Head pose estimation using improved label distribution learning with fewer annotations

Abstract

Head pose estimation in unconstrained environment remains a challenging task due to background clutter, illumination changes, and appearance variabilities. Multivariate label distribution has been successfully applied to head pose estimation. However, it is not applicable to unconstrained environments where assigning reasonable label distributions for images is difficult, and its performance significantly degrades when accurate grid information is unavailable (e.g., only yaw angles are known). To alleviate these problems, we propose an improved label distribution learning approach with fewer annotations. A data-driven weak learning strategy is first developed to construct label distributions to alleviate the problem of unreasonable label distributions. Regularization terms (e.g., L1,2 norm) are then introduced into the loss function induced by weighted Jeffreys divergence to avoid over-fitting. To further ameliorate the performance, positive correlation and negative competition are also introduced into the loss function to fine-tune the parameters of the corresponding model. Extensive experiments have been conducted on public databases: LFW and Pointing04. The proposed method achieves comparable performance over the state-of-art and possesses good generalization ability, but uses only fewer annotations, which suggests that it has strong potential for head pose estimation in unconstrained environments where sufficient annotations are routinely unavailable.