SD²-ReID: A semantic-stylistic decoupled distillation framework for robust multi-modal object re-identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multi-modal object re-identification (ReID) faces challenges in aligning style differences across modalities while maintaining identity consistency.
• Existing methods struggle to decouple semantic features from modality-specific styles, leading to noisy representations and reduced performance.

Purpose of the Study:

• To propose a novel multi-modal ReID framework, SD²-ReID (Semantic-Stylistic Decoupled Distillation for ReID), to improve modal consistency and cross-modal semantic discrimination.
• To address the limitations of existing methods in separating semantic and stylistic features for more robust ReID.

Main Methods:

• Developed a Hybrid Multi-modal Feature Extractor (HMFE) with shared shallow and deep modality-specific structures for fine-grained feature extraction.
• Introduced a Decoupled Distillation Module (DDM) using dual constraints for explicit semantic and style feature separation.
• Integrated an attention-guided masking strategy and contrastive learning within a Hierarchical Self-supervised Learning Module (HSLM) to enhance robustness.

Main Results:

• Achieved synergistic enhancement of semantic consistency, modal invariance, and feature robustness.
• Demonstrated effectiveness on three multi-modal object ReID benchmark datasets.
• Validated improved cross-modal semantic consistency and discriminative ability.

Conclusions:

• SD²-ReID effectively reconciles style discrepancies and semantic consistency in multi-modal ReID.
• The proposed method balances recognition performance and inference efficiency without requiring multi-modal fusion modules.
• The framework enhances robustness against local occlusions and style variations, outperforming existing approaches.