This study explores semantic style transfer to enhance animal facial landmark detection, focusing on cat faces as a case study. The researchers demonstrate that cropping facial images before applying style transfer significantly improves the quality and structural consistency of generated images. Furthermore, while training models solely on style-transferred images can degrade performance, Supervised Style Transfer (SST), which selects style sources based on landmark accuracy, effectively mitigates this issue. Ultimately, augmenting existing datasets with these carefully curated, style-transferred images proves to be a powerful data augmentation strategy, outperforming traditional methods and improving the robustness and accuracy of facial landmark detection models for animals.
Supervised Neural Style Transfer as an Augmentation Technique for Facial Landmark Detection
A major challenge in training accurate facial landmark detection models
