Export Scene as 3D Pointcloud

[brz1409]

How to export a 3D point cloud of the scene geometry from NeRFRAC?

Hi, and thank you for sharing this impressive work and code!

I would like to extract a 3D point cloud from a trained NeRFRAC model that represents the scene geometry below the refractive surface — not just the surface itself.

I have read Section 4.4 of the paper and understand that the geometry is modeled by sampling points along the refracted rays beneath the surface (x = Xs + t * v′), and that radiance and density are predicted at those sample points.

However, I'm having difficulty identifying the correct variables and method in the codebase to extract this. Most of my attempts result only in the flat refractive interface, or distorted point clouds that do not resemble the true scene (like the cow shown in the paper).

---

Request

Could you please provide guidance on:

• How to correctly extract 3D points that represent the actual underwater scene geometry?
• Which variables or parts of the code should be used for sampling and exporting these points?
• Whether a script or minimal working example exists for exporting the scene as a .ply or similar point cloud?

Any help or clarification would be highly appreciated.

Thanks again for making this work available

Best regards!

[Yifever20002]

How to export a 3D point cloud of the scene geometry from NeRFRAC?

Hi, and thank you for sharing this impressive work and code!

I would like to extract a 3D point cloud from a trained NeRFRAC model that represents the scene geometry below the refractive surface — not just the surface itself.

I have read Section 4.4 of the paper and understand that the geometry is modeled by sampling points along the refracted rays beneath the surface (x = Xs + t * v′), and that radiance and density are predicted at those sample points.

However, I'm having difficulty identifying the correct variables and method in the codebase to extract this. Most of my attempts result only in the flat refractive interface, or distorted point clouds that do not resemble the true scene (like the cow shown in the paper).

Request

Could you please provide guidance on:

• How to correctly extract 3D points that represent the actual underwater scene geometry?
• Which variables or parts of the code should be used for sampling and exporting these points?
• Whether a script or minimal working example exists for exporting the scene as a .ply or similar point cloud?

Any help or clarification would be highly appreciated.

Thanks again for making this work available

Best regards!

Thank you for your interest! I think you are trying to extract the geometry of the radiance fields, , which is a challenging task. Our work is based on vanilla NeRF, without precise modeling of underwater geometric surfaces, and the implicit representation of radiance fields further makes geometry extraction less intuitive.

However, there are some potential approaches to extracting coarse point clouds from radiance fields, such as sampling points and binarizing the density using a threshold. For example, you can uniformly sample a number of points (x,y,z) within a bounding box enclosing the scene and obtain their densities through the MLP. By setting a density threshold (e.g., 0.5), the points with densities above the threshold can be selected as the desired underwater scene points.

Please also note that the 3D points extracted in this way are in NDC space, and the transformation from NDC to real-scale 3D space is not Euclidean. You can find the function ndcp2camp in run_nerfrac_helpers.py to convert from NDC to camera coordinates.

Good luck!