Related Experiment Videos
Enhancing Unsupervised Multi-Source Domain Adaptation for Person Re-Identification via Mixture of Experts and
Hao Li1,2, Yuyang Feng1, Xin Zhao1
1School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China.
Sensors (Basel, Switzerland)
|June 26, 2026
Summary
This study introduces a new framework for person re-identification (re-ID) using a Mixture of Experts feature extraction network and a Graph-Based Relation module. The method enhances cross-camera matching by balancing domain-invariant features and domain-specific styles.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Person re-identification (re-ID) is crucial for matching individuals across different camera views.
- Existing multi-source unsupervised domain adaptation (UDA) re-ID methods struggle with balancing domain-invariant features and domain-specific styles.
- These methods also fail to adequately model correlations among diverse source domains, limiting cross-domain generalization.
Purpose of the Study:
- To propose a novel multi-source UDA re-ID framework addressing limitations in current methods.
- To enhance the balance between domain-invariant feature learning and domain-specific style preservation.
- To effectively model implicit correlations among diverse source domains for improved generalization.
Main Methods:
- A Mixture of Experts feature extraction (MEFE) network with mixed Instance and Batch Normalization (MIBN) for robust domain-invariant features.
- An embedded domain-specific style information (DSI) module to preserve style details.
- A Graph-Based Relation (GBR) module using cascaded Graph Attention and Graph Convolution Networks (GATs/GCNs) for multi-source feature fusion.
- Center maximum mean discrepancy loss to minimize cross-domain distribution discrepancies.
Main Results:
- The proposed MEFE network with the DSI module extracts robust domain-invariant features while preserving style details.
- The GBR module effectively fuses multi-source features by exploring implicit correlations using GATs/GCNs.
- The framework achieves state-of-the-art performance on large-scale datasets, outperforming mainstream UDA re-ID approaches.
Conclusions:
- The novel framework effectively balances domain-invariant feature learning and domain-specific style preservation in multi-source UDA re-ID.
- The proposed method significantly improves cross-domain generalization performance.
- This work offers a substantial advancement in person re-identification technology.