Training of Convolutional Networks on Multiple Heterogeneous Datasets for Street Scene Semantic Segmentation.