Thanks for organizing such a great challenge.
There are four datasets (i.e., Places2, ImageNet, FFHQ and wikiArt) and you have kindly provided the links to these datasets. I want to ask questions about the training/validation/testing splitting.
Firstly, which part of data are we allowed to use in the training process? Is there any overlap between the provided validation set (only input) and the images in the provided datasets (which we can download from the corresponding websites)? What do we need to avoid the data overlap between training and testing for this challenge?
Secondly, is it allowed to use additional data from other public datasets (e.g., CelebA-HQ)?
Thirdly, is there any requirement on the number of trained models? For example, train only one model for all the datasets and all kinds of masks, or train 4x7=28 models for the four datasets and 7 types of masks?
Could you please help to answer the above questions? Thanks!
Hello,
1. As we describe in the 'data overview' section, each dataset has publicly standardized splits for training, validation, and test (or training and validation). You *must not* use the validation or test set for training. Note that one of the objectives of the challenge is to establish a public benchmark for future research in this direction.
2. The use of additional resources is allowed, yet it has to be mentioned and discussed, in case it helps for your final performance.
3. That is a good point, and in case it was not clear, we will highlight it more in the challenge description. As one of our objectives is: "To perform a comprehensive analysis of the different types of masks", we expect one model per dataset.
We hope it is fine like this, let us know if there is anything else we can help with.
Posted by: afromero @ Feb. 16, 2022, 10:41 a.m.An additional comment for your third point. The purpose of this challenge is to foster research into mask-agnostic image inpainting, so each solution can potentially generate photo-realistic solutions in more complex or different scenarios beyond the 7 types of masks studied here.
Posted by: afromero @ Feb. 16, 2022, 10:52 a.m.Thanks for your quick reply. I have got the answers. Your explanations make the challenge rules more clear. Thanks!
Posted by: Alex_Zhou @ Feb. 16, 2022, 11:03 a.m.