josefbednar/world-streetview-500k

🌍 World StreetView 500k

World StreetView 500k is a large-scale computer vision dataset for visual geolocation estimation, spatial representation learning, and geographic scene understanding.

It pairs ~2 million street-level images from ~500,000 unique locations worldwide with geographic coordinates, country labels, capture dates, and elevation data. Each training location is captured from 4 compass headings (0°, 90°, 180°, 270°) — ideal for training GeoGuessr-style geolocation models, geo-embeddings, or country classifiers.

⚠️ Usage Restriction: This dataset is provided strictly for non-commercial academic, scientific, and educational research (CC BY-NC 4.0). See Terms of Use below.

---

🚀 Quickstart

from datasets import load_dataset

# ⚡ Recommended: stream to avoid downloading the full ~198 GB
ds = load_dataset("josefbednar/world-streetview-500k", streaming=True)

sample = next(iter(ds["train"]))
print(sample["country_code"], sample["latitude"], sample["longitude"])
sample["image"]  # PIL image

For a full local copy:

ds = load_dataset("josefbednar/world-streetview-500k")  # ~198 GB download

print(ds)
# DatasetDict({
#     train: Dataset({num_rows: 1999060}),
#     val:   Dataset({num_rows: 3000})
# })

Note: The evaluation split is named val (not validation): access it via ds["val"].

---

📊 Dataset Overview

Total images	2,002,060 (1,999,060 train + 3,000 val)
Unique locations	~503,000 geographic coordinates
Geographic coverage	Global, sampled randomly across public street-level coverage
Download size	~198 GB
Collection tooling	streetview-scrape (open source)

📌 Split Breakdown

Split	Unique Locations	Images per Location	Total Images	Overlap with Train
train	~500,000	4 headings (0°, 90°, 180°, 270°)	1,999,060	—
val	3,000	1 heading	3,000	0% (strictly disjoint locations)

The val split contains only locations (and panoramas) never seen in train, making it suitable for honest held-out evaluation of geolocation accuracy.

🗂️ Data Schema

Column	Type	Description
image_id	string	Unique identifier for the individual image file
panoid	string	Unique source panorama identifier (shared by the 4 headings of one location)
image	image	Street-level image (decoded as a PIL image)
country_code	string	Two-letter ISO 3166-1 country code (e.g., US, SE, OM, BR)
date	string	Capture date, formatted YYYY-MM
latitude	float64	Latitude in decimal degrees (WGS 84)
longitude	float64	Longitude in decimal degrees (WGS 84)
elevation	float64	Elevation in meters above sea level

Tip: To group the 4 views of a single location, group rows by panoid. To build a country classifier, use country_code as the label; for coordinate regression or geocell classification, use latitude/longitude.

🎯 Intended Uses

• Image geolocalization — predicting coordinates or geocells from a single image
• Country / region classification — 200+ class geographic scene classification
• Geographic representation learning — contrastive or self-supervised pretraining on globally distributed imagery
• Visual question answering — geography-grounded VQA and multimodal evaluation
• GeoGuessr-style agents and benchmarks

⚠️ Biases & Limitations

• Geographic coverage bias: Sampling follows public street-level mapping availability. Densely mapped regions (North America, Europe, Japan, etc.) are over-represented relative to sparsely mapped regions (parts of Africa, Central Asia); some countries have no coverage at all.
• Road bias: Imagery is captured from mapped roads and paths, so scenes skew toward drivable/urbanized areas rather than a uniform sample of terrain.
• Temporal spread: Capture dates span many years and vary by country according to original mapping update cycles; visual features (vehicles, signage, vegetation) may reflect different eras.
• Camera generation artifacts: Image quality, resolution, and stitching characteristics vary with the capture hardware generation used in each region.
• Incidental content: Despite provider-side blurring of faces and license plates, street imagery may incidentally contain people, vehicles, and private property.

📜 Terms of Use & Legal Disclaimer

Click to expand legal terms and copyright info

1. Non-Commercial Academic Use Only

This dataset is compiled and provided solely for non-commercial academic research, educational instruction, and scientific evaluation, consistent with its CC BY-NC 4.0 license. Commercial exploitation, redistribution for profit, or integration into commercial software is strictly prohibited.

2. Copyright & Intellectual Property

All underlying imagery remains the copyright and intellectual property of the original map imagery providers (e.g., Google LLC and its licensors). This repository does not claim ownership over the underlying image content and provides access for academic research purposes.

3. Privacy & Content Removal Requests

If you are a copyright owner or an individual who wishes to request the removal of a specific image or location record due to privacy or copyright concerns:

• Open an issue on the GitHub repository, or
• Contact the maintainer via the Community tab on Hugging Face.

Requested records will be reviewed and promptly removed from public splits.

4. Limitation of Liability

THIS DATASET IS PROVIDED "AS IS" AND WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED. IN NO EVENT SHALL THE CREATORS OR DISTRIBUTORS BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY ARISING FROM THE USE OF THIS DATASET.

📖 Citation

If you use this dataset in your research, please cite:

@misc{world_streetview_500k,
  author       = {Bednar, Josef},
  title        = {World StreetView 500k Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/josefbednar/world-streetview-500k}}
}