[OSM-talk] Announcing the Tabang-AI initiative
Eugene Alvin Villar
seav80 at gmail.com
Sat Aug 10 17:39:46 UTC 2019
On Sat, Aug 10, 2019 at 5:55 PM Frederik Ramm <frederik at remote.org> wrote:
> do I understand correctly that you and your local partners aim to
> recruit new mappers to OSM, who will not learn the "old fashioned"
> workflow of tracing stuff from imagery by hand, but be mainly taught to
> work with pre-processed Facebook road data?
Not at all. We have no plans to teach new contributors to *only* contribute
using the RapiD editor or using FB-provided data. We simply want to
incorporate these road-detections as an optional complement to the
trainings and workshops that we have been conducting over the past decade.
> How will you ensure that your partners give those new mappers a training
> that is good enough to know when to *not* trust the pre-processed AI
> data? All too often people automatically assume that "the computer is
> always right", and this would be especially the case in a mapathon setup
> where time is limited. Will local new recruits be taught to amend the
> raw machine-generated data with their own knowledge, like street names,
> road classification, surface...?
We are all too aware of the limitations of AI/ML-derived data and so we
plan to stress that AI is not infallible and to always use one's best
judgement. We will provide examples of false positives and false negatives
to show that the data still needs human judgement.
> [...] How will you ensure that you do not generate more contributions than
> you can ensure the quality for?
We have already browsed through the road-detections data provided by
Facebook and based on our assessment, most of the country's roads have
actually already been mapped. So we are pretty confident that the local
community can properly review/validate contributions and that we won't be
overwhelmed. We also plan to use the tasks-assisted instance of the HOT
Tasking Manager to coordinate mapping and validation. We have done a pilot
of mapping using RapiD with a small province as a test and we think this
process is feasible. In addition, we do not plan to initiate wholesale
mapping of the country with Facebook's AI-derived data. Our plan is to wait
for local mappers, local government units, or organizations to contact us
if they want to help complete the road network their local area. Only then
will we create tasks and provide trainings or workshops in case they need
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the talk