Question about the empty_space_loss
Net-Maker opened this issue · comments
CookMaker commented
I am learning the scene part in this excellent work.
But I found that the empty_space_loss in paper is like that:
But in our code,it's like:
coarse_empty_space_loss = torch.zeros_like(coarse_rgb_loss)
if self.penalize_empty_space > 0:
depth = batch['depth'][:, None].repeat(1, _n).to(device)
closer_mask = z_vals < (depth * self.opt.margin)
coarse_empty_space_loss += self.empty_space_loss_fn(
torch.tanh(torch.relu(out[closer_mask][:, 3])),
torch.zeros_like(out[closer_mask][:, 3])
) * self.penalize_empty_space
I am kind of confuse. could you tell me the meaning of the double activate functions?
And where to correspond the formula in the paper?