Deep Learning for Person Re-Identification: A Survey and Outlook

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (Re-ID) is crucial for intelligent video surveillance, leveraging deep neural networks.
• Existing research primarily focuses on the closed-world setting, which has limitations for real-world applications.
• The field is transitioning towards the more challenging open-world Re-ID setting.

Purpose of the Study:

• To provide a comprehensive overview and analysis of both closed-world and open-world person Re-ID.
• To introduce a novel baseline (AGW) and evaluation metric (mINP) for advancing person Re-ID research.
• To identify and discuss under-investigated open issues in the field.

Main Methods:

• Analysis of closed-world Re-ID through deep feature representation learning, deep metric learning, and ranking optimization.
• Summarization of open-world Re-ID methodologies across five key aspects.
• Development and evaluation of the AGW baseline and the mINP metric on multiple datasets and tasks.

Main Results:

• The AGW baseline achieves state-of-the-art or comparable performance on twelve datasets for four Re-ID tasks.
• The proposed mINP metric offers a valuable new criterion for evaluating Re-ID system efficiency in practical scenarios.
• The study highlights performance saturation in closed-world Re-ID, underscoring the importance of open-world research.

Conclusions:

• The AGW baseline and mINP metric represent significant advancements for person Re-ID.
• Open-world Re-ID is critical for practical surveillance applications, requiring further investigation.
• The paper identifies key future research directions in person Re-ID.