Open betatim opened 7 years ago
Another approach could be to use the rendered OSM tiles and and do pattern matching to decide if they are "empty" or not.
A further advantage of using QA tiles or OSM tiles is that we are not limited to training data that has been mapswiped.
https://github.com/tilezen/mapbox-vector-tile for reading QA tiles in python
Currently we use a majority vote of mapswipe users to decide the ground-truth of a tile.
From looking at some tiles which we classify as "boring" in the current setup I noticed a few contain things that should be mapped. The mapswipe user votes might also be noisy.
How to improve the ground truth?
One idea is to use OSM QA tiles. If the corresponding QA tile contains something this is a tile that is "interesting", if not we assume it is empty because it is boring (not because it hasn't been mapped yet).
Next step: download a country extract of QA tiles and see how easy it is to determine if a tile is "empty"