IDALib: a Python library for efficient image data augmentation

The accuracy obtained with deep learning-based systems usually depends on the availability of large image datasets, which is not always possible. Consequently, it is necessary to apply techniques to increase the size of these datasets and their variability in a reliable and efficient way. In this respect, a novel tool for image augmentation and the associated data, IDALib, is presented. It provides an automatic method to perform transformation operations jointly on the images and related data, such as landmarks or masks, with the main aim of minimising the computational overhead. Thus, it applies automatically a set of optimisations, such as the vectorisation and composition of operations. Furthermore, the transformations are performed in GPU, which leads to a notable speedup, specially in dual GPU setups. IDALib is publicly available in the PyPI repository,

keywords: data augmentation, computer vision, deep learning