Data augmentations in mixed reality machine learning applications

Machine learning models have accomplished much in the modern day. Nevertheless, they are reliant on big datasets to have practical relevance. Since it is not always possible to obtain masses of data, augmentations of already present data has become an appealing alternative. In this thesis, a data augmentation system is proposed that uses a mixed reality environment to create augmented image data for the task of classification. In order to achieve this, ArUco markers are tracked in a picture, which are then used to insert any virtual object onto the marker using a homography. Finally, the augmentations are evaluated by training a neural network with the augmented and real data as input datasets. The proposed system achieves augmentations, which can partly substitute real model data in a machine learning application. This indicates the possibility to create data augmentations using a mixed reality approach that can expand or substitute existing datasets with augmented pictures in an image classification task trained on a machine learning model.