search

barbararoessle / dense_depth_priors_nerf

Dense Depth Priors for Neural Radiance Fields from Sparse Input Views
MIT License
389 stars 49 forks source link

Depth loss is negative #26

Open YZsZY opened 1 year ago

YZsZY commented 1 year ago

Hello and thank you for your great work! I noticed that the paper is using GaussianNLLLoss to satisfy the true sampling distribution as much as possible, as this formula shows: image But I find that this loss is always negative, is this as expected?

Also I would like to ask, have you tried to do MSE loss directly on the rendered depth map with the GT depth?

Sorry to bother you and look forward you for your reply!

barbararoessle commented 1 year ago

Hi, due to the logarithm, the depth loss can be negative, so that is ok. We have ablations (named w/o GNLL) using MSE instead of GNLL in the paper, however they equally use the depth predicted by the depth completion (not the ground truth depth from the sensor).