Semi-Supervised Unconstrained Head Pose Estimation in the Wild.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing head pose estimation datasets have limitations, including unrealistic synthetic data, constrained collection, or small-scale natural images with manual annotations.
• Fully-supervised methods struggle due to reliance on extensive labeled data, which is scarce for unconstrained, in-the-wild scenarios.

Purpose of the Study:

• To propose the first semi-supervised method (SemiUHPE) for unconstrained head pose estimation, leveraging abundant unlabeled head images.
• To adapt semi-supervised rotation regression for the error-sensitive and label-scarce problem of unconstrained head pose estimation.

Main Methods:

• Utilizes aspect-ratio invariant cropping for wild heads, outperforming landmark-based alignment when landmarks are unavailable.
• Introduces dynamic entropy-based filtering to adaptively remove outliers from pseudo-labeled data during training.
• Develops novel head-oriented strong augmentations: pose-irrelevant cut-occlusion and pose-altering rotation consistency.

Main Results:

• SemiUHPE significantly outperforms existing methods on public benchmarks for both front-range and full-range head pose estimation.
• Ablation studies validate the effectiveness of the proposed components and approach.
• The method demonstrates strong performance and versatility in related tasks like generic object rotation regression and 3D head reconstruction.

Conclusions:

• SemiUHPE offers a robust and effective solution for unconstrained head pose estimation by utilizing unlabeled data.
• The proposed techniques, including dynamic filtering and novel augmentations, enhance semi-supervised learning in this domain.
• The method's adaptability extends to other computer vision tasks, highlighting its extensibility.