Person Re-Identification With Reinforced Attribute Attention Selection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (Re-ID) is crucial for video surveillance, matching individuals across different camera views.
• Existing Re-ID methods utilize attribute information to enhance performance, but identity-level attributes can be noisy and unhelpful.
• Image-level attribute annotations are labor-intensive, while identity-level attributes are more efficient but prone to noise.

Purpose of the Study:

• To develop a novel method for improving person re-identification (Re-ID) performance by effectively handling noisy identity-level attribute information.
• To introduce an Attribute Attentional Block (AAB) that can be integrated into existing Re-ID frameworks.
• To leverage reinforcement learning for attribute selection and attention mechanisms to boost Re-ID accuracy.

Main Methods:

• Proposed a novel Attribute Attentional Block (AAB) for person re-identification.
• Employed reinforcement learning with a designed reward system to identify and discard noisy attributes.
• Utilized aggregated attribute attention from the remaining discriminative attributes to improve Re-ID.
• Integrated the AAB into various backbone networks and frameworks.

Main Results:

• The proposed AAB method achieved state-of-the-art results on three benchmark person re-identification datasets.
• Demonstrated the effectiveness of reinforcement learning in filtering noisy attributes for improved Re-ID.
• Showcased the versatility of the AAB by integrating it into different Re-ID architectures.

Conclusions:

• The Attribute Attentional Block (AAB) offers a robust solution for noisy attribute handling in person re-identification.
• The proposed method significantly enhances Re-ID performance by focusing on relevant identity attributes.
• This approach provides a valuable contribution to the field of computer vision and surveillance.