They apply CNNs to label pixels. The result could be used to improve other computer vision tasks. Applied to maps they think it would be good to measure vegetation density or the availability and sidewalks.

Hmm... okay.? I dunno what I expected.

>Here is a short list of areas where semantic segmentation is available: Graz, Paris, San Francisco, Hong Kong, Dar es Salaam, and Mexico City.

If the method is general, why only these areas?

I'm guessing limited computing power while refining the algorithms.

It's because we need to upgrade and optimize some of the backend storage layout in order to cope with the data produced.

(I'm part of the Mapillary team)

Thanks for elucidating - makes sense.