Semi-Supervised Learning With Heterogeneous Distribution Consistency for Visible Infrared Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Visible Infrared Person Re-Identification (VI-ReID) faces challenges due to modality gaps between visible and infrared cameras.
• Fully-supervised methods require extensive labeled data, which is often unavailable in real-world scenarios, especially at night.
• Limited known identities and significant modality discrepancies hinder model effectiveness.

Purpose of the Study:

• To propose a novel Semi-Supervised Learning (SSL) framework for VI-ReID.
• To address the challenges of limited labeled data and modality discrepancies in VI-ReID.
• To improve the performance of VI-ReID systems in real-world, low-label scenarios.

Main Methods:

• Introduced a Semi-Supervised Learning framework with Heterogeneous Distribution Consistency (HDC-SSL).
• Developed a Gaussian Mixture Model-based Pseudo Labeling (GMM-PL) method to adaptively label identities based on confidence distributions.
• Proposed Modality Consistency Regularization (MCR) to ensure cross-modality prediction consistency and handle modality variance for unlabeled data.

Main Results:

• Demonstrated the effectiveness of HDC-SSL through extensive experiments on two VI-ReID datasets with varying label settings.
• Achieved competitive performance compared to state-of-the-art fully-supervised VI-ReID methods.
• Showcased remarkable results on the RegDB dataset using only one visible and one infrared label per class.

Conclusions:

• The proposed HDC-SSL framework effectively addresses the limitations of traditional VI-ReID methods.
• The GMM-PL and MCR components significantly improve the utilization of unlabeled data and handle modality discrepancies.
• HDC-SSL offers a promising solution for practical VI-ReID applications with scarce labeled data.