REMIND: Retrieval-Augmented Reconstruction With Dual Memories for Modality-Missing Object Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Modality-missing object Re-Identification (Re-ID) is challenging due to incomplete data.
• Existing reconstruction methods overlook modality-specific cues and use single-path strategies, leading to incomplete representations.

Purpose of the Study:

• To propose REMIND, a novel framework for modality-missing object Re-ID.
• To address limitations of existing reconstruction-based approaches by enhancing feature representation and reconstruction robustness.

Main Methods:

• Designed a Dual Memory Construction module extracting modality-specific and common features using information-theoretic insights.
• Developed a retrieval-augmented missing reconstruction module with multi-path reconstruction and perturbation.
• Employed adaptive fusion techniques for integrating features.

Main Results:

• REMIND achieves robust feature recovery and highly discriminative representations for reliable modality-missing Re-ID.
• Demonstrated effectiveness and superiority across multiple multi-modal object Re-ID benchmarks under various missing modality scenarios.

Conclusions:

• REMIND offers a superior approach to modality-missing object Re-ID.
• The framework's synergy of information-theoretic regularization and retrieval-enhanced reconstruction improves performance.