The chamfer loss is negative!

[zeal-up]

Does it normal to implement a negative loss? Since there is no constraint to make sure the Probabilistic Chamfer Loss is positive.

sigma_src_dst = (sigma_src + selected_sigma_dst) / 2
forward_loss = (torch.log(sigma_src_dst) + src_dst_min_dist / sigma_src_dst).mean()

The sigma is constrain to be positive but can be smaller than 1. Thus the log(sigma_src_dst) can be negative. And during my training, the loss do be negative. Is it normal?

[lijx10]

You are right, the loss can be negative, totally normal.

[zeal-up]

Thanks for your reply!