What techniques are currently available for e.g. locating the ZIP code on the envelope and cutting it to single digits prior to feeding them to digit recognition algorithm (trained on MNIST dataset)?
Or imagine a task of automatic recognition if a phone call recording contains a sound of a train passing a track switch (just yes or no). If present at all, it can be gentle or loud, depending if a window was open. It can be slow or fast, depending how quick the train was going through the switch.
I found so far Spatial Transformations Networks mentioned in this comment