Toward Robust and Unconstrained Full Range of Rotation Head Pose Estimation

Area of Science:

• Computer Vision and Machine Learning
• Robotics and Human-Computer Interaction

Background:

• Head pose estimation is vital for many applications but current methods struggle with full-range orientation prediction.
• Existing techniques often face challenges with ambiguous rotation labels and limited prediction capabilities.

Purpose of the Study:

• To develop a novel, unconstrained, end-to-end method for full range of orientation head pose estimation.
• To overcome limitations of current state-of-the-art methods by addressing ambiguous rotation labels and improving prediction robustness.

Main Methods:

• Introduced rotation matrix formalism for ground truth data to resolve label ambiguity.
• Proposed a continuous 6D rotation matrix representation for direct regression, enabling efficient learning of full rotation appearance.
• Utilized newly accumulated training data with full head pose rotation and a geodesic loss approach for stable learning.

Main Results:

• The developed model demonstrates significantly improved performance over state-of-the-art methods on public datasets.
• The method is efficient and robust in predicting head orientations across an extended range.
• The approach successfully overcomes limitations of previous head pose estimation techniques.

Conclusions:

• The novel 6D rotation matrix representation enables robust and efficient full range head pose estimation.
• The method's advanced prediction range expands potential applications in computer vision and human-computer interaction.
• Open-sourced code and models facilitate further research and development in head pose estimation.