Issues in indoor scene

[cantaible]

trafficstars

Thank you for your great work!

When I try to use TensoRF in indoor scene dataset, such as room 0 in pre-rendered Replica data used in Semantic-NeRF, it seems not worked well.

Here is the rendered image: Here is the extracted mesh: I use the DTU dataset file your provided in other issues, and use colmap to get the camera pose as shown in NeuS tutorial. The same workflow to produce customized data is worked in TensoRF for a building object.

Then I try to check the alignment of cameras and bounding box in nerf-studio GUI, the result is: So I try to adjust aabb as your mentioned in another issue, after do that, the alignment status seems good, but the result still not have improvement.

So I have several questions, I just a beginner on nerf:

1. Can you give me some advice on how to render a correct indoor scene by TensoRF?
2. TensoRF use voxel grid to store object information, is it still suitable for indoor scene? because for normal object, the TensoRF can try to learn a voxel grid to let the central voxels to have a high density component value, but voxel grid for indoor scene often have a higher density component value in edge voxels.
3. Can you give me some general advices on how to handle indoor scene dataset?

Thank you very much!