Composite Augmentation and Feature Integration Reconstruction for Occluded Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing occluded person re-identification (Re-ID) methods often struggle with diverse occlusion complexities.
• Current occlusion datasets exhibit significant training and testing data imbalances, hindering model generalization.

Purpose of the Study:

• To propose a novel method for occluded person Re-ID that addresses data imbalance and improves feature representation.
• To enhance the accuracy and robustness of person matching in challenging occlusion scenarios.

Main Methods:

• A Composite Data Augmentation Module was developed to increase occluded samples and mitigate data imbalance.
• A feature interaction module was introduced for bidirectional global and local feature interaction, reducing occlusion noise and redundancy.
• A Feature Reconstruction Module was implemented to reconstruct occluded body parts using nearest neighbor images.

Main Results:

• The proposed method demonstrates superior performance on five challenging occlusion datasets.
• The approach effectively alleviates data imbalance and improves feature representation efficiency.
• Experimental results confirm the method's capability in handling complex occlusion scenarios.

Conclusions:

• The developed method significantly advances occluded person Re-ID by addressing key limitations of existing approaches.
• The combination of data augmentation, feature interaction, and reconstruction leads to more complete and accurate person matching.
• This work provides a robust solution for person re-identification in real-world scenarios with occlusions.