Discriminative shared Gaussian processes for multiview and view-invariant facial expression recognition

Area of Science:

• Computer Vision
• Machine Learning
• Affective Computing

Background:

• Facial expression recognition often involves images from various viewpoints.
• Current methods typically train view-specific classifiers or a single classifier for all views, ignoring expression redundancy across views.

Purpose of the Study:

• To propose a novel Discriminative Shared Gaussian Process Latent Variable Model (DS-GPLVM) for multiview and view-invariant facial expression classification.
• To leverage the redundancy between different views of the same facial expression for more effective classification.

Main Methods:

• Learning a discriminative manifold shared across multiple views of facial expressions.
• Performing facial expression classification within this learned expression manifold.
• Enabling both view-invariant (single-view) and multiview classification, as well as principled feature fusion.

Main Results:

• The DS-GPLVM model demonstrated superior performance compared to state-of-the-art methods in multiview and view-invariant facial expression classification.
• Validation on three public datasets (MultiPIE, LFW, SFEW) showed significant improvements.
• The model also excelled in multiview learning and feature fusion tasks.

Conclusions:

• The proposed DS-GPLVM effectively addresses the challenges of multiview facial expression recognition by learning a shared, discriminative manifold.
• This approach leads to enhanced classification accuracy and robust performance across different datasets and conditions.
• The model offers a principled way to integrate information from multiple views and features for improved facial expression understanding.