README.md

---
license: other
license_name: loci-dataset-terms
pretty_name: LOCI Dataset
language:
  - en
task_categories:
  - image-feature-extraction
tags:
  - cross-view-localization
  - geolocalization
  - robotics
  - openstreetmap
  - satellite-imagery
extra_gated_prompt: >-
  By requesting access you agree to the LOCI Dataset Terms of Use
  (the LICENSE file in this repository), including its per-component licenses
  (ODbL 1.0 for OpenStreetMap-derived tables, CC BY-SA 4.0 for street-level
  imagery and derived annotations, CC BY 4.0 for checkpoints and evaluation
  artifacts, public-record terms for MassGIS/CT imagery) and its privacy
  clause: you will not use the Dataset in any manner to identify or invade
  the privacy of any person whose personal data may have been incidentally
  collected.
---

# LOCI Dataset

Dataset release for *Leveraging Semantic Maps for City-Scale Cross-View
Localization* ([arXiv:2607.25215](https://arxiv.org/abs/2607.25215)).
Code, docs, and reproduction guides live in the code repository linked
from the paper.

Every file is
listed with its sha256 in `RELEASE_MANIFEST.sha256`.

## Overview

| Folder | What's inside | Size |
|---|---|---|
| `satellite/` | Redistributable satellite imagery (public-record sources), per-city tars + pinned patch manifests | 9.2 GB |
| `panoramas/` | Street-level panoramas for the 8 trajectory environments (self-collected + Mapillary) | 22 GB |
| `osm_landmarks/` | OSM landmark tables, the exact per-environment versions used in the paper | 254 MB |
| `annotations/` | VLM-extracted panorama landmark annotations + embeddings, 11 environments | 4.0 GB |
| `correspondence/` | VLM-labeled correspondence dataset (Chicago/Seattle) + shared text-value embeddings | 0.9 GB |
| `checkpoints/` | Trained model weights (DINOv3 backbone stripped — see reassembly below) | 0.5 GB |
| `final_results/` | The paper's evaluation outputs | 2.0 GB |
| `evaluation_paths/` | The exact evaluation path files (pano-id sequences) | 115 MB |
| `verification/` | Similarity-matrix fingerprints + patch manifests for imagery that is not redistributed | 437 MB |
| `osm_baseline/` | Baked vector map tiles behind the WAG+OSM baseline | 58 MB |
| `human_eval_labels/` | Raw human labels behind the paper's Table IV | 3 MB |

## satellite/

| File | Contents | Size |
|---|---|---|
| `framingham.tar` | 40,401 patches, 640×640 (MassGIS 2025 aerial, 15 cm), VIGOR grid | 4.3 GB |
| `middletown.tar` | 39,601 patches, 640×640 (CT 2023 orthoimagery, 7.6 cm) | 4.9 GB |
| `*.patch_manifest.json` | Per-patch pixel sha256 + pinned source/grid | 10 MB ea. |

Only public-record imagery is redistributed here. Satellite imagery for
the other environments (Esri World Imagery, VIGOR) ships as pinned patch
manifests under `verification/` instead — see below.

## panoramas/

| File | Contents | Size |
|---|---|---|
| `boston_snowy.tar` | 1,674 GPS-stamped panoramas, 2048×1024 (as evaluated in the paper), collected during an active snowstorm | 0.6 GB |
| `boston_night.tar` | 1,378 panoramas, 7680×3840, night collect of the same route | 5.9 GB |
| `framingham.tar` | 478 panoramas, 4096×2048 (Mapillary) | 1.3 GB |
| `middletown.tar` | 263 panoramas, 5640×2820 (Mapillary, rain) | 0.8 GB |
| `san_francisco_mapillary.tar` | 300 panoramas, 5760×2880 (Mapillary) | 1.0 GB |
| `fort_myers.tar` | 1,073 panoramas, 12288×6144 (Mapillary, post-Hurricane-Ian) | 6.8 GB |
| `noordoostpolder.tar` | 1,916 panoramas, 4096×2048 (Mapillary) | 1.9 GB |
| `veluwe.tar` | 3,654 panoramas, 4096×2048 (Mapillary) | 4.6 GB |

Each tar contains `<city>/panorama/*.jpg` (filenames embed
`{pano_id},{lat},{lon}`) plus `pano_id_mapping.csv` and the collection
pipeline's provenance files (`extraction_log.csv`,
`pipeline_metadata.json`) where applicable. Panoramas for the VIGOR
train/eval cities (`chicago`, `seattle`, `new_york`) are not
redistributed — obtain the VIGOR dataset separately.

## osm_landmarks/

`<city>/<version>.feather` — one dated Geofabrik-derived OSM landmark
table per environment (ODbL), the exact versions used in the paper.
`boston.feather` appears under both Boston environments (shared table).

## annotations/

`<city>.tar` (11 environments) — VLM-extracted panorama landmark
annotations: Gemini batch outputs
(`sentences/results/**/predictions.jsonl`, image payloads replaced with
`sha256:` markers) plus `embeddings/embeddings.pkl` per environment.
Records are keyed by panorama stem (`{id},{lat},{lon},`) and correspond
1:1 with the released panoramas.

## correspondence/

| File | Contents | Size |
|---|---|---|
| `labels.tar` | VLM-labeled correspondence pairs (text-only Gemini batch outputs) | 221 MB |
| `text_value_embeddings.pkl` | text-embedding-005 value embeddings (768-d, 206,277 entries) serving classifier training, matrix export, and LOCI-EF training; covers every text value in the released landmark tables and annotations | 645 MB |

See `correspondence/README.md` for details.

