Datasets:
dataset_info:
features:
- name: image
dtype: image
- name: caption
dtype: string
splits:
- name: train
num_bytes: 10341164216.808
num_examples: 11707
download_size: 11002607252
dataset_size: 10341164216.808
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
language:
- da
pretty_name: Nordjylland News - Image caption dataset
size_categories:
- 10K<n<100K
license: apache-2.0
task_categories:
- image-to-text
- zero-shot-image-classification
- feature-extraction
task_ids:
- image-captioning
Dataset Card for "nordjylland-news-image-captioning"
Dataset Description
- Point of Contact: Oliver Kinch
- Size of dataset: 11 GB
Dataset Summary
This dataset is a collection of image-caption pairs from the Danish newspaper TV2 Nord.
Supported Tasks and Leaderboards
Image captioning is the intended task for this dataset. No leaderboard is active at this point.
Languages
The dataset is available in Danish (da
).
Dataset Structure
An example from the dataset looks as follows.
{
"file_name": "1.jpg",
"caption": "Bruno Sørensen og Poul Erik Pedersen er ofte at finde i Fyensgade Centret."
}
Data Fields
file_name
: astring
giving the file name of the image.caption
: astring
feature.
Dataset Statistics
Number of samples
11707
Image sizes
All images in the dataset are in RGB format, but they exhibit varying resolutions:
Width ranges from 73 to 11,830 pixels.
Height ranges from 38 to 8,268 pixels.
The side length of a square image with the same number of pixels as an image with height and width is approximately given as .
Plotting the distribution of gives an insight into the sizes of the images in the dataset.
Caption Length Distribution
Potential Dataset Issues
There are 14 images with the caption "Arkivfoto".
There are 37 images with captions consisting solely of a source reference, such as "Kilde: <name of source>".
You might want to consider excluding these samples from the model training process.
Dataset Creation
Curation Rationale
There are not many large-scale image-captioning datasets in Danish.
Source Data
The dataset has been collected through the TV2 Nord API, which can be accessed here.
Additional Information
Dataset Curators
Oliver Kinch from the The Alexandra Institute
Licensing Information
The dataset is licensed under the CC0 license.