Learning Modal-Invariant Angular Metric by Cyclic Projection Network for VIS-NIR Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Person re-identification across visible and near-infrared cameras (VIS-NIR Re-ID) is crucial for security and surveillance.
• Existing methods struggle with the modal gap between visible and near-infrared features, leading to poor performance.
• Heterogeneous image matching remains a significant challenge in VIS-NIR Re-ID.

Purpose of the Study:

• To address the performance degradation caused by the modal gap in VIS-NIR Re-ID.
• To propose a novel metric learning strategy for improved cross-modality feature matching.
• To enhance the accuracy and robustness of person re-identification systems.

Main Methods:

• A novel approach based on metric learning in a well-designed angular space is proposed.
• Features are mapped to a hypersphere manifold to eliminate norm variations and focus on angular relationships.
• A cyclic projection network (CPN) is introduced to transform features into an angle-related space while preserving identity information.
• Three angular metric learning loss functions (AICAL, LAL, DAL) are developed.

Main Results:

• The proposed method significantly outperforms existing state-of-the-art (SOTA) methods on benchmark datasets.
• Experiments on SYSU-MM01 and RegDB datasets demonstrate superior performance.
• The angular metric learning approach effectively bridges the modal gap between visible and near-infrared features.

Conclusions:

• The proposed angular metric learning method offers a promising solution for VIS-NIR person re-identification.
• Mapping features to an angular space effectively handles the modal gap and improves matching accuracy.
• The CPN and novel loss functions contribute to enhanced performance in cross-modality person re-identification.