search

facebookresearch / sapiens

High-resolution models for human tasks.
https://about.meta.com/realitylabs/codecavatars/sapiens/
Other
4.52k stars 255 forks source link

Depth Loss #38

Closed kkaytekin closed 2 months ago

kkaytekin commented 2 months ago

Hello, in the paper, in page 5, the equation 3 is referenced to 32: "Depth map prediction from a single image using a multi-scale deep network" from Eigen et. al. . But looking at this reference couldn't answer the following:

What is the reason to put the loss term in a square root there?

Sum of squares will always be $\geq$ square of sums, so the square root in this formulation will always be safe to use, but what is the underlying reason to do so?

Best regards,

rawalkhirodkar commented 2 months ago

Great question! This is a popular topic about the SiLog Loss. Please read this article for the explanation.