The rise of computer vision systems has reshaped many industries, driven by the powerful capabilities of deep neural networks. However, as the complexity of these systems grows, so does the demand for larger datasets. However, the manual annotation of large-scale datasets comprising the diversity required by these systems is labour-intensive and time-consuming. This dissertation delves into the exploration of procedural generation techniques for hand pose datasets to overcome this challenge while also investigating the impact of controlled variations in detection quality and reliability, encompassing joint angles, wrist orientations, texture, lighting, and background variations, aiming to make it capable of handling diverse real-world settings. To assess the efficacy of the generated datasets, a state-of-the-art computer vision system is trained to detect key points in hand images using both the procedurally generated dataset and traditionally annotated datasets. Comparative analyses evaluate the trained system’s performance on real-world data, comprehending the influence of procedural variations on its accuracy, robustness, and generalisation capabilities. In conclusion, this dissertation contributes to hand pose estimation by integrating innovative approaches for procedurally generating datasets. The findings underscore the importance of automated variations in dataset generation and offer insights into their impact on the quality of trained computer fabrication systems.

Access my MSc Dissertation