Receptive Multi-granularity Representation for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Person re-identification (Re-ID) relies on consistent local details for robust representation across diverse environments.
• Existing stripe-based Re-ID methods struggle with balancing feature diversity, locality, and robustness, leading to semantic inconsistencies due to rigid partitioning and misalignment.
• Challenges in Re-ID include handling variations in pose, illumination, and background clutter.

Purpose of the Study:

• To propose a novel receptive multi-granularity learning approach to enhance stripe-based feature learning for person re-identification.
• To improve the trade-off between feature diversity, locality, and robustness in Re-ID models.
• To achieve more comprehensive and efficient feature representations without increasing model storage costs.

Main Methods:

• Implemented a receptive multi-granularity learning approach by performing local partitions on intermediate representations.
• Utilized significance-balanced activations for adaptive pooling to create uniform stripes.
• Introduced random shifting augmentation to mitigate misalignment issues.
• Employed a two-branch network architecture to learn discriminative identity representations at different scales.

Main Results:

• The proposed approach enhances the representation of locality while maintaining proper local association.
• Achieved state-of-the-art accuracy of 96.2%@Rank-1 and 90.0%@mAP on the Market-1501 benchmark.
• Demonstrated effectiveness through extensive intra-dataset and cross-dataset evaluations.

Conclusions:

• The receptive multi-granularity learning approach significantly improves person re-identification performance.
• The method offers a more comprehensive and efficient feature representation for Re-ID systems.
• The findings suggest a promising direction for addressing limitations in current stripe-based Re-ID techniques.