Question about the BACKBONE

[XJTU-Haolin]

Hi,

SceneRF is a nice work and it is very helpful for me.

Currently, I'd like to revise SceneRF and integrate it into my work (I will definitely cite this work). I am meeting a challenge of how to switch the sphere backbone to standard ResNet elegantly. Will this modification directly lead to the failure of SceneRF? If the effect is acceptable, how to deal with it with minimal revision?

Could you give me any suggestions?

Thanks for your time!

Sincerely,
Haolin

[anhquancao]

Hi,

You can just set:
add_fov_ver = 0
add_fov_ver = 0
add_fov_hor = 0
add_fov_hor = 0
in the following code
https://github.com/astra-vision/SceneRF/blob/main/scenerf/models/scenerf.py#L83

Anh the sphere_W and sphere_H to the image width and height, respectively in the code

https://github.com/astra-vision/SceneRF/blob/main/scenerf/models/scenerf.py#L77

We tested without spherical mapping in the paper, the performance decreases not very significant. This is mentioned in the Sec. ablation studies of the paper, paragraph Spherical U-Net.