Stochastic style perturbation modelling for visible-Infrared person re-Identification with severely modality...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Visible-Infrared Person Re-Identification (VI-ReID) faces significant challenges due to data imbalance between visible (RGB) and infrared (IR) modalities.
• Limited IR data, often caused by cost and operational constraints, leads to underrepresentation during model training, hindering performance.

Purpose of the Study:

• To develop a novel framework, Co-Modality Balance Learning (CMBL), to effectively address data imbalance in VI-ReID.
• To improve the accuracy and robustness of person re-identification across different visual modalities.

Main Methods:

• Introduced a Stochastic Style Perturbation (SSP) module to generate synthetic IR samples, balancing the dataset.
• Developed a cross-distribution alignment loss for refined optimization of sparse modality features.
• Proposed Class-Aware Contrast Similarity Learning (CACS) to enhance feature discriminability.

Main Results:

• The CMBL framework demonstrated significant improvements on public VI-ReID datasets under imbalanced conditions.
• Ablation studies confirmed the efficacy of individual components (SSP, cross-distribution loss, CACS) in mitigating data imbalance effects.
• The approach successfully improved intra-class compactness and inter-class separation.

Conclusions:

• The proposed CMBL framework effectively tackles data imbalance in VI-ReID.
• The methods enhance feature extraction, leading to more accurate and robust cross-modality person re-identification.
• This work provides a valuable solution for real-world VI-ReID applications with inherent data disparities.