Tensor Dictionary Learning for Positive Definite Matrices.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Sparse modeling is highly effective in image processing and computer vision.
• Previous work focused on sparse vectors and low-rank matrices, with less attention to positive definite descriptors.
• Region covariances are popular, motivating sparse coding for these descriptors.

Purpose of the Study:

• To propose a novel dictionary learning approach for positive definite matrices.
• To develop a discriminative version for classification and clustering tasks.
• To provide a software library for practical implementation.

Main Methods:

• Dictionary learning via alternating minimization between sparse coding and dictionary update.
• Exploration of different atom update strategies.
• Development of a discriminative approach for multi-class learning.

Main Results:

• Experimental validation of dictionary learning from data for reconstruction and classification.
• Demonstration of improved performance compared to previous methods.
• Successful implementation in a C++ software library.

Conclusions:

• Learning dictionaries directly from data is advantageous for positive definite sparse coding.
• The proposed discriminative approach enhances classification and clustering.
• The provided library facilitates the application of these techniques.