AI + Undet + SketchUp: a different approach to point-cloud modelling

SketchUp
, ,
1

I have been thinking about how AI could realistically assist with point-cloud modelling inside SketchUp.

Rather than asking AI to interpret an entire E57 and automatically produce a complete building, I wonder if a more controlled approach could work better.

Many E57 surveys are already organised into individual scan/survey stations. In Undet, these stations can be activated or deactivated. Undet also provides Classes and ClipBox tools that allow the user to control what part of the point cloud is being displayed and worked with.

This suggests a possible workflow:

E57 / survey
↓
Undet survey stations
↓
Undet Classes
↓
Active ClipBox
↓
Controlled point-cloud information
↓
AI-assisted interpretation
↓
SketchUp / Ruby modelling
↓
QC against the original point cloud

For example, imagine modelling one window.

Instead of giving AI the entire building, we could:

  1. Activate the stations that have a good view of the window.
  2. Activate the relevant Undet Class(es).
  3. Define a ClipBox around the opening.
  4. Tell the AI: “Work only with the currently active point-cloud information.”
  5. Determine the opening dimensions and geometry.
  6. Generate the SketchUp model using Ruby.
  7. Compare the result against the point cloud.
  8. Correct any discrepancies.

The same principle could potentially be applied progressively to walls, doors, windows, rooms, façades and eventually larger parts of a building.

I have already been experimenting with the SketchUp/Ruby side of this idea.

For example, I have used ChatGPT to construct a detailed window assembly step by step: frame, double glazing, spacers, warm-edge material, gasket and materials, with the geometry organised into groups.

The interesting question for me now is whether the point-cloud side can be connected to this type of AI-assisted modelling.

I am also interested in how Gaussian Splats could complement the workflow. They provide another way of capturing and understanding the existing environment, while the point cloud can remain the more rigorous geometric reference.

So I am not thinking about:

“AI replaces Undet.”

I am thinking about:

“AI works with Undet.”

The goal would be an AI-assisted architectural modelling workflow in which the human controls the information being used, AI assists with interpretation and repetitive modelling, and the final geometry can be checked against the measured point cloud.

I would be very interested to hear from people working with Undet, point clouds, SketchUp Ruby and/or AI:

Could this approach be technically feasible with the tools that already exist?

And what would be the most useful API/workflow capabilities to make it possible?

“La connaissance n’a d’intérêt que si on la partage.”
— attribution uncertain

1 Like