nerfstudio icon indicating copy to clipboard operation
nerfstudio copied to clipboard
verified publisher icon nerfstudio-project

Monodepth regularization bug fix

Open AdamRashid96 opened this issue 2 years ago • 6 comments

  • Fixed a bug where we were not switching to true depth loss, when true depth data is available.

AdamRashid96 avatar Aug 08 '23 05:08 AdamRashid96

It seems as though #2312 has a different idea, FYI

machenmusik avatar Aug 08 '23 14:08 machenmusik

 @AdamRashid96 New bug: the indexes of the current image samples and the indexes of the cached depth maps may not match at different launches. Therefore, either fix the train and validation, or generate the depth for all images, and then read by indexes. image-3

@tancik maybe it's better to move the generation of depth maps from the dataset to a separate script?

ilbash avatar Aug 18 '23 20:08 ilbash

@Ilyabasharov Thanks for pointing out this bug. I'll figure out whether to move the depth map generation to a different script or not, but I'll fix this indexing bug.

AdamRashid96 avatar Aug 18 '23 21:08 AdamRashid96

@AdamRashid96 did you plan to realize spatial continuity regularization from SparseNeRF?

ilbash avatar Aug 23 '23 12:08 ilbash

@AdamRashid96 also I have found a bug that torch.nanmean useless in depth-ranking loss. if mask will be all False, torch.nanmean will give nan value.

ilbash avatar Aug 28 '23 13:08 ilbash

@AdamRashid96 What's the status on this PR? Seems like generating the depth maps in ns-process-data could be a good option? @brentyi @kerrj

kevin-thankyou-lin avatar Oct 23 '23 00:10 kevin-thankyou-lin