In this work we show how Physically Based Rendering tools can be used to extend the training image datasets of Machine Learning algorithms.