Head Pose Estimation via Manifold Learning

Abstract

For the last decades, manifold learning has shown its advantage of efficient non-linear dimensionality reduction in data analysis. Based on the assumption that informative and discriminative representation of the data lies on a low-dimensional smooth manifold which implicitly embedded in the original high-dimensional space, manifold learning aims to learn the low-dimensional representation following some geometrical protocols, such as preserving piecewise local structure of the original data. Manifold learning also plays an important role in the applications of computer vision, i.e., face image analysis. According to the observations that many face-related research is benefitted by the head pose estimation, and the continuous variation of head pose can be modelled and interpreted as a low-dimensional smooth manifold, we will focus on the head pose estimation via manifold learning in this chapter. Generally, head pose is hard to directly explore from the high-dimensional space interpreted as face images, which is, however, can be efficiently represented in low-dimensional manifold. Therefore, in this chapter, classical manifold learning algorithms are introduced and the corresponding application on head pose estimation are elaborated. Several extensions of manifold learning algorithms which are developed especially for head pose estimation are also discussed and compared.

Tensor Decomposition and Non-linear Manifold Modeling for 3D Head Pose Estimation

“…Many methods have considered the model of the underlying manifold structure of head pose variations (Sundararajan and Woodard, 2015;Wang et al, 2017a). The main idea behind these methods is that, regardless of the dimensionality of the input features representing the mesh, there should be at most three degrees of freedom for head pose variation, thus defining a 3D manifold (Raytchev et al, 2004).…”

“…angles) into separate subspaces, thus obtaining specific parametrizations for each of them. On the other hand, manifold learning (Wang et al, 2017a) can be used to find the low-dimensional manifold structure defined by the orientation angles.…”

Local Neighbor Propagation Embedding

“…O VER the past few decades, manifold learning has already caused broad attention and applied in biological science [1], [2], image reconstruction [3], pose estimation [4], [5], [6], etc. The history of manifold learning can be tracked to some local algorithms such as Locally Linear Embedding (LLE) algorithm [7], Local Tangent Space Alignment (LTSA) algorithm [8] and some global algorithms such as the Isometric Mapping (ISOMAP) algorithm [9] and the Maximum Variance Unfolding (MVU) algorithm [10].…”