Monocular custom video convergence issues

[aleatorydialogue]

Hello,
Thanks a lot for this work. I have a set of images that COLMAP successfully reconstructed into sparse model.

If i use ns-process-data on these images, i get a high quality result with nerfacto model.

But when I use colmap2nerf.py to convert into the dataset format for kplanes, with or without --dynamic flag to add time, I am unable to get any rendering convergence when attempting to train.

It appears that the images are oriented differently, like somehow the camera poses are being translated differently.

Appreciate your work on getting this accessible for us truly.

[aleatorydialogue]

Ok well I'm glad to say I've had more success now after changing near/far planes and now --pipeline.model.is-contracted True.

Any further info on how I could adjust scene scale correctly would be appreciated but this appears to be just requiring more tuning than I thought.

Thanks for your work on this, very cool project

[Giodiro]

Hi @aleatorydialogue ,

Sorry for the late response, but I think that your interpretation is correct: there can be a lot of tuning needed to adapt to different datasets, due to different parameters of the scene. Hopefully the nerfstudio GUI should help figure out how to adapt them more easily.

Scene contraction should help when you have both foreground or background.