Comparison of Training for Hand Gesture Recognition for Synthetic and Real Datasets

Human gesture recognition is often implemented in many HRI applications. Building datasets that involve human subjects, when aiming to capture comprehensive diversity and all possible edge cases is often both challenging and labor-intensive. While applying the concept of domain randomization to build synthetic datasets helps address the problem, an innate reality gap always exists that needs to be mitigated. In this paper, We present and discuss a comprehensive performance comparison of our synth datasets with real ones and demonstrate the results in this paper.