dataset_info:
features:
- name: image
dtype: image
- name: text
dtype: string
- name: id
dtype: string
- name: title
dtype: string
- name: description
dtype: string
- name: credits
dtype: string
- name: url
dtype: string
- name: Id
dtype: string
- name: Type
dtype: string
- name: Release date
dtype: string
- name: Related releases
dtype: string
- name: Size
dtype: string
- name: Name
dtype: string
- name: Distance
dtype: string
- name: Constellation
dtype: string
- name: Category
dtype: string
- name: Position (RA)
dtype: string
- name: Position (Dec)
dtype: string
- name: Field of view
dtype: string
- name: Orientation
dtype: string
- name: width
dtype: int64
- name: height
dtype: int64
- name: file_size
dtype: int64
- name: crop_w
dtype: int64
- name: crop_h
dtype: int64
- name: cropped
dtype: bool
- name: Related science announcements
dtype: string
- name: Related announcements
dtype: string
splits:
- name: train
num_bytes: 94474695584.124
num_examples: 2706
download_size: 61236366105
dataset_size: 94474695584.124
license: cc-by-4.0
task_categories:
- text-to-image
language:
- en
tags:
- space
pretty_name: ESA Hubble Deep Space Images & Captions
size_categories:
- 1K<n<10K
Dataset Card for ESA Hubble Deep Space Images & Captions
Table of Contents
- Dataset Description
- Dataset Structure
- Dataset Creation
- Considerations for Using the Data
- Additional Information
Dataset Description
- Homepage: ELI5 homepage
- Repository: ELI5 repository
- Paper: ELI5: Long Form Question Answering
- Point of Contact: Yacine Jernite
Dataset Summary
The ESA Hubble Deep Space Images & Captions dataset is composed primarily of Hubble deep space scans as high-resolution images, along with textual descriptions written by ESA/Hubble. Metadata is also included, which enables more detailed filtering and understanding of massive space scans. The purpose of this dataset is to enable text-to-image generation methods for generating high-quality deep space scans from prompts.
Examples
A grazing encounter between two spiral galaxies: In the direction of the constellation Canis Major, two spiral galaxies pass by each other like majestic ships in the night. The near-collision has been caught in images taken by the NASA/ESA Hubble Space Telescope and its Wide Field Planetary Camera 2.
Credit: NASA/ESA and The Hubble Heritage Team (STScI)
The magnificent starburst galaxy Messier 82: This mosaic image of the magnificent starburst galaxy, Messier 82 (M82) is the sharpest wide-angle view ever obtained of M82. It is a galaxy remarkable for its webs of shredded clouds and flame-like plumes of glowing hydrogen blasting out from its central regions where young stars are being born 10 times faster than they are inside in our Milky Way Galaxy.
Credit: NASA, ESA and the Hubble Heritage Team (STScI/AURA). Acknowledgment: J. Gallagher (University of Wisconsin), M. Mountain (STScI) and P. Puxley (NSF).
Extreme star cluster bursts into life in new Hubble image: The star-forming region NGC 3603 - seen here in the latest Hubble Space Telescope image - contains one of the most impressive massive young star clusters in the Milky Way. Bathed in gas and dust the cluster formed in a huge rush of star formation thought to have occurred around a million years ago. The hot blue stars at the core are responsible for carving out a huge cavity in the gas seen to the right of the star cluster in NGC 3603's centre.
Credit: NASA, ESA and the Hubble Heritage (STScI/AURA)-ESA/Hubble Collaboration
Statistics
- There are a total of 2,706 deep space images
- The complete uncompressed size of the dataset is 120 GB, so definitely make use of Streaming
- The average image is 44 MB, while the max image size is 432 MB
- The average image has a height of 2,881 pixels, and an average width of 3,267 pixels
Supported Tasks and Leaderboards
text-to-image
: The dataset can be used to train a model for conditional image generation from text. A conditional text-to-image generation model is presented with a text prompt, and is asked to generate an image which aligns with that text prompt. Model performance is typically measured by human judgement, as it is difficult to automatically measure the quality of generated images and how closely they match the text prompt. An example of a text-to-image model is Stable Diffusion v2-1.
Languages
The text in the dataset is in English, as written by the writers from ESA/Hubble at https://esahubble.org/. The associated BCP-47 code is en
.
Dataset Structure
Data Instances
A typical data point comprises a high-quality deep space scan as an image, along with a textual description of that image produced by ESA/Hubble.
The textual description was derived by combining the title
and the description
of the image from the ESA/Hubble website.
Additionally, each data point also contains significant metadata about the image, such as the type of image, credits, the URL, the release date, and more.
