Data Augmentation Does Not Necessarily Beat a Smart Algorithm

According to the “widely acknowledged truth”, more training data beats algorithmic improvements in machine learning tasks. We challenge this “widely acknowledged truth” in context of data augmentation of images and recognition tasks related to images. Our observations show that real training data may be much more valuable than augmented (i.e., artificially generated) data and – most importantly – the advantage of a sophisticated algorithm relative to a simple algorithm may not be easily compensated by data augmentation