The Cloudinary Image Dataset ’22 (CID22) is a large image quality assessment (IQA) dataset created in 2022, consisting of 22k annotated images based on 250 pristine images, compressed using (Moz)JPEG, WebP, AVIF, JPEG XL, JPEG 2000, and HEIC.
Quality range
Compared to other IQA databases like KADID-10k or TID2013, CID22 is relatively focused: distortions include only image compression, and the quality range is from medium quality to (near) visually lossless, e.g. mozjpeg q30 to q95. Previous datasets typically tended to focus on much lower qualities:
This is the range relevant for web delivery of images with various trade-offs between fidelity and bandwidth. It is also the quality range the new JPEG AIC-3 standard will focus on. CID22 is part of Cloudinary's response to the AIC-3 Call for Contributions on Subjective IQA.
License
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
Downloads
The full CID22 dataset consists of two parts:
- a training set based on 201 pristine images, and
- a validation set based on 49 pristine images.
-
CID22 dataset (7.2 GB)
250 reference images, 21903 distorted images -
CID22 validation set (1.4 GB)
49 reference images, 4292 distorted images -
CID22 validation set without distorted images (17 MB)
49 reference images -
CID22 validation set MCOS scores (103 KB)
just the CSV file (also included in any of the above)
Paper
The CID22 dataset is presented in this paper, including a detailed description and discussion of the test methodology that was used. An extended version of this paper was submitted as a contribution to the JPEG AIC-3 Call for Contributions on Subjective Image Quality Assessment.
If you use the CID22 dataset in your research, you can cite it as follows:
@article{CID22,
title={{CID22}: Large-Scale Subjective Quality Assessment for High Fidelity Image Compression},
author={Sneyers, Jon and Ben Baruch, Elad and Vaxman, Yaron},
journal={IEEE MultiMedia},
pubstate={Submitted},
year={2023},
doi={10.36227/techrxiv.22659061}}
Codec comparison
The following plot shows bitrate/distortion curves aggregated over the entire CID22 dataset:
Per-image plots are available for every image in the validation set; there are also aggregated plots available per image category, based on the full CID22 dataset: codec performance plots.
Objective metrics
Using the CID22 data to evaluate objective metrics, we get the following Kendall and Spearman rank-order correlation coefficients (KRCC and SRCC) and Pearson correlation coefficients (PCC). The sign only indicates whether the metric is of the “smaller is better” type (the number indicates amount of difference) or of the “bigger is better” type (the number indicates quality). Higher absolute values are better.
Metric | KRCC | SRCC | PCC |
---|---|---|---|
(SSIMULACRA 2) | 0.6934 | 0.882 | 0.8601 |
Butteraugli 2-norm | -0.6575 | -0.8455 | -0.8089 |
Butteraugli 3-norm | -0.6547 | -0.8387 | -0.7903 |
DSSIM 3.2 | -0.6428 | -0.8399 | -0.7813 |
VMAF | 0.6176 | 0.8163 | 0.7799 |
FSIM | 0.6089 | 0.8005 | 0.7676 |
PSNR-HVS | 0.6076 | 0.8100 | 0.7559 |
Butteraugli max-norm | -0.5843 | -0.7738 | -0.7074 |
SSIM | 0.5628 | 0.7577 | 0.7005 |
MS-SSIM | 0.5596 | 0.7551 | 0.7035 |
LPIPS | -0.5417 | -0.7316 | -0.6932 |
SSIMULACRA 1 | -0.5255 | -0.7175 | -0.6940 |
PSNR-Y | 0.4452 | 0.6246 | 0.5901 |
PSNR (ImageMagick compare -metric psnr ) | 0.3472 | 0.5002 | 0.4817 |
CIEDE2000 | 0.3154 | 0.4584 | 0.4096 |
Butteraugli, SSIMULACRA 1 and 2 are also part of libjxl. For SSIM, MS-SSIM, PSNR-Y, PSNR-HVS and CIEDE2000, the libvmaf implementation was used.