Attention-Based Neural Architecture Search for Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person reidentification (reID) has advanced using deep neural networks.
• Current reID models face challenges like high complexity, lengthy pretraining, and backbone task mismatch.
• Expert-designed architectures often do not optimally suit the reID task.

Purpose of the Study:

• To introduce Neural Architecture Search (NAS) for automated person reID backbone design (reID-NAS).
• To develop attention-based network architectures for reID from scratch.
• To address limitations of traditional NAS and improve reID model efficiency and performance.

Main Methods:

• Designed a reID-specific search space incorporating a lightweight attention module.
• Introduced a novel retrieval-based search objective tailored for reID tasks.
• Employed a hybrid optimization strategy to enhance search stability in reID-NAS.

Main Results:

• The reID-NAS searched architecture achieved state-of-the-art performance on three reID datasets.
• The new architecture demonstrated a significant reduction in parameters (one order of magnitude).
• Reduced the need for extensive pretraining, enabling direct search and training from scratch.

Conclusions:

• reID-NAS effectively automates the design of efficient and high-performing person reID models.
• The attention-based architectures discovered by reID-NAS offer a lightweight and effective solution.
• This approach streamlines the reID model development process by reducing complexity and pretraining dependency.