custom data results suffer from floaters

[emjay73]

Thank you for sharing your code!

I tried to run TensoRF using the dataset from the 'NeRF Object Removal' repository (https://github.com/nianticlabs/nerf-object-removal) and noticed there are lots of unpleasant floaters in the generated test set video.

video.mp4

video.2.mp4

I ran the code after changing the following settings.

self.near_far = [0.1,6.0]
self.scene_bbox = torch.tensor([[-1., -1., -1.], [1., 1., 1.]])

Any further suggestions?

[liuyubian]

Maybe you can change the config of fea2denseAct from softplus to relu, it works for me.

[emjay73]

@liuyubian Thank you for the suggestion, but relu actually worsens the situation in my case.