atomizer-text

OSM semantic overlay pilot for the Atomizer per-pixel token model.

This bucket is the canonical, read-only data store for the atomizer-text v1 pilot. It contains the raw source imagery, the metadata, the 3 extracted representative tiles, and the 9 CLI outputs. It is not a model — it is reproducible input + output data for a small Python tool that lives in the GitHub repo below.

---

What this pilot does (one-paragraph version)

For a satellite / aerial photo, the pilot queries OpenStreetMap for all buildings, roads, landuse polygons, etc. inside the photo's bounding box, maps each feature's raw OSM tags onto one of 10 fixed semantic categories (residential / office / commercial / industrial / other building, road, agricultural land, vegetation or park, water, unknown), paints those categories onto the photo's pixel grid (with a fixed priority that lets buildings win over landuse on overlaps), and writes a 3-panel figure, a JSON of summary statistics, and a per-feature record of raw OSM tags for a future text-embedding step.

The v1 plan, taxonomy rules, rasterization priority, and per-tile feasibility results are documented in the GitHub README.

---

Layout of this bucket

data/
├── D034-2021_AERIAL_RGBI.zip            4.77 GB  source aerial patches, D034
└── GLOBAL_ALL_MTD.zip                   2.11 GB  source metadata, all departments

sample_tiles/
├── residential.tif                      1.2 MB   Castelnau-le-Lez (suburban houses)
├── office.tif                           1.1 MB   Port Marianne / Antigone (towers)
└── mixed.tif                            1.0 MB   peri-urban fringe SW of center

results/
├── residential/
│   ├── overlay.png                      3-panel figure (RGB / category / RGB + outlines)
│   ├── stats.json                       coverage, category distribution, tag vocabulary
│   └── tag_records.json                 per-feature records for future embedding
├── office/                              (same 3 files)
└── mixed/                               (same 3 files)

Everything under data/ is the unaltered upstream data from IGNF/FLAIR-HUB (the zips are mirrored from the data/ directory of that dataset). The two top-level zips are mirrored here so the pilot can be reproduced without re-downloading from the original sources.

sample_tiles/ and results/ are produced by the v1 pipeline; their content is fully derivable from data/ + the code in the GitHub repo.

---

Per-tile feasibility summary

Headline: the residential-vs-office contrast is not present in OSM at 0.2 m over Montpellier — every building in all three tiles is tagged building=yes (a generic "yes, this is a building" stamp with no type information). The discriminative signal in the office tile is its road network and amenity tags, not its building types. See the GitHub README's "Results & interpretation" section for the full feasibility write-up.

---

Reproducing these results from scratch

You need: Python 3.12+, uv, the huggingface-cli (hf command), and outbound HTTPS access to Overpass (one-shot).

# 1. Clone the code.
git clone https://github.com/NoeFlandre/atomizer-text.git
cd atomizer-text

# 2. Install dependencies (creates .venv).
uv sync

# 3. Mirror the bucket locally so the source zips are present
#    without hitting the public IGNF/FLAIR-HUB dataset.
hf sync hf://buckets/NoeFlandre/atomizer-text ./bucket

# 4. Re-extract the 3 sample tiles into data/sample_tiles/. The
#    script downloads the zips to a temp dir under data/, picks
#    the 3 patches nearest to the LANDMARKS in the script, and
#    cleans up the temp dir.
ATOMIZER_BUCKET=buckets/NoeFlandre/atomizer-text \
    uv run python scripts/fetch_sample_tiles.py

# 5. Run the CLI on each tile. The default Overpass fetcher hits
#    the public API (~10 s per tile on the first run).
for tile in residential office mixed; do
    uv run python -m atomizer_osm_pilot.cli \
        --raster data/sample_tiles/$tile.tif \
        --out-dir results/$tile \
        --osm-tags building,highway,landuse,shop,office,amenity,natural,waterway,leisure
done

After step 5, results/{residential,office,mixed}/ should be byte-identical (up to PNG encoder nondeterminism in matplotlib) to the files in this bucket.

For a fully offline, no-Overpass smoke test, use the included scripts/offline_fetcher.py:

uv run python -m atomizer_osm_pilot.cli \
    --raster tests/fixtures/tiny.tif \
    --out-dir /tmp/atomizer_offline \
    --fetcher scripts.offline_fetcher:offline_fetcher

---

Tests (offline, no Overpass)

uv run pytest

Expected: 38 passed, 1 skipped (the real-Overpass test, marked @pytest.mark.network).

---

License and attribution

• The two top-level zips (data/D034-2021_AERIAL_RGBI.zip and data/GLOBAL_ALL_MTD.zip) are mirrored from IGNF/FLAIR-HUB. See that dataset card for the upstream license (BD ORTHO imagery comes from IGN's open-license policy; the metadata is published by IGN under an open license).
• The sample_tiles/ and results/ directories are produced by this pilot; they are released under the MIT license (same as the GitHub repo).
• OpenStreetMap data queried at runtime by the CLI is (c) OpenStreetMap contributors, licensed under the ODbL.

Citation

If you use this dataset, please cite the upstream FlairHUB paper:

@article{GARIOUD2026271,
  title  = {FLAIR-HUB: Large-scale multimodal dataset for land cover and crop mapping},
  author = {Garioud, Anatol and Giordano, S{\'e}bastien and David, Nicolas and Gonthier, Nicolas},
  journal= {ISPRS Journal of Photogrammetry and Remote Sensing},
  volume = {237},
  pages  = {271--300},
  year   = {2026},
  doi    = {10.1016/j.isprsjprs.2026.04.017}
}

And the Atomizer paper the pilot is designed to inform:

@misc{atomizer2025,
  title  = {Atomizer: a Foundational Model for Remote Sensing},
  author = {anonymous},
  year   = {2025},
  eprint = {2506.13542},
  archivePrefix = {arXiv}
}