A Functional Regression Approach to Facial Landmark Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Regression Analysis

Background:

• Linear regression is crucial for face detection and tracking algorithms, predicting shape changes from image features.
• Existing functional regression methods use B-splines or Fourier series, limiting real-time applications.

Purpose of the Study:

• To develop a novel functional regression solution for real-time incremental face tracking.
• To improve computational efficiency in face tracking algorithms.

Main Methods:

• Proposed Continuous Regression, approximating the input space via first-order Taylor expansion for a closed-form solution.
• Extended the continuous least squares problem to correlated variables.
• Integrated Continuous Regression into the cascaded regression framework, creating Cascaded Continuous Regression (CCR).
• Developed an incremental learning approach within CCR, named iCCR, for real-time performance.

Main Results:

• iCCR achieves real-time incremental face tracking, outperforming state-of-the-art methods by 20x.
• Demonstrated computational benefits of Continuous Regression in training and testing.
• Achieved state-of-the-art performance on the 300-VW face tracking benchmark.

Conclusions:

• Continuous Regression offers a computationally efficient and effective solution for face tracking.
• iCCR represents the first real-time incremental face tracker with state-of-the-art performance.
• The proposed method generalizes to correlated variables and shows significant speed improvements.