Heterogeneous Pseudo-Supervised Learning for Few-shot Person Re-Identification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (Re-ID) faces challenges with limited labeled data (few-shot learning).
• Pseudo-Supervised Learning (PSL) is a key approach, but its effectiveness hinges on pseudo-label quality.

Purpose of the Study:

• To enhance pseudo-label quality for improved few-shot person Re-ID performance.
• To introduce a novel Heterogeneous Pseudo-Supervised Learning (H-PSL) framework.

Main Methods:

• Developed an H-PSL framework utilizing an isomer as a feature extractor trained on pseudo-labeled data.
• Implemented a cross-level asynchronous match mechanism between the model and pseudo-supervised data.
• Designed a knowledge fusion strategy to integrate pseudo-labels and their confidence, filtering noisy data.

Main Results:

• The H-PSL framework significantly improved pseudo-label quality and feature expression.
• The knowledge fusion strategy effectively removed noisy pseudo-labeled samples.
• The proposed method demonstrated performance improvements across four benchmark datasets.

Conclusions:

• The H-PSL framework offers a simple yet effective solution for few-shot person Re-ID.
• Enhancing pseudo-label quality is critical for advancing PSL in low-data regimes.
• The method shows broad applicability and improves upon existing baseline works.