Open Antoine251 opened 9 months ago
Hi, the provided depth map is predicted from omnidata model and it's up to scale. You need to use some scale-invariant loss if use it as supervision.
Ohh okay I see, that's why the depth loss is scale and shift invariant. But what I don't get is why we need to do *50 + 0.5 on the ground truth depth before giving it to the depth loss. The loss should remove those from my understanding
Hi!
I was checking the values of depth maps in the scannet dataset and I found it strange: the max value is never more than 0.04. I checked for a scale up in the dataparser but I found nothing. Is it possible to know by how much the depth is scaled down ? (if it is scaled down)
Thanks in advance