Flexible Body Partition-Based Adversarial Learning for Visible Infrared Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Person re-identification (Re-ID) typically focuses on visible spectrum imagery, neglecting infrared (IR) data crucial for dark environments.
• Existing visible-infrared person Re-ID (VI-REID) methods often overlook detailed pedestrian information, limiting cross-modality retrieval performance.

Purpose of the Study:

• To develop an advanced VI-REID method that effectively utilizes fine-grained details from heterogeneous pedestrian images.
• To improve person retrieval accuracy in scenarios involving both visible and infrared camera views.

Main Methods:

• Proposed a Flexible Body Partition (FBP) model to automatically distinguish and represent pedestrian parts from feature maps.
• Introduced adversarial learning with a modality classifier to encourage modality-invariant feature extraction.
• Employed adaptive weighting for representation learning and a threefold triplet loss for metric learning to enhance feature discriminability.

Main Results:

• The FBP-AL method demonstrated superior performance on the SYSU-MM01 and RegDB cross-modality person Re-ID datasets.
• The approach effectively shrinks the cross-modality gap by learning more effective modality-sharable features.
• Fine-grained part representations significantly improved overall person re-identification accuracy.

Conclusions:

• The proposed FBP-AL method offers a significant advancement in visible-infrared person re-identification.
• Leveraging fine-grained details and adversarial learning is key to overcoming challenges in cross-modality person retrieval.
• The method shows strong potential for public security applications requiring robust person identification in diverse lighting conditions.