search

mattpoggi / mono-uncertainty

CVPR 2020 - On the uncertainty of self-supervised monocular depth estimation
234 stars 24 forks source link

Modifying Eq 12/13 to deal with negative loss #16

Open MohitLamba94 opened 3 years ago

MohitLamba94 commented 3 years ago

Hi, Thanks for this great work. In Eq. 12/13 you have a log(variance) that might go negative for variance<1. Assuming, variance to be always +ve can we modify the Eq. 12/13 to log(1+variance) to prevent negative loss and allowing better optimization? Will this modification hamper training or net output during inference?

mattpoggi commented 3 years ago

Hi, I never tried, because the negative loss never gave issues in my experiments. You can use sigmoids over the outputs, but I think your solution should work as well.