Data augmentation is especially useful to teach invariance properties to the network. Augmentation is applied when only a few training samples are available, or when the desired property is not present in the dataset.
The lack of samples does not pose a problem, since the images are simulated. On the other hand, possessing data with certain properties that cannot be simulated can be essential in a project. Labels are augmented together with the images with the purpose of having a larger dataset. These augmentation transformations can include rotations, translations, scaling, cropping, etc.
Image and label augmentation example. The applied transformations are translation and scaling.