Revisiting Person Re-Identification by Camera Selection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Surveillance Systems

Background:

• Person re-identification (Re-ID) is crucial for visual surveillance.
• Conventional Re-ID methods focus on pairwise similarity, risking inconsistent results across multiple cameras.
• Ensuring consistent retrieval across all views is vital for applications like epidemiological investigations.

Purpose of the Study:

• To address the challenge of consistently retrieving a target person across all camera views.
• To introduce the task of continuous person Re-ID and a metric, overall Rank-K accuracy.
• To develop a method for detecting defective cameras that degrade continuous Re-ID performance.

Main Methods:

• Proposed a continuous person Re-ID framework evaluating consistency across all camera views.
• Developed a relational deep Q-network to model visual and spatial camera relations.
• Collected a new dataset with camera topology information for evaluating spatial relations.

Main Results:

• The proposed method effectively detects defective cameras, outperforming random camera removal.
• Continuous person Re-ID performance is significantly degraded by defective cameras.
• The relational deep Q-network successfully selects properly deployed cameras.

Conclusions:

• Continuous person Re-ID is a more robust approach for surveillance than conventional methods.
• Identifying and addressing defective cameras is essential for reliable surveillance networks.
• The developed method provides a practical solution for assessing and improving camera network quality.