Visible-Infrared Person Re-Identification With Modality-Specific Memory Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visible-infrared person re-identification (VI-ReID) faces significant challenges due to modality discrepancies between visible and infrared images.
• Current methods often discard discriminative modality information by focusing on shared feature spaces.

Purpose of the Study:

• To propose a novel Modality-Specific Memory Network (MSMNet) for VI-ReID.
• To address the modality discrepancy by completing missing information and unifying feature spaces.
• To enhance VI-ReID performance by leveraging memory networks.

Main Methods:

• Developed a Modality-Specific Memory Network (MSMNet) that exploits missing modality information.
• Implemented three learning strategies: feature consistency, memory representativeness, and structural alignment.
• Utilized a memory network to propagate identity information between visible and infrared modalities.

Main Results:

• MSMNet effectively alleviates modality discrepancy, even with single-modality input.
• The proposed learning strategies guide the memory network for improved identity propagation.
• Experimental results on SYSU-MM01 and RegDB benchmarks show superior performance compared to state-of-the-art methods.

Conclusions:

• MSMNet offers a novel approach to VI-ReID by addressing modality discrepancies.
• The integration of missing modality completion and memory networks is a key innovation.
• The method demonstrates significant improvements in VI-ReID tasks on standard datasets.