Person Re-Identification by Dual-Regularized KISS Metric Learning

Area of Science:

• Computer Vision
• Machine Learning
• Intelligent Video Surveillance

Background:

• Person re-identification (re-ID) is crucial for intelligent video surveillance.
• Existing methods struggle with performance due to unstable covariance matrix estimation in metric learning, especially with small datasets.
• The Keep It Simple and Straightforward (KISS) metric learning method shows promise but has limitations.

Purpose of the Study:

• To introduce a novel metric learning method, dual-regularized KISS (DR-KISS), to address the instability issues in person re-identification.
• To improve the robustness and generalization of person re-ID algorithms.

Main Methods:

• Developed DR-KISS by regularizing two covariance matrices to prevent overestimation of large eigenvalues.
• Provided theoretical analyses to justify the necessity and robustness of the regularization.
• Conducted extensive experiments on benchmark datasets: VIPeR, GRID, and CUHK 01.

Main Results:

• DR-KISS guarantees the covariance matrix is irreversible, overcoming limitations of standard KISS.
• Theoretical analyses confirm the method's robustness for generalization.
• Achieved new state-of-the-art performance on three challenging person re-identification datasets.

Conclusions:

• DR-KISS offers a significant advancement in person re-identification by providing a stable and effective metric learning approach.
• The proposed regularization technique enhances the reliability and performance of person re-ID systems.
• This work contributes to more effective intelligent video surveillance solutions.