An example looks as follows:
{
'image': <encoded image>,
"text":"A grazing encounter between two spiral galaxies: In the direction of the constellation Canis Major, two spiral galaxies pass by each other like majestic ships in the night. The near-collision has been caught in images taken by the NASA/ESA Hubble Space Telescope and its Wide Field Planetary Camera 2.",
"id":"opo9941a",
"title":"A grazing encounter between two spiral galaxies",
"description":"In the direction of the constellation Canis Major, two spiral galaxies pass by each other like majestic ships in the night. The near-collision has been caught in images taken by the NASA/ESA Hubble Space Telescope and its Wide Field Planetary Camera 2.",
"credits":"NASA/ESA and The Hubble Heritage Team (STScI)",
"url":"https://esahubble.org/images/opo9941a/",
"Id":"opo9941a",
"Type":"Local Universe : Galaxy : Type : Interacting",
"Release date":"4 November 1999, 07:00",
"Size":"2907 x 1486 px",
"Name":"IC 2163, NGC 2207",
"Distance":"110 million light years",
"Constellation":"Canis Major",
"Category":"Galaxies",
"Position (RA)":"6 16 25.10",
"Position (Dec)":"-21° 22' 34.62\"",
"Field of view":"4.82 x 2.47 arcminutes",
"Orientation":"North is 191.2\u00b0 right of vertical",
"width":2907,
"height":1486,
"file_size":12959406,
"crop_w":0,
"crop_h":0,
"cropped":false
}
Data Fields
image
: encoded RGB.png
image of the deep space scantext
: text description of image, a combination oftitle
+ ': ' +description
id
: id of the image from ESA/Hubbletitle
: textual title of image from ESA/Hubble URLdescription
: textual description of image from ESA/Hubble URLcredits
: required credits for each image from ESA/Hubble URLurl
: ESA/Hubble URLId
: id of the image from ESA/Hubble (from website metadata)Type
: type of deep space scanRelease date
: release date of deep space scanSize
: size of original imageName
: name of celestial entities present in imageDistance
: distance from celestial entities present in imageConstellation
: constellation of celestial entities present in imageCategory
: category of celestial entities present in imagePosition (RA)
: coordinates for deep space scan used by Hubble telescopePosition (Dec)
: coordinates for deep space scan used by Hubble telescopeField of view
: coordinates for deep space scan used by Hubble telescopeOrientation
: coordinates for deep space scan used by Hubble telescopewidth
: width of image, same if the image did not need to be cropped, but otherwise could differ fromSize
height
: height of image, same if the image did not need to be cropped, but otherwise could differ fromSize
file_size
:width
xheight
x 3 bytes, used to estimate size of raw imagescrop_w
: width starting point of image if cropped, otherwise 0crop_h
: height starting point of image if cropped, otherwise 0cropped
: whether this image needed to be cropped or not
Data Splits
The data is only provided in a single training split.
Dataset Creation
Curation Rationale
The ESA Hubble Deep Space Images & Captions dataset was built to provide ease of access to extremely high-quality Hubble deep space scans. Images from the Hubble telescope have already inspired millions, and the hope is that this dataset can be used to create inspiring models and approaches to further push interest in space & cosmology.
Source Data
Initial Data Collection
All images were collected from https://esahubble.org/.
Fullsize Original images & metadata were crawled from the ESA Hubble website using Scrapy.
Images were downloaded as .tiff
files with the following code, while
additional metadata was later collected for each image using the following code.
As the ESA Hubble website collects images from a wide variety of sources, images were filtered to try to avoid any non-space scan images as follows: The ESA Hubble Advanced Image Search enables the following filtering parameters:
- images with Minimum size greater than or equal to 400x300
- Ranking greater than or equal to Fair or better
- Type containing 'Observation'
This reduced significantly the number of images which had nothing to do with Hubble deep space scans. A total of around 3,000 space images were collected with this method.
Filtering
Further automatic and manual filtering was performed to remove the following:
- improperly classified images
- space renders
- diagrams with text
- images of celestial bodies within our solar system
This brought the total number of deep space images down to 2,593. This process was not perfect, and there likely remain some images in the dataset that should be removed in the future.
Preprocessing
Some of the deep space scans were as large as 34,372x19,345, with a bit depth of 24 (nearly 2 GB).
Unfortunately, these images were too large to upload easily.
Therefore, images were automatically subdivided in half if they were above 12,000 pixels in either height or width.
Subdivided images were also tagged with additional metadata, such that users can reconstruct the original images if they would prefer.
Otherwise, metadata was copied across subdivided images.
Additionally, images were converted from RGB/RGBX .tiff
to RGB .png
files to avoid encoding issues.
This process resulted in 2,706 final deep space images.
Annotations
The dataset does not contain any additional annotations.
Annotation process
[N/A]
Who are the annotators?
[N/A]
Personal and Sensitive Information
[N/A]
Considerations for Using the Data
Social Impact of Dataset
The purpose of this dataset is to help inspire people to be interested in astronomy.
A system that succeeds at text-to-image generation would be able to generate inspiring deep space scans, providing interesting and inspiring art for those interested in space. This dataset provides a starting-point for building such a system by providing text and image pairs for Hubble deep space scans.
Discussion of Biases
It is unfortunate that we currently only have English captions for these deep space scans. In the future, expanding these captions to more languages could help spread interest in astronomy far and wide.
Additionally, these captions may be too technical for the average person to effectively utilize for a text-to-image model.
Other Known Limitations
[N/A]
Additional Information
Dataset Curators
The dataset was initially created by all the wonderful researchers, engineers, scientists, and more behind the Hubble Telescope, NASA, and the ESA. Maxwell Weinzierl collected, filtered, and preprocessed this data for ease of use.
Licensing Information
ESA/Hubble images, videos and web texts are released under the Creative Commons Attribution 4.0 International license and may on a non-exclusive basis be reproduced without fee provided they are clearly and visibly credited. See https://esahubble.org/copyright/ for additional conditions for reproduction and copyright.
Citation Information
If you use this dataset, please cite it as:
@misc{weinzierl2023hubble,
author = {Weinzierl, Maxwell A.},
title = {ESA Hubble Deep Space Images & Captions},
year={2023},
howpublished= {\url{https://huggingface.co/datasets/Supermaxman/esa-hubble}}
}
Contributions
Thanks to @supermaxman for adding this dataset.