MVRL/dense500-retrieval-instances

Dense500 retrieval split — Instance Annotations

Polygon-level object annotations anchored to the tiles of MVRL/dense500-retrieval: building footprints and OpenStreetMap area features, each with structured OSM tags and projected into the pixel frame of its deepest containing tile. These are the queries of the GeoDIVE object-to-tile retrieval benchmark (WACV 2026). This release holds annotations only; the imagery is in MVRL/dense500-retrieval.

Construction

Microsoft Global Building Footprints and OpenStreetMap polygons were clipped to each site and merged where both describe the same object. Each polygon is anchored to the deepest tile that fully contains it, with its vertices projected into that tile's 256×256 frame. OSM tags are stored as ids into a shared vocabulary. Captions were generated with a vision-language model from a padded crop of each object.

Files

File	Content
inst_metadata.npz	Instance arrays and a per-tile reverse index
tag_vocab.pt	{(key, value): id} for OSM tags
tag_pairs.parquet	The same vocabulary as a parquet table
site_index.parquet	region_id <-> site_id
site_registry.parquet	Per-site metadata
all_tiles.parquet	Tile inventory
captions.parquet	Captions, prompt v7
captions_v8.parquet	Captions, prompt v8; the GeoDIVE text queries
caption_sample.json	Instance ids of the captioned sample
scripts/load_dataset_example.py	Render-one-instance demo

Schema (inst_metadata.npz)

Indexed by instance id i:

source[i]                  uint8     0=ms_only  1=ms+osm_merged  2=osm_building  3=osm_area
geom_type[i]               uint8     1=polygon
multipolygon_group_id[i]   int32     -1 if singleton; siblings share a value
anchor_zoom[i]             int8      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   building height (m); NaN if unknown
ms_confidence[i]           float32   footprint 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
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    "region_zoom_x_y"
tile_inst_start[t]         int64   first instance anchored at this tile
tile_inst_end[t]           int64   one past the last

Captions

Both caption files share the columns inst_local_id, anchor_tile_id, source, anchor_zoom, pixel_area_at_anchor, region_id, sample_label, tags_str, caption; join to instances on inst_local_id. Prompt v7 describes only what is visible in the crop. Prompt v8 also gives the model the instance's OSM tags, a land-cover summary, and a settlement class, so a caption can name a function that is not visible in the pixels; the tags themselves are never the query.

How to use

from huggingface_hub import snapshot_download
snapshot_download("MVRL/dense500-retrieval-instances", repo_type="dataset", local_dir="./dense500-retrieval-instances")

from datasets import load_dataset
imagery = load_dataset("MVRL/dense500-retrieval")

Join annotations to imagery on the tile key "{zoom}_{x}_{y}" (the imagery location column). See scripts/load_dataset_example.py.

Citation

@inproceedings{iqbal2026geodive,
  author    = {Iqbal, Hamza and Cher, Dan and Xing, Eric and Kolouju, Pranavi and Stylianou, Abby and Pless, Robert and Jacobs, Nathan},
  title     = {GeoDIVE: Geospatial Descent Informed by Value Estimation},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2026}
}

@inproceedings{khanal2026genesis,
  author    = {Khanal, Subash and Cui, Yangzhi and Cher, Daniel and Xing, Eric and Wei, Brian and Sastry, Srikumar and Jacobs, Nathan},
  title     = {Genesis: A Generative Engine for Hierarchical Satellite Image Synthesis},
  booktitle = {ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (ACM SIGSPATIAL)},
  year      = {2026},
  doi       = {10.1145/3841645.3843313}
}