Overcoming data limitations: Few-shots classification for Radar-Based Hand Gesture Recognition

Few-shot learning is a cutting-edge machine learning paradigm designed to address the challenge of limited data availability. This research introduces a novel approach to alleviate data constraints in radar-based hand gesture recognition. By enabling models to effectively learn new gestures with only a few labeled examples, few-shot learning opens up new possibilities for gesture recognition systems with reduced data dependencies. In this study, we employed a meta-learning algorithm, Model-Agnostic Meta-Learning (MAML), to tackle this task. A new dataset, recorded from scratch, was utilized to train and test the model on the given task. Several benchmark configurations are presented in this work, serving as valuable resources for future studies on this dataset. Additionally, an extensive experiment was conducted to explore the adaptability of the proposed methodology to new tasks. The adaptation performance was then compared with the one of a conventionally trained model, clearly highlighting the strengths of MAML in comparison. Lastly, the model was tested in a live-inference scenario, although improvements are needed in this aspect.