Closed DapengFeng closed 1 year ago
Hello! Thank you for your comment.
Sat-NeRF uses the index of each input image in the training set to learn a transient embedding.
Validation images are not part of the training set, so there are different ways to handle them: e.g., do not use any transient embedding, or assign them a random index or a predefined index.
With predefined_val_ts
, I simply fixed the selected image indexes for the validation views, but you can use any of the choices I listed, there is no correct way to do it.
Best,
Roger
Dear Roger,
Thank you for your reply. Does it mean that the value of predefined_val_ts
has no effect on the validation performance?
Best, Dapeng
Exactly, predefined_val_ts
does not alter the output rgb values or the output geometry. PSNR and elevation MAE will not be affected.
Best,
Roger
Well, that is cool. Thank you.
Hi @rogermm14 sorry for asking this again. Does it mean for any other AOI, we can fix the predefined_val_ts to any number, and the validation metrics wouldn't be affected?
Hi, thank you for the excellent work. After reading the paper and the code, I am still confused about the definition of
predefined_val_ts
. Could you give suggestions for applying to other scenes inDFC2019
? https://github.com/centreborelli/satnerf/blob/de7bd6ec1cbd6f3895949350d6ca98ba56e87bdd/eval_satnerf.py#L186-L209