In modern tasks of Computer Science like robotics or automated driving systems Machine Learning algorithms are heavily used. For example, when creating a gesture-based control system for a remote-control robot, the classification of the gestures is usually achieved by utilizing Artificial Intelligence. These algorithms learn a certain function from a provided dataset and therefore rely on the quantity and quality of the data. In most cases this data is created by hand and might be limited. Training with limited data, however, leads to Overfitting. This work evaluates methods to reduce Overfitting mainly by Data Augmentation, i.e. the synthesizing of new images for the dataset form existing ones. Simple images processing methods are tested. These include linear transformations, bluring, sharpening, adding noise and rescaling. Additionally, the class labels are manipulated to force the network to learn specific features. This work attempts to measure the richness of the dataset, but the measurement is only a weak indicator for accuracy of the network. The recently developed generative models also are suitable for use in Data Augmentation. The class accuracies are balanced by generating samples for the less diverse classes. Segmentation data can also be generated by extracting the attention from the classification network.