Can ImageNet pre-trained backbones be used?

Dear Organizers,

With the ease of access for ImageNet pre-trained backbones, it is highly probable that their utilization does not entail unfairness. However, I still want to make sure whether they belong to the restricted “pre-training nets”.

Best regards, Yiwen Zhang

Hi Yiwen,

Thank you for reaching out. If you wish to use ImageNet to pre-process the data: no problem. However, if you are going to use it for image synthesis, this is not allowed. Please, also double check the answer to the previous forum question https://grand-challenge.org/forums/forum/synthrad2023-643/topic/external-data-use-of-predictions-from-other-models-1490/.

If you wish to share a bit more details on your situation but do not feel like sharing details on the model publicly here on the forum, feel free to reach out to the organizers.

I hope this answer your question.

Best Regards,

Matteo

They are not used in preprocessing, I mean that the backbones (e.g., ResNet, VGGNet, DenseNet...) pre-trained on ImageNet dataset are used as the initialization of the synthesis model in training stage.

As mentioned in your previous forum question, the pre-trained models are not allowed to be used because of fairness. I quite agree that private or public task-related datasets and pre-trained models cannot be used in this competition.

However, the backbones pre-trained on ImageNet are widely used in neural network training, and these pre-trained models can be easily obtained by everyone. So I don't think that using these models will affect the fairness of the competition. But I still respect the decision of the organizers.