Kernel Distance Metric Learning Using Pairwise Constraints for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Person re-identification is crucial in computer vision but challenged by viewpoint variations.
• KISSME is an effective linear metric learning method but lacks power for complex scenarios.

Purpose of the Study:

• To enhance person re-identification by kernelizing the KISSME method for nonlinear transformations.
• To develop an efficient incremental version for sequential data processing.

Main Methods:

• Kernelization of the KISSME algorithm to achieve nonlinear transformations.
• Development of a fast incremental version for learning dissimilarity functions in feature space.
• Validation on five public datasets.

Main Results:

• The kernelized KISSME method demonstrates effectiveness for complex person re-identification tasks.
• The incremental version achieves competitive results efficiently without inverse covariance matrix estimation.
• The approach is suitable for learning metrics from structured objects without vectorization.

Conclusions:

• Kernelizing KISSME significantly improves person re-identification performance by enabling nonlinear modeling.
• The incremental variant provides an efficient and effective solution for real-time or sequential person re-identification applications.