Cascaded Structural Representation for Face Alignment
Wenjuan Xia, Jingwen Qin, Xiaofang Hong · 2020
This paper presents a highly intuitionistic, very efficient and accurate approach for face alignment. We directly gives an insight into the relationship between the feature space and the coordinate space. We first define a nonlinear function from local features extracted around landmarks to the deviation between estimated and ground truth landmarks. Then Taylor Expansion is used to approximate this nonlinear function. As this approximation works well in a short range of argument, we use a cascaded way to progressively refine the results. This is finally induced to a representation problem. Furthermore, our approach clearly demonstrates the structure and role of each cascaded module. A structured Schatten p-norm is defined to generate the Cascaded Structural Representation problem. Three kinds of structures are defined in this framework in the cascaded process to simulate the ideal situation for updating landmarks. Our approach achieves the state-of-the-art results when tested on current challenging benchmarks with fast speed on a desktop satisfying real-time performance.