About L1 loss design?

[Luh1124]

This is an exciting piece of work. I am curious about how the loss is specifically designed. The article mentions using a log transformation to convert the colors obtained from nn.ELU(alpha=1e-3) into log space before computing the L1 loss. What details are there to avoid log(zero)? For instance, is log1p used?

[iamNCJ]

The log-encoded HDR space used in the paper is log(x+1), which inherently avoids negative values.

[Luh1124]

The log-encoded HDR space used in the paper is log(x+1), which inherently avoids negative values.

Thank you for your reply. I would like to confirm with you which coordinate system is used for the camera-to-world (c2w) transformation in the provided data (demodata): is it the OpenGL coordinate system or the OpenCV coordinate system?

[iamNCJ]

As noted in README, we use the same camera coordinate system as Blender (-Z = view direction, +Y = up, +X = right).