Identity-Guided Collaborative Learning for Cloth-Changing Person Reidentification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Pattern Recognition

Background:

• Cloth-changing person reidentification (ReID) faces challenges due to significant appearance variations from clothing changes and pose variations.
• Existing ReID methods struggle with dynamic appearance changes and underutilize human semantic and stable identity information.

Purpose of the Study:

• To propose a novel identity-guided collaborative learning scheme (IGCL) for robust cloth-changing person ReID.
• To effectively leverage human semantic information and invariant identity features to overcome appearance variations.

Main Methods:

• Developed a clothing attention degradation stream to minimize clothing interference using clothing attention and mid-level collaborative learning.
• Introduced a human semantic attention and body jigsaw stream to emphasize background-invariant semantic information and simulate pose variations.
• Proposed a pedestrian identity enhancement stream to bolster identity importance and extract robust identity features.
• Integrated these streams into an end-to-end framework guided by identity optimization.

Main Results:

• The proposed IGCL method demonstrated superior performance across six public cloth-changing person ReID datasets.
• Achieved state-of-the-art results, outperforming existing methods on multiple benchmarks.
• Extracted features exhibit enhanced representation and discrimination capabilities, with reduced correlation to clothing.

Conclusions:

• The IGCL scheme effectively addresses the challenges of cloth-changing person ReID by integrating semantic and identity information.
• The proposed approach offers a more robust and discriminative feature representation for person reidentification tasks.
• This work advances the field of person ReID by providing a novel framework for handling significant appearance variations.