Data-driven neural network models trained on human motion data facilitate human activity recognition and identity verification applications. However, large annotated and processed human motion datasets are scarce, leading to the overfitting of models to training data. Thus, it is important to investigate data augmentation techniques to generate additional data to facilitate model generalisation. In addition, the choice of augmentation techniques severely impacts the performance of self-supervised learning. Thus, this work experiments and evaluates various augmentation techniques on seven sensor-based human activity datasets. Three supervised neural network models and one self-supervised learning model were experimented with. We note that due to the high variability in the performance of algorithmic augmentation techniques on time-series human activity datasets, generative data is highly influential in this domain.