search

nerfstudio-project / nerfstudio

A collaboration friendly studio for NeRFs
https://docs.nerf.studio
Apache License 2.0
8.87k stars 1.18k forks source link

Best way to get the highest quality and lowest training time ? #3217

Open brahimfarhat opened 2 weeks ago

brahimfarhat commented 2 weeks ago

Dear all,

I'm trying to train a model over my custom setup with 48 fixed camera. I have the camera positions and the preprocessing step is stable. I tried many models (nerfstudio models), nerf-based (nerfacto, -big, -huge) or gs-based (splatfacto, -big), however, the final quality is simply not good enough, my doubt is that the number of input images are few (not sure), I wanted to ask the community for any recommendations to get the highest quality over my custom data ? also is there any recommendations to get the best trade off between training time and final quality ?

Thank you all. Brahim Farhat

sumanttyagi commented 2 weeks ago

1.use all images for reconstruction rather than best one, -- set colmap to exhaustive mode as it will link all the images to other 2.if you have less images go with original size as no downscale factor

brahimfarhat commented 2 weeks ago

Thank you for the response, what do you mean by using all images for reconstruction ? I generated the camera position with metashape and i dont use colmap for my pipline .. do you have a cmd line to recommend (splatfacto for example)?

ichsan2895 commented 2 weeks ago

Its difficult question. My suggestion is, please try it with good images.

1000 images that blurred < 100 images that focused