Robust facial landmark detection in the wild

Facial landmark detection is a well-studied topic in the field of computer vision that aims to find important key points in human faces. In the wild, the task is particularly challenging due to the high variability of shapes, expressions, poses, lighting conditions, and occlusions. This work presents a state-of-the-art approach to robustly solve the problem of facial landmark detection even under such difficult conditions. A key novelty of the presented approach lies in the fact that it is based on a fullyconvolutional architecture, making it invariant to translation. Translation invariance is particularly useful when a separate face detector is not available, desirable, or reliable (enough). Fully-convolutional architectures, however, suffer from a comparatively narrow receptive field. This shortcoming is mitigated by a novel implicit kernel convolution. Multiple experiments verify that the implicit kernel convolution improves both landmark detection performance and convergence speed in comparison to other state-of-the-art approaches. Moreover, a proof of concept for face detection-free landmark detection based on the novel approach is provided. High resolutions are handled by a pyramid-like multi-resolution fusion approach, whereas low resolutions are handled by a super resolution mechanism. The presented approach therefore constitutes a generalizable way of robustly detecting facial landmarks in the wild.