MIDD: Meikoudai Image Distortion Dataset

Subjective assessment dataset of visually lossless ratio by absolute binary decision (ABD) for IQA of near-lossless compression.

Dataset type

Link

Meikoudai is an abbreviation of Nagoya Institute of Technology in Japanese.

The dataset name MIDD (QoMEX2023) has been renamed to MIDD-GAR due to the addition of a new dataset (MIDD-CAP) (2/7/2025).

1. MIDD-GAR (Gray and Resolution, JPEG, WebP, HIEF), QoMEX2023

[1] Soichiro Honda, Yoshihiro Maeda, and Norishige Fukushima, “Dataset of Subjective Assessment for Visually Near-Lossless Image Coding based on Just Noticeable Difference,” in Proc. International Conference on Quality of Multimedia Experience (QoMEX), 2023. [IEEE Xplore] [pdf]

@INPROCEEDINGS{honda2023dataset,
  author={Honda, Soichiro and Maeda, Yoshihiro and Fukushima, Norishige},
  booktitle={Proceedings of International Conference on Quality of Multimedia Experience (QoMEX)}, 
  title={Dataset of Subjective Assessment for Visually Near-Lossless Image Coding based on Just Noticeable Difference}, 
  year={2023},
  pages={236-239},
  doi={10.1109/QoMEX58391.2023.10178524}
}

2. MIDD-CAP (Color and Patch, JPEG), submitted 2025.

[2] Soichiro Honda, Yoshihiro Maeda, Osamu Watanabe, and Norishige Fukushima, “A Multi-Level Patch Dataset for JPEG Image Quality Assessment by Absolute Binary Decision,” submitted, 2025.

@INPROCEEDINGS{honda2023dataset,
  author={Honda, Soichiro and Maeda, Yoshihiro and Watanabe, Osamu and Fukushima, Norishige},
  booktitle={submitted}, 
  title={A Multi-Level Patch Dataset for JPEG Image Quality Assessment by Absolute Binary Decision}, 
  year={2025},
  pages={9 page},
}

Specification for Each Dataset

1. MIDD-GAR, QoMEX2023

Codec Size QPs
JPEG 512x512 70, 60, 50, 35, 20
JPEG 1024x1024 90, 80, 70, 60, 50
WebP(on) 512x512 70, 60, 50, 35, 20
WebP(on) 1024x1024 90, 80, 70, 60, 50
WebP(off) 512x512 70, 60, 50, 35, 20
WebP(off) 1024x1024 90, 80, 70, 60, 50
HEIF 512x512 45, 40, 35, 30, 25
HEIF 1024x1024 55, 50, 45, 40, 35

example plot

2. MIDD-CAP, submitted 2025.

Codec images Size QPs
JPEG 24 496x496 inf, 90, 80, 70, 60, 50, 40, 30, 20
JPEG 48 248x248p inf, 90, 80, 70, 60, 50, 40, 30, 20
JPEG 96 124x124p inf, 90, 80, 70, 60, 50, 40, 30, 20

Explanation of 1. MIDD-GAR directory structure

source image

The 10 source images are contained in the following directory.

source_image/kodim(image number).png

Source images are losslessly compressed by OptiPNG.

compression image

The distorted images are contained in the following directory.

compression_image/kodim(image number)\_(quality rate)\_(compression metrix).(filename extention)

HEIF images are decoded and then lossless compressed by OptiPNG for usability. JPEG and WebP are their own formats.

identification_ratios_data

The identification ratios between the original and compressed images are listed. The unit of identification ratio values are %.

data_per_participants

The data for each of the 30 participants is listed individually. The images that participants judged to be the same are marked as “1”, and those judged to be different are marked as “0”.

Explanation of 2. MIDD-CAP directory structure

source image

The following directory contains 24 pristine images (496x496), 48 pristine patches (248x248), and 96 pristine patches (124x124).

source_image_496/kodak(image_number1-24).png
source_image_248/kodak(image_number1-24)_(patch_number1-2).png
source_image_124/kodak(image_number1-24)_(patch_number1-4).png

Pristine source images/patches are losslessly compressed by OptiPNG.

distortion image

The following directory contains 192 distorted images (496x496), 384 distorted patches (248x248), and 768 pristine patches (124x124).

distortion_image_496/kodak(image_number1-24)_(qp20-90).png
distortion_image_248/kodak(image_number1-24)_(patch_number1-2)_(qp20-90).png
distortion_image_124/kodak(image_number1-24)_(patch_number1-4)_(qp20-90).png

JPEG-encoded images are decoded and then lossless compressed by OptiPNG for usability.

data_vlrate

The visually lossless ratio values for each data are stored in CSV format. The unit of the ratio values are %. Each csv file (result_496.csv, result_248.csv, result_124.csv) has 1D vector representation.

The order is as follows

for image_number=1:24
 for patch_number=1:(1 or 2 or 4)
  for QP = 20:90:10

If QP reaches 90, the next patch_numer index to start from is QP20, and if the patch number reaches MAX, the next image_number to start from is.

For example, result_496.csv are

96.667  (kodak1_20.png)
90      (kodak1_30.png)
70      (kodak1_40.png)
...

In the case of result_124.csv,

96.667  (kodak1_1_20.png)
90      (kodak1_1_30.png)
90      (kodak1_1_40.png)
...

data_per_participants

The data for each of the 30 participants is listed individually. The images that participants judged to be the same are marked as “1”, and those judged to be different are marked as “0”.

under construction

Related Project

GitHub link

On this page, the results of upsampling using the same protocol and subjective assessment are published. Rather than simple upsampling, high-resolution guided upsampling such as joint bilateral upsampling is used. The results of subjective assessment of the approximate processing that accelerates processing speed by guided upsampling of the low-resolution image processed with the high-resolution image before processing.

[3] Teppei Tsubokawa, Hiroshi Tajima, Yoshihiro Maeda, and Norishige Fukushima “Local Look-Up Table Upsampling for Accelerating Image Processing,” Multimedia Tools and Applications, 2023.