One example: OpenSolarMap... We first start by crowdsourcing building roof orientations using a very simple webapp (no need to register, open to anybody). When enough contribution match they are considered OK (at least 3 more than all other contributions).
Then, these contributions were used to train a neural network. Then the nueral network was used to classify other roofs... and the result has been put back as robot contribution to the crowdsourcing webapp counting for 1 or 2 contributions depending on the level of confidence (raw data is also available for download). In all cases, there is always at least one human contribution, before putting anything back to OSM. It is also interesting to compare when human and robot do not agree ;) Links... http://opensolarmap.org/ https://github.com/opensolarmap Next step is to use the same technique on other kind of challenges, like: - landuse boundaries (to speedup/simplify Corine Land cover import improvements) - check road alignment with aerial imagery on "old" OSM traced contributions etc... The potential of deep learning mixed with human contributions can give very good things if done properly. 2016-12-22 17:48 GMT+01:00 Mikel Maron <mikel.ma...@gmail.com>: > > Frederik, all > > > an editor plugin were to help the mapper trace buildings that the > mapper identifies or at least individually verifies, that would probably > be ok > > This feels like the consensus across the board -- machine learning has > potential to be useful when integrated into a human editor workflow. Maybe > we can work on guidelines that encapsulates this. With something written > up, we'll be able to stop "spinning wheels" on whether this is useful or > not, and focus on experimenting and implementing promising approaches. > > -Mikel > > * Mikel Maron * +14152835207 @mikel s:mikelmaron > > > On Wednesday, December 21, 2016 7:59 PM, Frederik Ramm < > frede...@remote.org> wrote: > > > > Hi, > > On 12/22/2016 01:10 AM, john whelan wrote: > > Do we have any guidelines in the wiki etc? > > Nothing specific, no. > > Automated editing and/or import guidelines would apply to any such > process and I would ask everyone who overhears discussions about > "uploading" machine-detected data to OSM to point this out to those > discussing. We've already had to revert a couple hundred thousand such > edits (roads though, not buildings). > > If, OTOH, an editor plugin were to help the mapper trace buildings that > the mapper identifies or at least individually verifies, that would > probably be ok, at least until HOT trains an army of monkeys with > typewriters, er keyboards, to rubber-stamp everything the algorithm puts > out ;) > > More generally speaking, in my opinion the human-centered aspect of > mapping is a key property that sets us apart from other map databases. > You can safely assume that any algorithm we can run to detect buildings, > Google can run 1000 times faster and with a fraction of the error rate, > leading to 1000 times more and 10 times better data of that kind than we > can accumulate. This is not a field in which we can, or should attempt > to, compete. > > Bye > Frederik > > -- > Frederik Ramm ## eMail frede...@remote.org ## N49°00'09" E008°23'33" > > > _______________________________________________ > talk mailing list > talk@openstreetmap.org > https://lists.openstreetmap.org/listinfo/talk > > > > _______________________________________________ > talk mailing list > talk@openstreetmap.org > https://lists.openstreetmap.org/listinfo/talk > > -- Christian Quest - OpenStreetMap France
_______________________________________________ talk mailing list talk@openstreetmap.org https://lists.openstreetmap.org/listinfo/talk
