KonIQ-10k extended with annotations about the presence and type of defect

KonIQ++ contains labels for 10,073 images regarding the presence of four categories of distortion, as an extension of the original KonIQ-10k dataset. KonIQ++ is meant to explore how this information is beneficial to NR-IQA (quality score prediction), and also to evaluate how well image distortions could be identified.

Each participant in the study provided a 5-point rating for the technical quality, from bad (1) to excellent (5), and a binary annotation for the presence and type of distortion visible. The categories of distortion were blurs, artifacts, colors, contrast, and a category for "other" (multiple could be selected). This yielded a mean opinion score in [0,1] for each of the distortions, from an average of 58 ratings per image.

In the paper we also introduce a NR-IQA model, with two side networks connected to a shared backbone network that predict both image quality and distortion magnitude. You can find the code on Github.

​Cite us

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

@inproceedings{su2021koniq++,
author={S. {Su}, V. {Hosu}, H. {Lin}, Y. {Zhang}, D. {Saupe}},
booktitle={The 32nd British Machine Vision Conference (BMVC)},
title={KonIQ++: Boosting No-Reference Image Quality 
Assessment in the Wild by Jointly Predicting Image 
Quality and Defects},
year={2021}}

Downloads

annotations (.CSV)

1.6 Mb download

The annotations file header

• filename: the image file name from KonIQ-10k
• qmos: quality mean opinion score from the KonIQ++ experiment
• sd: the standard deviation of the quality ratings
• votes: number of ratings received, after user screening
• artifacts, blur, contrast, colors, other: frequency at which each of the degradation types were reported 
• degraded.amount: the mean of all five degradation frequencies