pcvlab/justzoomin

Just Zoom In Cross-View Geo-Localization Dataset

This dataset accompanies Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming. It supports cross-view geo-localization, where a ground-level street-view image is localized using geo-referenced overhead imagery. The dataset pairs crowd-sourced, limited-field-of-view street-view images with a multi-scale satellite/aerial tile hierarchy over Washington, D.C., making it suitable for realistic evaluation of both retrieval-based and coarse-to-fine localization methods.

The current release contains approximately 300k ground-level images over a 10 km × 10 km Washington, D.C. area, together with the corresponding multi-scale overhead tile hierarchy.

The data is stored as TAR shards to keep the Hugging Face repository manageable. After extraction, the directory layout matches the path-based PyTorch/tiledwebmaps loader used in the paper.

Links

• 📄 Paper: Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming
• 💻 GitHub: Official Repository
• 🌐 Project Website

Motivation

Most cross-view geo-localization benchmarks are built around single-shot retrieval from fixed satellite crops. This setup is useful, but it can underrepresent key challenges of real-world localization: street-view images may have unknown orientation, limited field of view, variable image quality, and visual cues that fall outside a single overhead crop. In addition, flat retrieval databases do not explicitly expose the geographic hierarchy of maps.

This benchmark is designed to address these limitations by combining realistic crowd-sourced street-view imagery with a multi-scale overhead tile structure. It enables evaluation of methods that reason from coarse geographic context to fine local detail, including autoregressive zooming models, hierarchical search methods, and standard retrieval baselines.

Data Sources and Licensing

This dataset contains data derived from two sources.

Component	Source	License
Street-view images	Mapillary	CC BY-SA 4.0
Street-view metadata derived from Mapillary imagery	Mapillary / dataset authors	CC BY-SA 4.0
Aerial orthophotography / satellite tile imagery	Open Data DC / Government of the District of Columbia	CC BY 4.0
Split files and derived benchmark metadata	Dataset authors	CC BY-SA 4.0

Because the repository contains Mapillary-derived imagery, the dataset is distributed under CC BY-SA 4.0.

Required source attributions:

Street-view imagery derived from Mapillary, licensed under CC BY-SA 4.0.
Aerial orthophotography derived from Open Data DC / Government of the District of Columbia, licensed under CC BY 4.0.

License references:

• Mapillary open imagery license: https://help.mapillary.com/hc/en-us/articles/115001770409-CC-BY-SA-license-for-open-data
• CC BY-SA 4.0: https://creativecommons.org/licenses/by-sa/4.0/
• Open Data DC: https://opendata.dc.gov/
• CC BY 4.0: https://creativecommons.org/licenses/by/4.0/

Users are responsible for complying with the licenses of the underlying data sources.

Repository Layout

.
├── README.md
├── metadata/
│   ├── large_area_train_map.csv
│   └── large_area_val_map.csv
├── streetview/
│   ├── metadata_satellite_covered.csv
│   └── metadata_satellite_covered.parquet
├── satellite/
│   └── layout.yaml
└── archives/
    ├── streetview_images_000.tar
    ├── streetview_images_001.tar
    ├── streetview_images_002.tar
    ├── streetview_images_003.tar
    ├── satellite_level_0_000.tar
    ├── satellite_level_0_001.tar
    ├── ...
    ├── satellite_level_m1_000.tar
    ├── satellite_level_m2_000.tar
    └── ...

The archive names use m1, m2, etc. for negative satellite levels. For example, satellite_level_m1_000.tar contains files under satellite/-1/.

After extraction, the data has this structure:

extracted/
├── streetview/
│   └── images/
│       ├── <image_id>_undistorted.jpg
│       └── ...
└── satellite/
    ├── layout.yaml
    ├── 0/
    ├── -1/
    ├── -2/
    ├── ...
    └── -9/

Metadata

The split CSV files contain the fields used by the paper's training/evaluation loader.

Field	Description
image_id	Ground-level image identifier
sequence	Ground-truth zoom-action sequence
latitude	Ground-truth latitude
longitude	Ground-truth longitude

Ground images are named:

<image_id>_undistorted.jpg

Download and Extract

from pathlib import Path
import shutil
import tarfile

from huggingface_hub import snapshot_download

repo_dir = Path(snapshot_download(
    repo_id="pcvlab/justzoomin",
    repo_type="dataset",
))

extract_dir = Path("./justzoomin_data")
extract_dir.mkdir(parents=True, exist_ok=True)

for tar_path in sorted((repo_dir / "archives").glob("*.tar")):
    print(f"Extracting {tar_path.name}")
    with tarfile.open(tar_path, "r") as tar:
        tar.extractall(extract_dir)

# Place the tiledwebmaps layout file beside the extracted satellite folders.
(extract_dir / "satellite").mkdir(exist_ok=True)
shutil.copy2(
    repo_dir / "satellite" / "layout.yaml",
    extract_dir / "satellite" / "layout.yaml",
)

Loader Paths

After extraction, use the following paths in the paper's dataset loader:

metadata_csv = repo_dir / "metadata" / "large_area_train_map.csv"
ground_root = extract_dir / "streetview" / "images"
tile_layout = extract_dir / "satellite" / "layout.yaml"

For validation:

metadata_csv = repo_dir / "metadata" / "large_area_val_map.csv"

The satellite tile folders are loaded through satellite/layout.yaml using tiledwebmaps.

Samples from the Dataset

Each sample consists of a ground-level street-view image, its latitude/longitude, and a ground-truth zoom-action sequence over the satellite tile hierarchy. The standard task is to predict the correct overhead cell, or equivalently the zoom-action sequence, for a given street-view query. However, users are not limited to use discrete sequences, since dataset supports many different zoom levels.

Dataset Notes

The benchmark uses limited-field-of-view, non-panoramic, crowd-sourced street-view images and a multi-scale overhead tile hierarchy. It is intended for within-area cross-view geo-localization over the Washington, D.C. region.

The data is not stored as loose image files on the Hub. The TAR shards should be extracted before using the original path-based loader.

Citation

If you use this dataset, cite the paper:

@article{erzurumlu2026justzoomin,
  title={Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming},
  author={Erzurumlu, Yunus Talha and Kwag, Jiyong and Yilmaz, Alper},
  journal={arXiv preprint arXiv:2603.25686},
  year={2026}
}

Please also attribute the underlying data sources:

Street-view imagery: Mapillary, CC BY-SA 4.0.
Aerial orthophotography: Open Data DC / Government of the District of Columbia, CC BY 4.0.