nerfstudio
nerfstudio copied to clipboard
nerfstudio-project
Cameras to TensorDataclass
Cameras is now a tensordataclass and is about as generalized as I think we can get before launch. Updated tensordataclass to make this happen, keeping a list of fields that will have custom tensor sizes before the batch dimensions. Had to bug fix some issues with TensorDataclass in general even before making these modifications.
ns-train and ns-eval appear to work just like normal, performance seems to also be at normal
The one bug I can't fix is that the input images don't show up on the viewer anymore, and I'm not sure how to fix this at this point.
Will do one last sweep tomorrow to make sure that nobody is accessing any of the fields in an outdated way, but should be fine given that everything works as normal pending the viewer bug
I'll be working on this in person tomorrow at noon-ish so I can explain more then too if this is a lot to read through
Checklist
- [x] remove ravel from the code
- [x] speed up the test cases (they are taking over 4 minutes locally compared to 30 sec on main)
- [x] make the code work for variable height and width
- [x] remove all unused code/comments
Hi @jake-austin, @tancik
Thank you all for the amazing package and I know that your TODO lists are probably overflowing like crazy. I just wanted to ask what the timeline on this PR is and if it already works with adapted transforms.json (see https://github.com/nerfstudio-project/nerfstudio/issues/868). If not, no problem :)
Hey @lkskstlr, we are refactoring the dataloader and then we will merge PR #961. It's quite a big change, so we just want to make sure it doesn't break anything. I'll close this PR since #961 implements everything here and has a test with Phototourism data.