Complementary Data Augmentation for Cloth-Changing Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Person re-identification (Re-ID) is challenging due to appearance variations, especially clothing changes.
• Collecting diverse datasets with varied clothing for Re-ID is data-intensive and impractical.
• Existing methods struggle with cloth-invariant feature learning.

Purpose of the Study:

• To develop effective data augmentation strategies for cloth-changing person re-identification.
• To reduce the need for extensive data collection in Re-ID tasks.
• To enhance the robustness and accuracy of person Re-ID systems.

Main Methods:

• Designed complementary positive and negative data augmentation strategies.
• Positive augmentation: Randomly patches person images with different clothes to simulate diverse appearances.
• Negative augmentation: Generates out-of-distribution synthetic samples by combining appearance and posture factors.

Main Results:

• Proposed augmentation strategies significantly improve feature learning for Re-ID.
• Demonstrated superior performance on both cloth-changing and cloth-unchanging Re-ID tasks.
• Consistently achieved higher accuracy compared to various baseline methods.

Conclusions:

• The developed data augmentation techniques effectively address the challenges of cloth-changing person re-identification.
• These methods enhance model robustness and reduce reliance on large, varied datasets.
• The approach offers a practical solution for improving Re-ID system performance in real-world scenarios.