Attribute-Guided Alignment Model for Person Re-Identification With Feature Distillation and Enhancement

Abstract

This paper proposes an attribute-guided alignment model for person re-identification with feature distillation and enhancement. The proposed attribute-guided part-level distillation model learns spatial body-part information, enabling the alignment of attribute features with their corresponding body regions. Hence, the localization ability of the attention mechanism in spatial information is fully exploited to address the feature misalignment issue in person re-identification tasks. Furthermore, an attribute-guided part-level feature enhancement method is proposed, which re-weights local features by selecting the top-3 most distinctive features for person re-identification during the inference phase. Experimental results on person re-identification datasets demonstrate the effectiveness of the proposed model.

Keywords

• Attention mechanisms
• convolutional neural networks
• deep learning
• feature extraction
• fine-grained image recognition
• identification of persons