Unsupervised Person Re-Identification With Stochastic Training Strategy.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised person re-identification (re-ID) is crucial for scalable real-world applications.
• Current methods rely on clustering for pseudo-labeling but suffer from unreliable centroids and inconsistent features.
• These limitations hinder performance by accumulating errors and introducing training bias.

Purpose of the Study:

• To propose a novel unsupervised re-ID approach using a stochastic learning strategy.
• To address the issues of unreliable cluster centroids and inconsistent instance features in existing methods.
• To improve the accuracy and robustness of unsupervised re-ID systems.

Main Methods:

• A stochastic updated memory mechanism is introduced, using random instances for cluster-level memory updates.
• A sole last-seen sample directly updates each cluster center for consistency.
• A unified distance matrix is proposed to mitigate camera domain bias during clustering.

Main Results:

• The proposed stochastic learning strategy effectively avoids training bias from unreliable pseudo-labels.
• The updated memory ensures consistency and up-to-date cluster centers for improved classification.
• The unified distance matrix reduces camera variance and emphasizes identity variations, outperforming state-of-the-art methods.

Conclusions:

• The stochastic learning strategy offers a significant advancement in unsupervised person re-ID.
• The method demonstrates superior performance across various unsupervised and domain adaptation re-ID tasks.
• This approach provides a more robust and accurate solution for real-world person re-identification challenges.