Robust Correlated and Individual Component Analysis.

Area of Science:

• Computer Vision
• Data Analysis
• Signal Processing

Background:

• Extracting correlated and individual components from datasets is crucial for image, vision, and behavior computing.
• Real-world data often suffers from non-Gaussian noise and temporal misalignments, complicating component extraction.

Purpose of the Study:

• To develop a robust method for analyzing two potentially misaligned data sets.
• To address the challenges of gross non-Gaussian noise and temporal incongruities in component analysis.

Main Methods:

• Propose Robust Correlated and Individual Component Analysis (RCICA) to handle gross, sparse errors in two data sets.
• Extend RCICA to manage temporal misalignments by solving two optimization problems.

Main Results:

• Demonstrate the method's generality across four applications: heterogeneous face recognition, multi-modal behavior analysis, face clustering, and facial expression temporal alignment.
• Experimental results on synthetic and real-world datasets show superior performance compared to state-of-the-art methods.

Conclusions:

• RCICA is a robust and effective method for extracting correlated and individual components from noisy and temporally misaligned data.
• The proposed approach significantly advances multi-modal fusion, predictive analysis, and clustering in computer vision and behavior computing.