Closed dispoth closed 1 year ago
The main results in the paper use the NeRF-synthetic dataset which has 100 training views and produces good results on new/test views. We also include an ablation study with 25 training views, where the trick for retaining reasonably good test results is to increase the weight on the TV regularization.
Thanks @sarafridov, I definitely want to push the minimum samples so will give the paper a more thorough read and try out the TV regularization.
Hi,
My objects are captured from an upper semi-sphere in relatively equally spaced elevations/azimuths and I have ~600 images per object. I'm interested in training on <100 images.
Will this produce quality renderings on novel (test) views?
What techniques, applicable to this codebase, are there for improving rendering quality (PSNR) with a low number of input samples?
Thanks!