Multi-pseudo Regularized Label for Generated Data in Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep learning models require extensive labeled data, which is costly and time-consuming to acquire.
• Generative Adversarial Networks (GANs) can create synthetic data, but these samples often lack crucial annotation labels.
• Limited real-world data restricts the performance of deep neural networks in various applications.

Purpose of the Study:

• To address the challenge of limited labeled data in deep learning by proposing a novel labeling strategy for synthetic data.
• To enhance the training of deep neural networks using generated data in a semi-supervised learning framework.
• To improve the accuracy and effectiveness of person re-identification (re-ID) systems.

Main Methods:

• Introduction of Multi-pseudo Regularized Label (MpRL), a virtual label assigned to generated data.
• MpRL assigns a set of weight-based values (multi-pseudo labels) to each generated sample, indicating its affiliation likelihood to predefined classes.
• Integration of MpRL-labeled generated data with real training data for semi-supervised learning using state-of-the-art Convolutional Neural Networks (CNNs).

Main Results:

• The proposed MpRL method significantly improved person re-ID performance across five benchmark datasets (Market-1501, DukeMTMC-reID, CUHK03, VIPeR, and CUHK01).
• Rank-1 accuracy improvements ranged from +3.48% to +6.30% over a strong CNN baseline.
• The method demonstrated superior performance compared to existing state-of-the-art approaches in person re-identification.

Conclusions:

• MpRL effectively leverages generated data to supplement limited real data, boosting deep network training.
• The proposed virtual labeling strategy enhances semi-supervised learning for computer vision tasks, particularly person re-ID.
• This approach offers a viable solution for improving model performance when faced with data scarcity.