Dataset Card for Div2k

Dataset Summary

DIV2K is a dataset of RGB images (2K resolution high quality images) with a large diversity of contents.

The DIV2K dataset is divided into:

• train data: starting from 800 high definition high resolution images we obtain corresponding low resolution images and provide both high and low resolution images for 2, 3, and 4 downscaling factors
• validation data: 100 high definition high resolution images are used for genereting low resolution corresponding images, the low res are provided from the beginning of the challenge and are meant for the participants to get online feedback from the validation server; the high resolution images will be released when the final phase of the challenge starts.

Install with pip:

pip install datasets super-image

Evaluate a model with the super-image library:

from datasets import load_dataset
from super_image import EdsrModel
from super_image.data import EvalDataset, EvalMetrics

dataset = load_dataset('eugenesiow/Div2k', 'bicubic_x2', split='validation')
eval_dataset = EvalDataset(dataset)
model = EdsrModel.from_pretrained('eugenesiow/edsr-base', scale=2)
EvalMetrics().evaluate(model, eval_dataset)

Supported Tasks and Leaderboards

The dataset is commonly used for training and evaluation of the image-super-resolution task.

Unofficial super-image leaderboard for:

• Scale 2
• Scale 3
• Scale 4
• Scale 8

Languages

Not applicable.

Dataset Structure

Data Instances

An example of train for bicubic_x2 looks as follows.

{
    "hr": "/.cache/huggingface/datasets/downloads/extracted/DIV2K_valid_HR/0801.png",
    "lr": "/.cache/huggingface/datasets/downloads/extracted/DIV2K_valid_LR_bicubic/X2/0801x2.png"
}

Data Fields

The data fields are the same among all splits.

• hr: a string to the path of the High Resolution (HR) .png image.
• lr: a string to the path of the Low Resolution (LR) .png image.

Data Splits

name	train	validation
bicubic_x2	800	100
bicubic_x3	800	100
bicubic_x4	800	100
bicubic_x8	800	100
unknown_x2	800	100
unknown_x3	800	100
unknown_x4	800	100
realistic_mild_x4	800	100
realistic_difficult_x4	800	100
realistic_wild_x4	800	100

Dataset Creation

Curation Rationale

Please refer to the Initial Data Collection and Normalization section.

Source Data

Initial Data Collection and Normalization

Resolution and quality: All the images are 2K resolution, that is they have 2K pixels on at least one of the axes (vertical or horizontal). All the images were processed using the same tools. For simplicity, since the most common magnification factors in the recent SR literature are of ×2, ×3 and ×4 we cropped the images to multiple of 12 pixels on both axes. Most of the crawled images were originally above 20M pixels. The images are of high quality both aesthetically and in the terms of small amounts of noise and other corruptions (like blur and color shifts).

Diversity: The authors collected images from dozens of sites. A preference was made for sites with freely shared high quality photography (such as https://www.pexels.com/ ). Note that we did not use images from Flickr, Instagram, or other legally binding or copyright restricted images. We only seldom used keywords to assure the diversity for our dataset. DIV2K covers a large diversity of contents, ranging from people, handmade objects and environments (cities, villages), to flora and fauna, and natural sceneries including underwater and dim light conditions.

Partitions: After collecting the DIV2K 1000 images the authors computed image entropy, bit per pixel (bpp) PNG compression rates and CORNIA scores (see Section 7.6) and applied bicubic downscaling ×3 and then upscaling ×3 with bicubic interpolation (imresize Matlab function), ANR [47] and A+ [48] methods and default settings.

The authors randomly generated partitions of 800 train, 100 validation and 100 test images until they achieved a good balance firstly in visual contents and then on the average entropy, average bpp, average number of pixels per image (ppi), average CORNIA quality scores and also in the relative differences between the average PSNR scores of bicubic, ANR and A+ methods.

Only the 800 train and 100 validation images are included in this dataset.

Who are the source language producers?

The authors manually crawled 1000 color RGB images from Internet paying special attention to the image quality, to the diversity of sources (sites and cameras), to the image contents and to the copyrights.

Annotations

Annotation process

No annotations.

Who are the annotators?

No annotators.

Personal and Sensitive Information

All the images are collected from the Internet, and the copyright belongs to the original owners. If any of the images belongs to you and you would like it removed, please kindly inform the authors, and they will remove it from the dataset immediately.

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

• Original Author: Radu Timofte

Licensing Information

Please notice that this dataset is made available for academic research purpose only. All the images are collected from the Internet, and the copyright belongs to the original owners. If any of the images belongs to you and you would like it removed, please kindly inform the authors, and they will remove it from the dataset immediately.

Citation Information

@InProceedings{Agustsson_2017_CVPR_Workshops,
    author = {Agustsson, Eirikur and Timofte, Radu},
    title = {NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study},
    booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    url = "http://www.vision.ee.ethz.ch/~timofter/publications/Agustsson-CVPRW-2017.pdf",
    month = {July},
    year = {2017}
} 

Contributions

Thanks to @eugenesiow for adding this dataset.