autonomousvision/factor-fields

Reproducing nerf generalization experiments

TangYucopper opened this issue · 4 comments

TangYucopper commented

Hi there, thanks for your amazing codes!
I'm trying to reproduce the nerf generalization experiments, for which I run

python train_across_scene.py configs/nerf_set.yaml

, followed by changing the ckpt in nerf_ft.yaml to the pretrained checkpoint obtained in the first stage, setting with_dropout in nerf_ft.yaml as False, setting basis_type as 'mlp' and running:

python train_across_scene_ft.py configs/nerf_ft.yaml

However, the results reported by the latter are 21.84 PSNR, which is worse than that reported in your paper. I wonder if my reproducing procedure is correct. And I would appreciate it if I can get your valuable guidance.

Thanks.

apchenstu commented

Hi there, which scene and which setting (how many views) are you working on?

TangYucopper commented

Hi, I am working on scene 100, as it is the default in nerf_ft.yaml. And the factor field was pretrained on scene [0, 100)

apchenstu commented

Hi, I am working on scene 100, as it is the default in nerf_ft.yaml. And the factor field was pre-trained on scene [0, 100)

Okay, how many views? here are the training and testing script: https://github.com/autonomousvision/factor-fields/blob/main/run_batch.py#L52-L84

the scores are an average score of 8 random sample scenes.

TangYucopper commented

250 views. Thank you for your reply. I will give it a try.