Discrete Metric Learning for Fast Image Set Classification.

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• Image set classification faces challenges in efficiently handling discriminative features.
• Existing methods often struggle with computational complexity and memory costs.

Purpose of the Study:

• To develop a novel, efficient approach for image set classification using hashing and Riemannian manifolds.
• To introduce Discrete Metric Learning (DML) and Bilinear Discrete Metric Learning (BDML) for fast and accurate classification.

Main Methods:

• Modeling image sets as points on a Riemannian manifold.
• Jointly learning a metric in an induced space and a compact Hamming space.
• Developing DML and BDML for efficient optimization via geodesic mean computation.

Main Results:

• Proposed DML and BDML achieve competitive accuracy and efficiency in visual recognition tasks.
• Methods demonstrate effectiveness in face, object, and action recognition.
• BDML overcomes DML's vectorization limitations by directly manipulating Riemannian representations.

Conclusions:

• DML and BDML offer a computationally efficient and memory-saving solution for image set classification.
• The proposed methods provide a significant advancement over conventional Riemannian metric learning techniques.
• The approach is broadly applicable to various visual recognition challenges.