adaptation.
• To propose a novel Threshold-based Hierarchical clustering with Contrastive loss (THC) method.

Main Methods:

• THC treats outliers as individual clusters, preserving their information without predefining cluster numbers.
• The method employs contrastive loss to integrate information from source-class centroids, target-cluster centroids, and single-sample clusters.
• Extensive experiments were conducted on Market-1501, DukeMTMC-reID, and MSMT17 datasets.

Main Results:

• THC effectively preserves outlier information by treating them as single-sample clusters.
• The contrastive loss strategy maximizes the utility of all available data, enhancing model performance.
• The proposed method achieved state-of-the-art results on benchmark datasets.

Conclusions:

• The Threshold-based Hierarchical clustering with Contrastive loss (THC) method offers a significant advancement in unsupervised domain adaptation for person re-ID.
• THC's novel approach to handling outliers and utilizing comprehensive data information leads to superior performance compared to existing methods.