SymNet: A Simple Symmetric Positive Definite Manifold Deep Learning Method for Image Set Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Manifold Learning

Background:

• Visual classification using image sets often represents data as covariance matrices on the symmetric positive definite (SPD) manifold.
• A significant challenge in this domain is managing large within-class variability in data representations.
• Existing SPD matrix learning methods aim to address this by embedding data into lower-dimensional manifolds, preserving geometric structures.

Purpose of the Study:

• To propose a novel deep learning network, SymNet, for robust image set classification.
• To address the challenge of large within-class variability in SPD matrix representations.
• To leverage Riemannian geometry for improved discriminative feature learning.

Main Methods:

• SymNet employs SPD matrix mapping layers for dimensionality reduction.
• Rectifying layers introduce nonlinearity, and pooling layers compress SPD matrices.
• A log-map layer embeds matrices into the tangent space for Euclidean learning, utilizing (2-D)^2 PCA for weight learning and Kernel Discriminant Analysis (KDA) for classification.

Main Results:

• The proposed SymNet effectively reduces within-class variability in SPD matrix representations.
• The network successfully preserves the Riemannian geometrical structure of the data manifold.
• Experiments on six visual classification tasks demonstrate SymNet's superior performance compared to state-of-the-art methods.

Conclusions:

• SymNet offers a feasible and valid approach for image set classification, outperforming existing methods.
• The network's design facilitates easier training and implementation through efficient weight learning.
• This work highlights the potential of deep learning on SPD manifolds for advanced visual recognition tasks.