Transfer Learning with Style Transfer between the Photorealistic and Artistic Domain

Transfer Learning is an important strategy in Computer Vision to tackle problems the face of limited training data. However, this still heavily depends on amount availabl data, which a challenge for small heritage institutions. This paper investigates various ways enrichingsmaller digital collections boost performance deep learningmodels, using identification musical instruments as case study. We apply traditional data augmentation techniques well use external, photorealistic collection, distorted by Style Transfer. are capable artistically stylizing images, reusing style from any other given image. Hence, can be easily augmented with artificially generated images. introduce distinction between inner and outer transfer show that images both scenarios consistently improve classification results, top techniques. counter-intuitively, such artistic depictions works surprisingly hard classify. In addition, we discuss example negative within non-photorealistic domain.