Reubencf/streetview-global

StreetView Global

A globally-sampled street-view image dataset with rich scene annotations and visual question-answer pairs. All photographs are sourced from Mapillary, the open street-level imagery platform, via its public image API.

Each example pairs a street-level photograph with geographic metadata (latitude, longitude, compass, capture time, region), a free-form scene description, structured scene classification (setting, weather, time of day, road type, infrastructure), and five visual question-answer pairs covering area type, vehicles, signs/text, cultural/regional cues, and infrastructure.

Dataset Summary

• Total examples: 10,209
• Image source: Mapillary public image API
• Regions covered: 140 distinct regions spanning every inhabited continent
• Annotations: Scene description + structured tags + 5 VQA pairs per image
• Annotation model: Generated with Google Gemma 4 31B (google/gemma-4-31B-it) served via Together AI
• Format: HuggingFace datasets (Arrow), images stored inline

Features

Field	Type	Description
image	Image	Street-view photograph (JPEG)
id	string	Mapillary image id
latitude	float64	WGS84 latitude
longitude	float64	WGS84 longitude
compass_angle	float64	Camera heading in degrees (0 = north)
captured_at	string	Capture timestamp (ISO 8601)
is_pano	bool	Whether the source is a 360° panorama
width	int32	Source image width in pixels
height	int32	Source image height in pixels
region	string	Macro-region or Country - City label
sequence	string	Mapillary capture sequence id
scene_description	string	Natural-language description of the scene
setting	string	urban / suburban / rural / highway / forest / coastal / desert / mountain
weather	string	sunny / overcast / cloudy / rainy / snowy / foggy
time_of_day	string	day / dusk / dawn / night
road_type	string	paved / unpaved / cobblestone / dirt
infrastructure	string	Free-form infrastructure tag (e.g. minimal, moderate, well-maintained)
vqa	string	JSON-encoded list of 5 {question, answer} pairs

Example

from datasets import load_dataset
import json

ds = load_dataset("Reubencf/streetview-global", split="train")

ex = ds[0]
print(ex["region"], ex["setting"], ex["weather"])
print(ex["scene_description"])
for qa in json.loads(ex["vqa"]):
    print("Q:", qa["question"])
    print("A:", qa["answer"])
ex["image"].show()

Distributions

Macro-region coverage (top groupings)

Region	Count
Europe	2,245
North America	1,814
South America	830
Southeast Asia	780
East Asia	663
Africa	553
Middle East	442
South Asia	387
Oceania	325
Central America	323
Remote / uninhabited	285

Plus city-level samples from 129 additional Country - City buckets.

Setting

Setting	Count
urban	4,059
suburban	3,625
rural	1,429
highway	785
forest	132
coastal	60
desert	60
mountain	54

Weather

Weather	Count
sunny	6,069
overcast	3,025
cloudy	785
rainy	86
snowy	45
foggy	30

Time of day

Time	Count
day	9,797
dusk	199
night	185
dawn	23

Road type

Road type	Count
paved	9,392
unpaved	521
cobblestone	62
dirt	16

Intended Uses

• Visual question answering on real-world outdoor scenes
• Image captioning / scene description with geographic diversity
• Scene classification (setting, weather, time of day, road surface)
• Geo-aware vision-language model evaluation
• Street-view retrieval and clustering

Limitations and Biases

• Sampling is heavier in Europe and North America than in lower-income regions, reflecting Mapillary contributor distribution.
• The overwhelming majority of images are daytime and sunny/overcast; low-light, night, and adverse-weather scenes are under-represented.
• Scene descriptions, tags, and VQA pairs are model-generated with Google Gemma 4 31B (google/gemma-4-31B-it) served via Together AI, and may contain hallucinations, mislabeling, or biased cultural inferences — especially when inferring country from signage or architecture. Downstream users should treat them as weak supervision rather than ground truth.
• Some images contain incidentally-captured faces, license plates, or signage; Mapillary applies automated blurring, but coverage is not perfect.
• region labels mix coarse macro-regions with fine Country - City buckets depending on how the sample was collected.

Source and License

Images are sourced from the public Mapillary API and are licensed under CC BY-SA 4.0, the license under which Mapillary contributors release their photographs. You must attribute Mapillary and its contributors, and redistribute any derivative dataset under the same license.

The annotations (scene descriptions, scene classifications, and VQA pairs) in this repository are released under the same CC BY-SA 4.0 license.

Citation

If you use this dataset, please cite Mapillary as the image source:

@misc{mapillary,
  title        = {Mapillary: Street-level imagery, powered by collaboration and computer vision},
  author       = {{Mapillary contributors}},
  howpublished = {\url{https://www.mapillary.com/}},
  note         = {Images licensed under CC BY-SA 4.0}
}

And this dataset:

@misc{streetview_global,
  title        = {StreetView Global: annotated street-level imagery with scene descriptions and VQA, sourced from Mapillary},
  author       = {Fernandes, Reuben},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Reubencf/streetview-global}}
}