Cross-Modality Person Re-Identification via Modality-aware Collaborative Ensemble Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visible Thermal Person Re-Identification (VT-ReID) is a complex cross-modality retrieval task.
• Existing methods primarily address feature-level modality discrepancies, neglecting classifier-level differences.
• This limitation results in suboptimal discriminative power for VT-ReID systems.

Purpose of the Study:

• To propose a novel modality-aware collaborative ensemble (MACE) learning method for VT-ReID.
• To address modality discrepancies at both feature and classifier levels.
• To enhance the discriminative capability and overall performance of VT-ReID systems.

Main Methods:

• Introduced a middle-level sharable two-stream network (MSTN) to capture discriminative, sharable features in convolutional layers.
• Developed modality-specific and modality-sharable identity classifiers to manage classifier-level discrepancies.
• Implemented an ensemble learning scheme and a collaborative learning strategy to integrate classifier outputs and regularize predictions.

Main Results:

• The proposed MACE method significantly outperforms existing state-of-the-art VT-ReID approaches.
• Achieved Rank-1/mAP accuracy of 51.64%/50.11% on the SYSU-MM01 dataset.
• Attained Rank-1/mAP accuracy of 72.37%/69.09% on the RegDB dataset.

Conclusions:

• The proposed method effectively handles modality discrepancies at both feature and classifier levels.
• MSTN captures superior middle-level sharable features compared to existing methods.
• The modality-aware collaborative ensemble approach demonstrates superior performance in cross-modality person re-identification.