palapapa/iqa-project-dataset

IQA Project Dataset

This is the dataset used in our capstone IQA project. Most of the labels.csvs are produced by further processing the labels provided by LIQE into a more convenient format, using our dataset preprocessing scripts.

labels.csv Columns

The meanings of the columns that can appear in labels.csv are as follows:

1. filename: The filename inside the images directory. If there are subdirectories inside images, this can contain more than one path segment.

2. mos: The mean opinion score.

3. stddev: The standard deviation of the mean opinion score.

4. distortion: One of the 11 distortion types defined by LIQE of the distorted image. The possible values are: blur, color, contrast, jpeg compression, jpeg2000 compression, noise, other, overexposure, quantization, spatial, and underexposure.

5. scene1 ~ scene3: One to three of the 9 distortion types defined by LIQE of the distorted image. If the image has less than 3 scene types, the extra ones will be empty strings. The possible values are: animal, cityscape, human, indoor, landscape, night, plant, still_life, and others.

6. reference: For synthetic distortion datasets, this is the filename inside the reference-images directory from which a distorted image is generated from.

7. set: This indicates the split an image is in. The possible values are: training, testing, and validation. If there is official splits information for a dataset, that is used.

Dataset Modifications

Here we try to list the different ways that the datasets uploaded here are different from how they originally were. This includes the modifications already made by LIQE.

1. For the CSIQ and LIVE datasets, they originally used DMOS (differential mean opinion score) with higher values indicating lower image qualities. LIQE reversed this by subtracting the original scores from the maximum of the original scores to make higher scores indicate higher image qualities.

2. The LIVE dataset contains subdirectories that each represent a distortion type. However, they also put the reference images from which the distorted images were generated into those subdirectories, despite the fact that they already exist in the refimgs directories. Those duplicate images have been removed.

3. If it is possible, all image samples from a dataset have been merged into a single images directory, and all reference images for the synthetic distortion datasets have been merged into a single reference-images directory.

4. Extra files that are not images and are not needed have been removed.

5. For the LIVE dataset, LIQE originally gave the images under the directory fastfading the white noise distortion type, even though their code already supports the fastfading distortion type. In our dataset, their distortion types have been changed to fastfading. Note that due to the pre-mapping, as will be mentioned below, the fastfading distortion type will actually then be mapped back to jpeg2000 compression.

6. The BID (Blurred Image Database) dataset from "No-reference blur assessment of digital pictures based on multi-feature classifiers" claims to have 585 authentic distortion images in the paper, but when downloaded from https://huggingface.co/datasets/chaofengc/IQA-PyTorch-Datasets/blob/main/BID.zip, it actually contains 586 labels while having 590 images. The 4 extra images (DatabaseImage0002.JPG, DatabaseImage0346.JPG, DatabaseImage0413.JPG, and DatabaseImage0439.JPG) have been manually removed since they don't have their associated labels.

7. LIQE actually provides 33 distortion types in their labels, instead of the 11 as claimed in their paper, but they still map them back to the 11 types at runtime. The 11 distortion types are also named differently than what is written in their paper. Specifically, color-related is renamed to color, JPEG compression to jpeg compression, JPEG2000 compression to jpeg2000 compression, over-exposure to overexposure, under-exposure to underexposure, spatially-localized to spatial, and others to other. The scene type indoor scene has also been renamed to indoor, night scene to night, and still-life to still_life. The distortion types have been pre-mapped back to the 11 types, and the naming for the scene and distortion types follows the source code of LIQE instead of the paper.

Notes

1. Even though the README file in the LIVE dataset claims that:

   Currently raw scores and standard deviations are not being released. They will be released in the future.

   The standard deviations are actually available in the dmos_realigned.mat file.

2. The official train/test/val splits for the KonIQ-10k dataset is downloaded from https://github.com/subpic/koniq/blob/master/metadata/koniq10k_distributions_sets.csv.

3. The csiq_labels.csv file in the CSIQ dataset doesn't contain the standard deviation of the scores. They are available at https://s2.smu.edu/~eclarson/csiq/csiq.DMOS.xlsx.

4. For the KADID-10k and BID datasets, the LIQE authors forgot to put a file that contains the distortion and scene types of all the image samples (no file named kadid10k_all_clip.txt under IQA_Database/kadid10k/splits2/ and no file named bid_all_clip.txt, unlike the other datasets), so you need to manually combine the three .txt files from any of the directories named 1 to 10 under IQA_Database/kadid10k/splits2/ and IQA_Database/BID/splits2/ before passing the merged file as the second argument to the script process_liqe_labels.py.

5. The KonIQ-10k dataset has more images than it has labels (10373 images vs 10073 labels), but that's how it originally is.