Deep Coattention-Based Comparator for Relative Representation Learning in Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (re-ID) aims to recognize individuals across disjoint camera views using discriminative representations.
• Current methods often rely on pair-wise similarity learning with fixed region features, lacking correlation between paired images.
• Existing region-based approaches independently detect and align local features, neglecting inter-image relationships.

Purpose of the Study:

• To introduce a novel deep coattention-based comparator (DCC) for enhanced person re-identification.
• To fuse codependent representations of paired images by correlating relevant parts and producing relative representations.
• To improve the accuracy and robustness of re-identification systems in surveillance scenarios.

Main Methods:

• The proposed Deep Coattention-based Comparator (DCC) fuses codependent representations of paired images.
• It correlates the best relevant image parts and generates relative representations, mimicking human foveation.
• The method concurrently detects distinct regions across images and fuses them for similarity learning.

Main Results:

• The DCC method demonstrates improved performance in person re-identification tasks.
• Achieved a 1.2-point gain in mean average precision (mAP) on the DukeMTMC-reID dataset.
• Obtained a 2.5-point gain in mAP on the Market-1501 dataset, showcasing state-of-the-art results.

Conclusions:

• The DCC approach effectively learns representations relative to test shots for robust re-identification.
• The method is well-suited for re-identifying pedestrians in complex surveillance environments.
• The proposed coattention mechanism significantly enhances similarity learning for person re-ID.