Bridging the gap between image quality and resolution

IISA-DB is a dataset of 785 high-resolution images (>2048px width) paired with their Image Intrinsic Scale (IIS) — the optimal viewing scale where perceived quality is maximized. Created through rigorous crowdsourcing with 10 expert annotators providing 20 total opinions per image via a custom web-interface, the intrinsic scales are highly reliable (95% CI: 0.057) with values ranging from 0.060 to 0.811.

The dataset addresses a critical gap in image quality research by quantifying how scale affects perceived quality, enabling development of scale-aware vision models. The dataset includes WIISA, a training strategy that generates weak labels from downscaled images, demonstrating consistent performance improvements (up to 5%) across multiple IQA architectures when adapted for the IISA task.

​Cite us

IISA-DB is freely available to the research community. If you use our database in your research, you can cite it as follows:

 @article{hosu2025IISA, title={Image Intrinsic Scale Assessment: Bridging the Gap Between Quality and Resolution}, author={Vlad Hosu and Lorenzo Agnolucci and Daisuke Iso and Dietmar Saupe}, year={2025}, eprint={2502.06476}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2502.06476}, } 

Downloads

DATABASE (.zip)

  3.5 GB file

The 'annotations.csv' fields

• img_name: the image file name
• iis: the aggregated Image Intrinsic Scale