DATA.md

Preparing Datasets

Deraining

	All-in-One comparison	Single-task comparison
Train	Rain100L	Rain200L, Rain200H, DID, DDN
Test	Rain100L	Rain200L, Rain200H, DID, DDN

Dehazing

	All-in-One comparison	Single-task comparison
Train	RESIDE	RESIDE
Test	SOTS-outdoor	SOTS-outdoor

Note that, in RESIDE dataset, there has some overlap between training data and testing data, before training we exclude these overlapped sampels in training set.

Desnowing

	All-in-One comparison	Single-task comparison
Train	Snow100K	Snow100K
Test	Snow100K-test	Snow100K-test

Note that, we evaluate the performance on the whole Snow100K-test. Previous works may evaluate their performance on a susbet of Snow100K-test, but they did not clarify how to construct testing data.

Motion-Deblurring

	All-in-One comparison	Single-task comparison
Train	GoPro	GoPro
Test	GoPro	GoPro

Gaussian Color Image Denoising

	All-in-One comparison	Single-task comparison
Train	DIV2K, Flickr2K, WED, BSD	DIV2K, Flickr2K, WED, BSD
Test	BSD68	Urban100

Compression Artifacts Removal

	All-in-One comparison	Single-task comparison
Train	DIV2K, Flickr2K, WED, BSD	DIV2K, Flickr2K, WED, BSD
Test	LIVE1	LIVE1

Lowlight Enhancement

	All-in-One comparison	Single-task comparison
Train	LOLv1	LOLv1
Test	LOLv1	LOLv1

We follow dataset usage as Retinexformer.

Multiple Mixed Degradation

	All-in-One comparison
Train	DIV2K, Flickr2K
Test	DIV2K-val

The way of degradation follows Real-ESRGAN, BSRGAN, MiOIR, and OneRestore.