Dually Distribution Pulling Network for Cross-Resolution Person Reidentification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (Re-ID) faces challenges with varying image resolutions (high-resolution/HR and low-resolution/LR) in real-world scenarios.
• Existing cross-resolution Re-ID methods primarily address resolution mismatch, often neglecting distribution discrepancies between HR and LR images.

Purpose of the Study:

• To propose a novel Dually Distribution Pulling Network (DDPN) to mitigate distribution mismatch in cross-resolution person Re-ID.
• To enhance person Re-ID performance by aligning distributions of LR and HR images from both image and feature perspectives.

Main Methods:

• The DDPN comprises a super-resolution module and a person Re-ID module.
• Maximum Mean Discrepancy (MMD) losses are optimized to pull LR image distributions closer to HR image distributions.
• The approach tackles distribution mismatch at both the image and feature levels.

Main Results:

• Experiments on three benchmark datasets validate the effectiveness of the DDPN.
• The proposed DDPN achieved a rank-1 accuracy of 76.9% on the VR-Market1501 dataset.
• DDPN demonstrated a significant performance improvement, outperforming state-of-the-art methods by 10%.

Conclusions:

• The DDPN effectively addresses the distribution mismatch problem in cross-resolution person Re-ID.
• The dual-module approach enhances Re-ID accuracy by aligning image and feature distributions.
• DDPN represents a significant advancement in cross-resolution person re-identification technology.