Nonnegative Decompositions for Dynamic Visual Data Analysis.

Area of Science:

• Computer Vision
• Behavior Computing
• Machine Learning

Background:

• Analyzing dynamic, time-varying visual data is crucial in image, vision, and behavior computing.
• Traditional facial analysis often relies on facial landmarks, limiting its scope.
• Unsupervised analysis of temporal dynamics in facial expressions remains challenging.

Purpose of the Study:

• To develop a novel approach for unsupervised analysis of facial expression temporal phases and action units (AUs).
• To enable temporal alignment of facial behaviors between different individuals using raw pixel intensities.
• To improve the accuracy and robustness of dynamic facial behavior analysis.

Main Methods:

• Exploitation of raw pixel intensities instead of facial landmarks for face representation.
• Proposal of Slow Features Nonnegative Matrix Factorization (SFNMF) for learning part-based representations of temporal sequences.
• Extension of SFNMF with Dynamic Time Warping for temporal alignment of misaligned data sequences.

Main Results:

• Successfully achieved unsupervised detection of temporal phases for both posed and spontaneous facial events.
• Demonstrated effective temporal alignment of facial expressions between two individuals.
• Outperformed state-of-the-art methods in both unsupervised phase detection and temporal alignment tasks.

Conclusions:

• The proposed SFNMF method provides an effective framework for analyzing temporal dynamics in facial expressions using raw pixel data.
• The integration of Dynamic Time Warping enables robust temporal alignment of facial behaviors.
• This approach advances the field of dynamic facial behavior analysis, offering superior performance over existing methods.