Robust Head-Pose Estimation Based on Partially-Latent Mixture of Linear Regressions

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Head-pose estimation is crucial for applications like human-robot interaction and driving assistance.
• Accurate head-pose estimation faces challenges including illumination variations, facial appearance changes, and occlusions.

Purpose of the Study:

• To develop a robust head-pose estimation method.
• To address the challenges of varying conditions and errors in head-pose estimation.

Main Methods:

• A mixture of linear regressions with partially-latent output is proposed.
• The method maps high-dimensional feature vectors from face bounding boxes to head-pose angles and bounding-box shifts.
• Unsupervised manifold learning techniques are combined with mixture of regressions.

Main Results:

• The proposed regression method demonstrates robust head-pose prediction.
• The algorithm is validated on three public datasets.
• Four variants of the algorithm were benchmarked against state-of-the-art methods.

Conclusions:

• The developed method offers a robust solution for head-pose estimation.
• The approach effectively handles unobservable phenomena impacting prediction accuracy.
• This technique advances the field of head-pose estimation for various applications.