Relighting Training

[Marcel Santana]

@lmurmann Hello, thanks for releasing the dataset and the code to evaluate the models.
Could you please provide some additional details about the training of the relighting model? For my understanding, for training the relighting model you use both the input and the target images
as HDR in the log domain. So, you preprocess these images before you feed them in the training pipeline. Furthermore, you use for the loss function akin
L(I, Î) = || \nabla I - \nabla Î ||_1,
where \nabla is the spatial gradient. Am I missing anything?

Thank you!

[Lukas Murmann]

Hi Marcel, that looks correct. Log is applied as a preprocess (as in the eval script), and the loss is computed the way you wrote, still in the log domain. Let me know if you have more questions

For gradients, let me clarify:

L(I, Î) = || \nabla_x I - \nabla_x Î ; \nabla_y I - \nabla_y Î ||_1

So it is the L1 loss over x derivatives and y derivatives. In practice, we stack x direction and y direction into one tensor and then reduce everything with a single mean(abs(grads(I) - grads(Î))) call