SPD Manifold Deep Metric Learning for Image Set Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Manifold Learning

Background:

• Image set classification faces challenges with high intraclass variability and interclass similarity.
• Existing methods using symmetric positive definite (SPD) manifolds have limitations due to shallow, linear feature transformations.
• Capturing complex geometric features in visual data remains an open problem.

Purpose of the Study:

• To propose a novel SPD manifold deep metric learning (SMDML) approach for improved image set classification.
• To address limitations of shallow feature transformations in existing SPD manifold methods.
• To enhance the capture of pivotal geometric information for visual scene understanding.

Main Methods:

• Utilized a SPD manifold neural network (SPDNet) as an encoder for nonlinear SPD matrix representation.
• Incorporated a Riemannian decoder with a reconstruction error term (RT) to preserve structural information.
• Introduced a ReCov layer to regularize local statistical information and a novel metric learning regularization term.

Main Results:

• Demonstrated the feasibility of using Euclidean distance instead of Riemannian metric in the reconstruction error term.
• Showcased enhanced effectiveness of the learning process through the ReCov layer.
• Achieved superior performance on three typical visual classification tasks, validating the proposed SMDML approach.

Conclusions:

• The proposed SMDML approach effectively addresses limitations in existing image set classification methods.
• The novel deep metric learning framework enables powerful Riemannian representations and effective classifier training.
• SMDML offers a robust solution for complex visual classification tasks by capturing intricate geometric data variations.