Summary

We collected a dataset of binary human judgement of image quality of 500 images adjusted to various configuration using brightness and contrast settings. The images were taken from the MIT-Adobe FiveK Dataset. We made sure that each image is seen by multiple people, and that we have dense sampling in the brightness-contrast adjustment space. The judgement is collected through Amazon Mechanical Turks. The experiment participants were compensated.

This dataset was collected for our project: Predicting Range of Acceptable Photographic Tonal Adjustments. If you use this dataset, please cite it using the this Bibtex.

You can quickly look at our dataset using Data Explorer.

Data

Our data is provided in csv format. Each row represent a data point. Here is the explanation of the columns:

• img_no: Image number. Refer to the MIT-Adobe FiveK Dataset for the actual images.
• label: Acceptability provided by the Amazon Mechanical Turkers. 1 indicates acceptable adjustment, while 0 indicates not acceptable.
• brightness: Brightness setting, refer to our paper for detail.
• contrast: Contrast settings, refer to our paper for detail.
• id: Id of the data points. Use this to refer to the pre-adjusted images (see below).

The dataset can be downloaded here.

If you don't want to reimplement our adjustment pipeline, the preadjusted images is provided in a tar file. Since we have over 300,000 data points, the file is quite large (~45GB). They can be downloaded here (SHA1:4eba28660984b9a4c8f514d5ec189762f7a9f571).

Acknowledgement

The author would like to thank Quanta Computer Inc. for their generous support of this project.