thomasht86/road-images-and-embeddings

Norwegian Road Images with Embeddings (Trondheim Area)

A dataset of 34,908 road images from the Trondheim region of Norway (~40km radius), captured by Statens vegvesen (Norwegian Public Roads Administration) in 2025. Each image is paired with rich geospatial metadata, nearest address information, and a 3072-dimensional image embedding from Google's gemini-embedding-2-preview model.

Dataset Structure

Each example contains:

Field	Type	Description
image	Image	Road camera JPEG image (4011x2018)
doc_id	string	Unique identifier from Vegbilder
lat	float	Latitude (WGS84)
lon	float	Longitude (WGS84)
year	int	Capture year (2025)
timestamp	string	Capture time (ISO 8601)
road_category	string	Road type: E (European), R (National), F (County)
road_number	int	Road number
road_section	string	Road section (e.g., "S2D1")
meter	float	Meter position along road segment
lane	string	Lane code (1 or 2, indicating direction)
heading	float	Camera heading in degrees
county_number	int	Norwegian county number (50 = Trondheim region)
image_type	string	Camera type (Planar)
detected_objects	string	Auto-detected objects as JSON (e.g., {"car": "1"})
address_text	string	Nearest address from Geonorge (e.g., "Innherredsveien 1, 7014 TRONDHEIM, TRONDHEIM")
embedding	list[float]	3072-dim image embedding from gemini-embedding-2-preview

Road Category Distribution

Category	Count	Description
F (County)	31,412	County roads
E (European)	3,499	European highways (e.g., E6, E39)
R (National)	183	National roads

~82% of images have a resolved nearest address from Geonorge.

Usage

Load the dataset

from datasets import load_dataset

ds = load_dataset("thomasht86/road-images-and-embeddings", split="train")

# Access a single example
example = ds[0]
print(example["address_text"])   # "Kvamsveien 72, 7336 MELDAL, ORKLAND"
print(example["image"].size)     # (4011, 2018)
print(len(example["embedding"])) # 3072

Stream the dataset (recommended for large datasets)

from datasets import load_dataset

ds = load_dataset("thomasht86/road-images-and-embeddings", split="train", streaming=True)

for example in ds:
    image = example["image"]
    embedding = example["embedding"]
    lat, lon = example["lat"], example["lon"]
    # Process...

Use embeddings for similarity search

import numpy as np
from datasets import load_dataset

ds = load_dataset("thomasht86/road-images-and-embeddings", split="train")

# Build embedding matrix
embeddings = np.array(ds["embedding"])  # (34908, 3072)

# Find similar images to the first one
query = embeddings[0]
similarities = embeddings @ query / (np.linalg.norm(embeddings, axis=1) * np.linalg.norm(query))
top_k = np.argsort(similarities)[-5:][::-1]

for idx in top_k:
    print(f"  {ds[int(idx)]['address_text']} (similarity: {similarities[idx]:.3f})")

Filter by location or road type

# Only European highways
e_roads = ds.filter(lambda x: x["road_category"] == "E")

# Only images near Trondheim city center
import math
def near_center(example):
    dlat = example["lat"] - 63.43
    dlon = example["lon"] - 10.40
    return math.sqrt(dlat**2 + dlon**2) < 0.05

city_center = ds.filter(near_center)

Data Collection

1. Image metadata was collected via the Vegbilder WFS endpoint, tiling the bounding box into 0.01-degree chunks
2. Address enrichment was performed using the Geonorge punktsok API, with coordinate-grid caching at ~100m resolution
3. Image thinning was applied at 100m minimum spacing along each road segment to reduce redundancy (original dataset: 243,418 images)
4. Embeddings were generated using Google Gemini Batch API with the gemini-embedding-2-preview multimodal embedding model at 3072 dimensions

Intended Uses

• Visual road condition monitoring and analysis
• Geospatial image search and retrieval
• Multimodal search applications (text-to-image via shared embedding space)
• Training and evaluation of road scene understanding models
• Urban and infrastructure planning research

Limitations

• Images are from 2025 only (single year snapshot)
• Coverage is limited to the Trondheim area (~40km radius)
• ~18% of images lack address information (rural/remote areas)
• 186 images from the original selection could not be downloaded (0.5%)
• Embeddings are from a preview model (gemini-embedding-2-preview) which may change

Citation

If you use this dataset, please credit the original data source:

Statens vegvesen (2025). Vegbilder. Norwegian Public Roads Administration.
Licensed under NLOD 2.0: https://data.norge.no/nlod/en/2.0