Large-Scale Stochastic Scene Generation and Semantic Annotation for Deep Convolutional Neural Network Training in the RoboCup SPL

Object detection and classification are essential tasks for any robotics scenario, where data-driven approaches, specifically deep learning techniques, have been widely adopted in recent years. However, in the context of the RoboCup standard platform league these methods have not yet gained comparable popularity in large part due to the lack of (publicly) available large enough data sets that involve a tedious gathering and error-prone manual annotation process. We propose a framework for stochastic scene generation, rendering and automatic creation of semantically annotated ground truth masks. Used as training data in conjunction with deep convolutional neural networks we demonstrate compelling classification accuracy on real-world data in a multi-class setting. An evaluation on multiple neural network architectures with varying depth and representational capacity, corresponding run-times on current NAO-H25 hardware, and required sampled training data is provided.