Stable Median Centre Clustering for Unsupervised Domain Adaptation Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) for person re-identification (re-ID) addresses domain shift by transferring knowledge from labeled source domains to unlabeled target domains.
• Pseudolabel-based UDA methods achieve state-of-the-art performance but are hindered by noisy labels generated during clustering, limiting model optimization.

Purpose of the Study:

• To propose a novel Stable Median Centre Clustering (SMCC) method to enhance unsupervised domain adaptation for person re-ID.
• To mitigate the impact of noisy labels and outliers in the clustering process for more robust model training.

Main Methods:

• Developed SMCC, a clustering algorithm that adaptively identifies credible samples for model optimization.
• Employed intracluster distance confidence and K-reciprocal nearest neighbor cluster proportion to select reliable samples.
• Assigned adaptive weights based on intracluster sample distance confidence to improve inter-cluster distance measurement and robustness.

Main Results:

• SMCC effectively mines credible and stable samples for training unsupervised domain adaptation models.
• The proposed method demonstrates improved performance in person re-ID tasks compared to existing approaches.
• Experimental results validate the robustness and effectiveness of SMCC in handling label noise.

Conclusions:

• SMCC offers a robust solution for unsupervised domain adaptation in person re-ID by improving sample selection and reducing noise.
• The method enhances the performance and stability of deep learning models in cross-domain person re-ID scenarios.
• This work contributes to advancing person re-ID techniques by providing a more reliable clustering approach.