A high-resolution NR-IQA database for high-quality authentic photos 

The dataset comprises 6,073 UHD-1 (4k) images. These were selected from the top photos on Pixabay.com based on popularity indicators, out of a total of 1.2 million photos available at the time of indexing. The dataset ensures high photo aesthetics and technical quality, and excludes synthetic images, such as purely computer generated graphics, drawings or those that are heavily edited.

​​Features

• Resolution: all images are originally larger than UHD-1 (3840x2160 pixels) and are downscaled to a width of 3840px maintaining their aspect ratio.
• Authenticity: synthetic images were manually excluded to ensures all are genuine photos.
• Reliability: 10 thoroughly selected professional photographers and graphic artists rated each photo twice several days apart, leading to highly reliable mean opinion scores (MOS) based on 20 ratings per image.

​Cite us

The UHD-IQA database is freely available to the research community. If you use our database in your research, please cite us:

@misc{hosu2024uhdiqa-database,
      title={UHD-IQA Benchmark Database: Pushing the Boundaries of Blind Photo Quality Assessment}, 
      author={Vlad Hosu and Lorenzo Agnolucci and Oliver Wiedemann and Daisuke Iso and Dietmar Saupe},
      year={2024},
      eprint={2406.17472},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2406.17472}}

@misc{hosu2024uhdiqa-challenge,
      title={AIM 2024 Challenge on UHD Blind Photo Quality Assessment}, 
      author={Vlad Hosu and Marcos V. Conde and Lorenzo Agnolucci and Nabajeet Barman and Saman Zadtootaghaj and Radu Timofte},
      year={2024},
      eprint={2409.16271},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2409.16271}}

Image license
All images are available under the CC0 license. 
Images have been published to Pixabay.com prior to 2019.

Acknowledgements
The dataset has been created at the University of Konstanz, funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project-ID 251654672 – TRR 161.

Browse images

The dataset supported the development of models for the UHD-IQA challenge held at the ECCV AIM 2024 Workshop.
The challenge has ended, read more about the results and best performing approaches in our paper.

Downloads

images + meta-data (.zip)

10.7 Gb @ UHD-1

META-data (.zip)

Paper (.PDF)

Meta-data: ratings, popularity, indicators, tags (.csv files) 

uhd-iqa-metadata.csv

• image_name: name of the image, based on a dataset specific ID
• quality_mos: the quality mean opinion score (MOS) of the image
• set: the membership of an image to the training, validation and test sets
• subset: membership to the exclusive categories which are only part of the test and validation sets
• orig_width, orig_height: original image dimensions​

• views, likes, favorites, downloads, comments: popularity indicators from the time the image was indexed
• image_bytes: size of the original image in bytes
• user_name: the author account name on Pixabay.com
• page_url: the address of the image post on Pixabay.com
• image_name_orig: the name following the naming convention used on Pixabay.com (used in the tags file below)

uhd-iqa-tags.csv

• image_name: name of the image, based on a Pixabay naming convention 
• tag: name of machine tag, per Amazon Rekognition
• machine_confidence: confidence in the tag prediction
• parents: greater abstraction level tags related to the current

Example images: highest quality

Example images: lowest quality

Even if the lowest quality images are highly aesthetic, they show degradation that is technical in nature such as blurs, loss of contrast or under-exposure. Removing the confound of aesthetics is a good reason why the database is an interesting challenge for NR-IQA.