Supplementary Materials for Unifying Image Quality Assessment Datasets

User Interface for the subjective test

• One image is shown to a subject at a time (without showing the reference)
• A subject is asked to rate the quality of the image on a scale from 0 to 100, either by dragging the slider or by typing the score on keyboard.
• A subject may modify the score of the last image by clicking the “Previous” button.

Incorporating Image Features During Score Mapping

One may ask whether can be further improved using an image-feature-dependent mapping:

As shown in the table below, improves for some datasets. However, this approach departs substantially from the original goal of learning a content-independent realignment function. Moreover, we have no reliable way to determine whether scores create more chances for IQAs to overfit when trained on them.

Method	AVA	CID	CSIQ	KADID	KonIQ	LIVE	CLIVE	LIVE-MD	SPAQ	TID08	TID13	Mean
Selected	.731	.766	.848	.317	.919	.823	.721	.893	.924	.866	.449	.751
Logistic 4	.732	.758	.832	.320	.920	.820	.734	.885	.927	.856	.323	.737
Logistic 5	.730	.758	.829	.284	.918	.821	.736	.903	.927	.880	.327	.737
CDF Matching	.783	.792	.806	.138	.934	.784	.717	.901	.922	.826	.375	.725
Selected +	.841	.855	.915	.447	.920	.862	.863	.876	.906	.821	.491	.800