Multi-Aligned and Multi-Scale Augmentation for Occluded Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Occluded person re-identification (Re-ID) is challenging due to occlusion noise and limited realistic data.
• Existing data augmentation methods suffer from intra-sample (misaligned occluders) and inter-sample (information loss) inconsistencies.
• These inconsistencies lead to unrealistic artifacts and weaken model robustness in Re-ID.

Purpose of the Study:

• To propose a unified Multi-Aligned and Multi-Scale Augmentation (MA-MSA) framework to address dual inconsistencies in occluded Re-ID data augmentation.
• To enhance the realism of synthetic occluded data by aligning with real-world data characteristics.
• To improve the robustness and performance of Re-ID models operating under occlusion.

Main Methods:

• Introduced the Frequency-Style-Position Data Augmentation (FSPDA) module with an occluder library, adaptive instance normalization, and hierarchical position rules.
• Developed the Multi-Scale Crop Data Augmentation (MSCDA) strategy using multi-scale cropping and dynamic view fusion to prevent information loss.
• Integrated FSPDA and MSCDA in parallel within the MA-MSA framework for collaborative resolution of dual inconsistencies.

Main Results:

• MA-MSA achieved state-leading performance on benchmark datasets: 73.3% Rank-1 and 62.9% mAP on Occluded-Duke.
• Achieved 87.3% Rank-1 and 82.1% mAP on Occluded-REID.
• Demonstrated superior robustness of the proposed method without relying on auxiliary models.

Conclusions:

• The MA-MSA framework effectively resolves dual inconsistencies in data augmentation for occluded Re-ID.
• The proposed method significantly improves Re-ID performance and robustness in challenging occlusion scenarios.
• MA-MSA offers a promising approach for generating realistic occluded data and enhancing Re-ID model capabilities.