Also, the reason that ReCAPTCHA use street number images is that they are difficult for machines to read. It wouldn’t surprise me if what they are actually doing is throwing in the occasional unrecognized number and actually learning how people recognize it. People expect to get a proportion of these images wrong., and maybe it even trusts some low risk users to do the initial recognition. (There may also be an element of Google having vastly better image recognition than organised crime, or other enterprises, such as OSM.)
In the UK, as well as the large variety of presentation of the numbers, in my experience only about one in three houses still has a number on display.
I’d certainly agree that some people would find cataloguing street numbers much more interesting than blinding taking geotagged photographs.
Yes, you are quite correct there but even fragmented data is useful because in some cases, e.g. terraced housing, the rest can be interpolated (useful for geocoding, for instance). I am just throwing house numbers out there as an example but are heaps of features which could be recognised using CV to improve attribute data, thus cartography, thus the map. E.g. whether a tree is coniferous or deciduous.