Same grid, five feature types. Each number is a per-cell total across the whole city. Ratio is OSM ÷ Overture, and r shows whether OSM and Overture agree on where each feature type is dense, even if not on how many there are.
| Theme | OSM total | Overture total | Ratio (OSM/Overture) | Correlation (r) |
|---|
Matching counts don't prove the buildings are in the right place. This check pulls real OSM building outlines for a sample of cells and matches each one against Overture's footprints using intersection-over-union (IoU), so it's checking whether the same building exists in both datasets, not just whether the totals line up. It still only shows whether OSM and Overture agree with each other, not which one is correct. Turn on satellite imagery above and zoom into a low-match cell below to judge that for yourself.
| Cell (row, col) | OSM buildings | Overture buildings | OSM match rate | Overture match rate | Mean IoU |
|---|
Every grid cell for the theme selected above. Click a column header to sort.
| Cell | OSM | Overture | Diff (OSM − Overture) | Ratio (OSM/Overture) |
|---|
Grid: the city is split into ~1.8 km cells (see
data/full/osm_grid_counts.py). OSM counts come from live
Overpass count queries per cell. Overture counts come from
the downloaded theme files, bucketed by feature centroid into the same
grid.
Schema mismatch is real: OSM and Overture don't tag things
the same way, so these mappings are best-effort:
road = OSM highway=* vs Overture's road segments;
rail = railway=* vs rail segments; water =
natural=water + waterway=* vs Overture's
water theme; landuse = landuse=* +
leisure=* vs Overture's land_use theme (which folds in
parks that OSM usually tags as leisure, not
landuse).
Roads and rail counts are especially rough: Overpass counts any way with a node in the cell, so one long OSM way crossing several cells gets counted in each one. Overture's segments are pre-split at intersections and much shorter, so this isn't an apples-to-apples count. Treat the road and rail correlation as a rough direction, not a precise number.
Geometric accuracy is sampled, not exhaustive: pulling full building outlines for all 642k+ Overture and ~310k OSM buildings city-wide would be a much bigger, slower job against the shared public Overpass server. So the accuracy section above samples a spread of cells across the density range instead.