Real-time synthetic-to-real human detection for robotics applications

During the recent years, Deep Learning achieved exceptional performance in various computer vision tasks, paving auspicious research directions for its application in robotics. A key component for its exceptional performance is the availability of sufficient training data. However obtaining such amount of training data constitutes a challenging task, especially considering robotics applications. Thus, synthetic data have recently been regarded as a promising tool to overcoming the data availability problem. In this work we first build a synthetic human dataset, and then we train a lightweight model, capable of operating in real-time for high-resolution input on low-power GPUs, for discriminating between humans and non-humans. The target of this work is to assess the generalization of the model trained on synthetic data, to real data, and also to explore the effect of using (few) real images in the training phase. As it is shown through quantitative and qualitative results the use of only few real images can beneficially affect of the performance of the synthetic-to-real real-time model.