Depth CNNs for RGB-D scene recognition: learning from scratch better than transferring from RGB-CNNs