GSOC Application: Data Augmentation + An extra addition

[alexis.dra...@gmail.com]

I am posting to make you aware of my interest in [project idea 21](https://github.com/opencv/opencv/wiki/GSoC_2020#idea-data-augmentation).

I am an MSc student in computational applied mathematics with a background in computer vision. I started programming in 2017 and now use Python & R on a daily basis. My background in programming is primarily focused on scientific computing. My background in computer vision comes in the form of two internships + BSc thesis project. I will attach a research project I completed here: https://github.com/alexisdrakopoulos/ising_cnn

I have recently become interested in data augmentation due to this paper in CVPR 2019: http://openaccess.thecvf.com/content_CVPR_2019/papers/Cubuk_AutoAugment_Learning_Augmentation_Strategies_From_Data_CVPR_2019_paper.pdf

I know the project is focused on implementing core data augmentation features, but I am also interested if time permits to look into adding functionality for an implementation of the above paper (or similar).

The core functionalities discussed in project 21 would be implemented.

I would like to get input from an experienced engineer (as my experience in design patterns/OOP is limited) in terms of the best approach for implementing a general purpose image augmentation class.

This would then allow for users to easily select the types of augmentations seemlessly, while selecting respective probabilities in a pipeline. Importantly I would like this to work with existing pipelines to fascilitate the common issue with augmenting data pulled from disk in batches.

The current problem, for example using TF2, is that augmentation is incredibly easy when data fits in memory as well as when using the basic Keras generator but extremely painful when using the tf.data API currently.

I would also like to add functionality for a comprehensive report output (log file or something similar) so that the user can know exactly what augmentations were applied, to which images and in what percentages.

Thanks in advance and apologies for posting this so close to the deadline.