Remix.run Logo
DoctorOetker 12 hours ago

I've always been confused why OSM doesn't design a cheap open source stereo imager, some finetuned modern YOLO neural network could detect and blur the glass parts of windows, doorbells, nameplates and humans in the scene.

It could use GPS for rough geolocation, and 3D models of all the scenery could be generated. New contribution traces would contain changes compared to the past. Volunteers could request new paths it would like to see explored, and OSM could propose paths that cut through or ride along segments of recently submitted recordings of other users, to check if those changes are real, without checking the whole suspected recording.

Separate ground truth recording from its interpretation into mappable concepts, going out to make a recording or observation is a different task from deciding how to canonicalize the content.

ponorin 13 minutes ago | parent | next [-]

OpenStreetMap is a 2D map with limited 3D features, not a streetview alternative. For crowdsourced street view there are services like Mapillary (now owned by Meta) and Panoramax, both of which are cleared for use in mapping in OSM (and in case of Panoramax which is self-hostable, often run by a local OSM community).

nullisland00 12 hours ago | parent | prev | next [-]

It’s a volunteer org. Sounds like you’re volunteering to build that and pay for the imagery etc. Let us know when you’re done.

PetitPrince 6 hours ago | parent | prev | next [-]

Are you describing what Mapillary is doing ? It's a crowdsourced street level imagery provider (owned by Facebook since 2020, but still contributing to OSM), but they do auto-detection of feature (you can see that in the online OSM editor: Map Data > Photo Overlays > Mapillary > Map Features).

I vaguely remember a blog post about doing photogrammetry out of their images, but I'm not entirely sure about it.

DoctorOetker 5 hours ago | parent [-]

I was talking about photogrammetry, preferably with 2 global shutter synchronous image sensors, attached to the ends of say a 1 m pole.

namibj 30 minutes ago | parent [-]

Don't need global shutter just use a bit more math and IMU data with frequency in excess of the frame rate.

Make it work with the recording capabilities of a good modern phone and handle the compute in non-real-time.

Rygian 4 hours ago | parent | prev | next [-]

Have you suggested this to the folks at Panoramax? https://panoramax.fr/

They're already covering the blurring, as well as detecting useful objects in imagery (traffic signs for a start).

nemomarx 11 hours ago | parent | prev [-]

Well, first step here is to demonstrate that you can do that neural network part. I think that would be a really valuable tool for more projects than openstreetmaps - google might even want to buy it.