dcher95/global-dense-instances

Global Dense Satellite — Instance Annotations (v1.0)

932,427 polygon-level object annotations across 275 hierarchical sites, anchored to global-dense-satellite satellite tiles at zoom levels 10-18. Each instance is a polygon (building footprint or OSM area feature) with structured OSM tags, projected into the pixel frame of its deepest containing tile.

This is an annotations-only release — the satellite imagery is in dcher95/global-dense-satellite and must be downloaded separately.

What's in the box

File	Content
inst_metadata.npz	All instance arrays + per-tile reverse index (mmap-friendly)
tag_vocab.pt	{(key, value): int_id} dict, 4,828 OSM tag pairs
tag_pairs.parquet	Same vocab as a queryable parquet
site_index.parquet	region_id <-> site_id map (275 sites)
site_registry.parquet	Per-site metadata: lon/lat, tag, continent, zoom range, bbox, country
all_tiles.parquet	Tile inventory: tile_id, zoom, x, y, site_id, is_root/is_leaf, hf_indices
scripts/load_dataset_example.py	End-to-end render-one-instance demo

Quick stats

• 932,427 polygon instances
• 12,384 unique tiles with annotations
• 6,271,936 polygon vertices
• 1,196,954 tag-ID assignments (CSR form)
• 4,828 unique OSM (key, value) tag pairs
• 275 hierarchical sites, globally distributed
• Source mix: 456,890 ms_only, 63,179 ms+osm_merged, 365,964 osm_building, 46,394 osm_area

Schema (inst_metadata.npz)

Indexed by global instance id i ∈ [0, N):

source[i]                  uint8     0=ms_only  1=ms+osm_merged  2=osm_building  3=osm_area
geom_type[i]               uint8     1=polygon (always)
multipolygon_group_id[i]   int32     -1 if singleton; siblings share a value
anchor_zoom[i]             int8      10..18 — deepest tile fully containing the polygon
anchor_x[i], anchor_y[i]   int32     tile coords at anchor_zoom
pixel_area_at_anchor[i]    int32     polygon area in px on the 256x256 anchor tile
height[i]                  float32   MS-supplied building height (m); NaN if unknown
ms_confidence[i]           float32   MS source confidence; NaN if unknown
point_px[i, :]             float32   (2,) centroid in anchor-tile pixel coords
bbox_px[i, :]              float32   (4,) xmin, ymin, xmax, ymax in pixel coords
region_id[i]               int32     index into site_index.parquet

CSR variable-length:
verts_px[verts_ptr[i]:verts_ptr[i+1], :]   float32 (k, 2)  polygon exterior ring (pixel coords)
tag_ids[tag_ids_ptr[i]:tag_ids_ptr[i+1]]   int32           OSM tag IDs (decode via tag_vocab.pt)

Per-tile reverse index:
tile_ids[t]                <U40    e.g. "0007_17_106504_55302"  (region_id_zoom_x_y)
tile_inst_start[t]         int64   first instance idx anchored at this tile
tile_inst_end[t]           int64   one-past-last instance idx

How to use

from huggingface_hub import snapshot_download
snapshot_download("dcher95/global-dense-instances", repo_type="dataset", local_dir="./global-dense-instances")

from datasets import load_dataset
imagery = load_dataset("dcher95/global-dense-satellite")

Join annotations to imagery on the tile location string "{zoom}_{x}_{y}" (the imagery dataset's location column). See scripts/load_dataset_example.py for a full demo.

Pipeline

1. Site registry: each site's geographic bbox + country, from the tile hierarchy.
2. OSM polygons: per-site extraction from Geofabrik country extracts.
3. MS buildings: per-site Microsoft Global Building Footprints.
4. Instance assembly: MS footprints unioned with OSM polygons. OSM with IoU >= 0.5 OR (cont_osm >= 0.85 AND cont_ms >= 0.5) merge their tags onto the MS row. Unmatched OSM polygons remain as osm_building / osm_area instances.
5. Anchor zoom: deepest tile fully containing each polygon; project vertices to its 256x256 frame.
6. Tag vocab + CSR pack + global merge.

Source licenses

Derivative database under ODbL-1.0, inheriting OpenStreetMap share-alike.

Component	License	Attribution
Imagery (separate)	see dcher95/global-dense-satellite	per that dataset
OpenStreetMap tags	ODbL-1.0	"© OpenStreetMap contributors"
MS Global Building Footprints	ODbL-1.0	Microsoft + OpenStreetMap

Citation

@misc{global_dense_instances_2026,
  author = {Cher, Daniel},
  title  = {Global Dense Satellite — Instance Annotations (v1.0)},
  year   = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/dcher95/global-dense-instances}},
}