A Systematic Evaluation and Benchmark for Person Re-Identification: Features, Metrics, and Datasets

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (re-id) is crucial for video analytics in security and surveillance.
• Rapid progress in re-id algorithms is hindered by inconsistent evaluation protocols and metrics.

Purpose of the Study:

• To provide an extensive review and performance evaluation of single- and multi-shot person re-identification algorithms.
• To establish a standardized benchmark for comparing re-id methods.

Main Methods:

• Developed a unified code library implementing 11 feature extraction algorithms and 22 metric learning/ranking techniques.
• Evaluated algorithms using a new large-scale dataset and 16 other public datasets.
• Incorporated recent advances in feature extraction and metric learning into the experimental protocol.

Main Results:

• Facilitated direct comparison of diverse re-id algorithms under a consistent experimental setup.
• The evaluation codebase and results will be publicly released to benefit the research community.

Conclusions:

• A standardized evaluation framework is essential for advancing person re-identification research.
• The provided benchmark enables fair comparison and accelerates progress in surveillance and video analytics.