Distance Metric Learning via Iterated Support Vector Machines.

Area of Science:

• Machine Learning
• Computer Vision
• Pattern Recognition

Background:

• Distance metric learning enhances data sample similarity evaluation for classification.
• Existing metric learning methods often rely on inefficient custom optimizers for large-scale problems.

Purpose of the Study:

• To reformulate metric learning as a kernel classification problem with positive semi-definite constraints.
• To develop efficient and globally optimal metric learning models.

Main Methods:

• Formulated metric learning as a kernel classification problem with positive semi-definite constraints.
• Solved the problem using iterated training of Support Vector Machines (SVMs).
• Developed two novel models: Positive-Semidefinite Constrained Metric Learning (PCML) and Nonnegative-Coefficient Constrained Metric Learning (NCML).

Main Results:

• The new formulation is easy to implement and efficient using off-the-shelf SVM solvers.
• Both PCML and NCML guarantee global optimality.
• Experimental results on general classification, face verification, and person re-identification show comparable accuracy to state-of-the-art methods with improved training efficiency.

Conclusions:

• The proposed kernel classification approach offers an efficient and effective alternative for metric learning.
• PCML and NCML provide globally optimal solutions and demonstrate strong performance in various computer vision tasks.