## checkpoints/

| Model | Contents | Size |
|---|---|---|
| `wag/`, `wag_plus_osm/` | Trained weights per side (`best_panorama/`, `best_satellite/` — the layout `--checkpoint best` expects), **DINOv3 backbone stripped**, + `weights_manifest.json` (per-tensor sha256 incl. the backbone), a VIGOR-free `input_output.tar` consistency anchor, and `train_config.yaml` | ~72 MB ea. |
| `loci_ef/` | Same layout; transformer + tag-bundle encoders + frozen SAFA (backbone stripped) | ~370 MB |
| `correspondence_classifier/` | `best_model.pt` (674 KB, complete — no backbone) + `config.yaml` | <1 MB |

**Reassembly**: the released weights omit the frozen DINOv3 backbone
(obtained from Meta under the DINOv3 License) and the pickled `model.pt`.
Run `bazel run //tools:reassemble_checkpoints -- --data-root <this
download>` from the code release: it rebuilds each model from
`train_config.yaml` (torch.hub downloads DINOv3), verifies every tensor —
including the downloaded backbone — against `weights_manifest.json`, and
writes `model.pt` + full `model_weights.pt` in place. Released configs use
a literal `{DATA_ROOT}` placeholder resolved by the tool. After reassembly
the released training/eval entry points load these checkpoints unchanged.

## final_results/

`{wag,wag_plus_osm,loci,loci_ef}.tar` (0.53 GB each) — the paper's
evaluation outputs: per-path `{error, mode_error, prob_mass_by_radius,
path, var, distance_traveled_m}.pt` for 1,000 paths per environment
(5,000 for New York and Seattle) across all 10 evaluation environments
plus Seattle (calibration), with per-environment
`summary_statistics.json` and the exact eval configs (`args.json`,
`aggregator_config.yaml`). Every `average_final_error` in the shipped
`summary_statistics.json` matches Table V of the paper exactly.

`sigma_calibrations/` holds the four Seattle sigma-calibration fits
(JSON + args) used by the eval configs.

The environment set includes `framingham_mixed_sat` — the paper's
"Framingham Mixed-Sat" row. That row's satellite imagery (a Google mosaic)
can not be redistributed. All other environments' matrices are regenerable
from the released checkpoints and imagery. Code is provided to collect the current google mosaic of Framingham, but this may drift over time.

## evaluation_paths/

`<env>.json` — the paper's evaluation path files: 1,000 3-km paths per
trajectory environment; 5,000 5-km goal-directed paths for New York and
Seattle. Paths are pano-id sequences; the 8 trajectory files reference
exactly the released panoramas (the Seattle/New York files reference
VIGOR pano ids).

## verification/

`fingerprints/<env>__<matrix>.fingerprint.npz` — numerical fingerprints
for every regenerable similarity matrix (10 environments × 4 matrices):
quantiles, seeded samples, random-projection sketch, per-row top-k.
Check regenerated matrices with the code release's
`tools/verify_artifacts.py matrix` (tolerance/rank-based — GPU float
nondeterminism means bit-exact comparison is not expected).

`patch_manifests/<env>.patch_manifest.json` — per-patch decoded-pixel
sha256 + pinned source/grid for the satellite imagery that is **not**
redistributed: Esri-pinned manifests (`boston_snowy`/`boston_night`,
`fort_myers`, `noordoostpolder`, `veluwe`) drive bit-exact re-download via
`download_tiles.py --manifest`; foreign-source manifests (`chicago`,
`new_york`, `seattle`, `san_francisco_mapillary`) verify a user's own
VIGOR copy.

`boston_snowy` and `boston_night` share the same satellite imagery (their
manifests are identical) — download once and reuse. `framingham` /
`middletown` canonical patch manifests live with their imagery under
`satellite/`. `framingham_mixed_sat` has no fingerprints (see
`final_results/` above).

## osm_baseline/

`<region>.mbtiles` — baked vector tiles (planetiler, OpenMapTiles schema)
behind the WAG+OSM baseline: the exact paper-era bakes from the pinned
dated Geofabrik dumps, one per environment plus `illinois` / `washington`
/ `new_york` for the VIGOR train/eval cities. Rasterize into
`<city>/satellite_osm/` with the code release's `render_osm_tiles.py`;
`boston_night` reuses `boston_snowy`'s renders (same grid).

## human_eval_labels/

`annotations/` — per-rater judgments (`yes`/`minor`/`no`) on 500
VLM-extracted panorama annotations per city (Chicago/Seattle) plus
third-pass consensus for the 87 disagreements. `correspondence/` — 1,000
blind same/different judgments per rater per city on proposed
correspondence pairs plus the full 2,000-pair consensus (112
adjudicated).

Table IV is reproducible from these files digit-for-digit: annotation
rows (n=500 per city: Chicago 83.6/4.2/12.2 κ 0.70, Seattle 78.2/5.0/16.8
κ 0.75) and correspondence metrics (κ 0.88/0.90; P/R/F1 0.976/0.859/0.914
Chicago, 0.976/0.850/0.909 Seattle).

## Licensing

Composite — per-component licenses are listed in
[LICENSE](LICENSE): ODbL 1.0 (OpenStreetMap-derived tables
and the correspondence dataset), CC BY-SA 4.0 (street-level imagery and
VLM annotations), CC BY 4.0 (checkpoints and evaluation artifacts), and
public-record terms for the MassGIS / CT ECO imagery. Esri World Imagery
satellite tiles and VIGOR imagery are **not redistributed**; pinned patch
manifests allow re-download.