Supervised Neural Style Transfer as an Augmentation Technique for Facial Landmark Detection

Abstract Neural Style Transfer (NST) is a technique that applies the visual characteristics of one image to another while preserving the structural content. Traditionally used for artistic transformations, NST has more recently been adopted in new fields, such as domain adaptation and data augmentation. This study explores a novel application of the technique as a data augmentation strategy for landmark detection. To validate the approach as a feasibility study for a complex cat facial landmark detection task, we make two main contributions. First, we demonstrate that applying style transfer to cropped facial images, rather than full-body images, improves the structural consistency of generated images, which is necessary for the landmark detection task. Secondly, we suggest a Supervised Neural Style Transfer (SNST) technique specifically adapted for the landmark detection task, which outperforms traditional image augmentation methods and unsupervised NST. Our findings propose SNST as an effective augmentation strategy for improving the performance of landmark detection models in the animal domain and potentially beyond.