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

A major challenge in training accurate facial landmark detection models for animals is the scarcity of diverse data, as classical image augmentations often disrupt the crucial spatial alignments required for this task. To address this gap, this study introduces a novel data augmentation technique called Supervised Neural Style Transfer (SNST). Focusing on cats, the researchers synthetically expanded their training dataset by transferring textural styles from top-performing examples onto cropped facial images, which perfectly preserved the original geometric structure of the faces. This specific approach effectively decouples appearance from shape, forcing the models to rely on robust structural features rather than surface-level textures. Ultimately, the study demonstrates that SNST significantly outperforms traditional augmentation methods, substantially reducing detection errors and improving the models’ overall robustness.