But really, if Google Street View data (or similar) is entirely part of the training dataset it is more than expected that it has this capability.