Question about perceptual loss

Hi, I'm wondering if perceptual loss can be used for training in this challenge. Perceptual loss typically relies on a network pre-trained on imagenet for loss computation, but I did notice in a previous post that the pre-trained networks were not allowed to use as backbones. Thus I would like to check if using perceptual loss is considered as using pre-trained backbones. Thanks!

Dear Han,

Thank you for the question. Perceptual loss can be used as loss function assuming that the weights are not updated during the training phase. With pre-trained models, we aim to avoid transfer learning for the image synthesis.

Best Regards,

Matteo

Dear Matteo,

just to clarify, in this sentence: " Perceptual loss can be used as loss function assuming that the weights are not updated during the training phase" Do you mean that the weights of the pre-trained network used for the perceptual loss cannot be updated or the weights of the network used for image synthesis?

best Thomas

Hi Thomas,

The weights of the perceptual loss cannot be updated. The weights of the synthesis architecture can be updated but should be trained from scratch on the provided data.

I hope this clarifies,

Matteo