WC2FEst-Net: Wavelet-Based Coarse-to-Fine Head Pose Estimation from a Single Image

Zhen Li, Wei Li, Zifei Jiang, Peng Jiang, Xueqing Li, Yan Huang, Jingliang Peng

Abstract: This paper proposes a novel head pose estimation scheme that is based on image and wavelets input and conducts a coarse to fine regression. As wavelets provide low-level shape abstractions, we add them as extra channels to the input to help the neural network to make better estimation and converge. We design a coarse-to-fine regression framework that makes coarse-grained head pose classification followed by fine-grained angles estimation. This framework helps alleviate the influence of biased training sample distribution, and combines segment-wise mappings to form a better global fitting. Further, multiple streams are used in the neural network to extract a rich feature set for robust and accurate regression. Experiments show that the proposed method outperforms the state-of-the-art methods of the same type for the head pose estimation task.