An Improved Tiered Head Pose Estimation Network with Self-Adjust Loss Function.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Head pose estimation is crucial for applications like tiredness detection.
• Existing methods often treat head angles (roll, yaw, pitch) separately, ignoring their interdependence.
• Angle estimation discontinuity reduces accuracy in current head pose estimation algorithms.

Purpose of the Study:

• To propose a novel model, THESL-Net (tiered head pose estimation with self-adjust loss network), to overcome limitations in head pose estimation.
• To enhance the accuracy and consistency of estimating head pose angles.
• To investigate the impact of different angle ranges on model performance.

Main Methods:

• Introduced a tiered estimation approach using distinct network layers for greater angle estimation freedom.
• Identified causes of angle discontinuity, including data labeling and loss function design.
• Implemented a self-adjustment constraint on the loss function for improved angle consistency.

Main Results:

• The proposed THESL-Net model demonstrated superior performance compared to state-of-the-art methods.
• Experiments on BIWI, AFLW2000, and UPNA datasets validated the model's effectiveness across different angle ranges.
• The tiered estimation and self-adjusting loss function significantly improved angle estimation accuracy and consistency.

Conclusions:

• THESL-Net effectively addresses key challenges in head pose estimation.
• The novel approach offers improved accuracy and consistency for real-world computer vision applications.
• This work advances the field of head pose estimation with a more robust and accurate methodology.