They have a list :

https://arxiv.org/pdf/1801.05585.pdf

It's not dataset. Author just take some images from somewhere/taken by their camera. And the mask is not given.

What mask ? Shouldn't you define the mask ??

Mask defines the corrupted parts on image.

In the paper link , they have used dilation and CNN for feature extraction and then patched it (simple way of saying) .

Blurring and dilation seems to be the way to inpaint when mask is not there ...

Read section 4.3 :) They just make a white part at center of image (this is a mask, but it is very simple). So this is easy way to detect the corrupted parts. It didn't relate to any deep learning. The deep learning is used for another task (inpainting), not for detecting the corrupted parts.

Well, CNN (convolution) for checking holes in the image :)

:) Actually, corrupted parts are not only white holes, they are more complex :) Using deep learning to extract them is necessary. However, this paper didn't study it.