search

ngageoint / hootenanny

Hootenanny conflates multiple maps into a single seamless map.
GNU General Public License v3.0
359 stars 74 forks source link

Conflate Mapillary point data with road ways #1871

Open brianhatchl opened 7 years ago

brianhatchl commented 7 years ago

Mapillary's computer vision algorithms extract metadata like road surface type and street sign speed limits from their street level imagery.

Is it possible for Hoot to conflate this point data (which is registered at points along the road geometry) with the neighboring way to add surface type and speed limit tags?

@amahon I opened this ticket to flesh out requirements and investigate feasibility with the team.

drew-bower commented 7 years ago

If so that would be useful.

amahon commented 7 years ago

@brianhatchl - thanks for kicking this off!

Regarding conflating object detections (points) to road segments, we've actually started some work on this front. While I'm not too familiar with that work myself, I will follow up with the folks who are involved, and report back. There may be some work and concepts that can be folded into Hoot.

The other interesting challenge is conflating object detections to non-linear geometries that may not be colocated with the detection. I'm thinking along the lines of conflating a detection of a brick building, observed from a roadway, with a building geometry that is along the roadway such that we can apply a descriptive attribute to that building geometry. Does this make sense?

In the meanwhile, how can I help this move forward? I'll follow up here once I've done a bit of digging on my end.

brianhatchl commented 7 years ago

@amahon I think if you could provide a sample of object detection data, that could get our team thinking in a more specific fashion.

For instance, is the bearing of the source camera included? I can see that being helpful / necessary to determine a proper road way match for signage. I can see your building scenario also benefiting from bearing data, with ray tracing being performed on detections like "brick building" and "stone wall" to search for possible conflation match candidates in the osm data.

With the cool new Mapillary features I saw, like georeferenced feature extraction from the street level imagery (lamp post demo), your AI may even pull a distance estimate that along with bearing would provide an absolute georeference estimate for object detections. In that case, the object detections could become a point-in-polygon conflation that we already support.

I'm looking forward to exploring this set of use cases more.