Bringing 3D ML to commercial CAD programs

[Arthurg27]

Guys, with all the magnificent work done here is it possible to start bringing together the 3D ML and the traditional CAD programs (Creo, SolidWorks, Catia, etc...)
This would be a real breakthrough in machine design world, reinforced with ML capabilities

[XiaoshuiHuang]

What do you mean? Can you explain your idea?

[Arthurg27]

1. Geometry part/assembly classification and detection
2. Identifying key characteristics of the model: size, volume, material, weight, etc.
3. Parametric and geometric search

Currently there is no such capabilities in any commercial CAD program, and there's a great need for such application, especially in companies with a big 3D dataset, it will give a chance of smart search, based on geometry and CAD key properties, and ofcourse a major move towards design automation.
As I see it the application shoud be integrated smoothly in currently commercial CAD program

[erinaldi]

It's possible that the accuracy is not high enough for commercial CAD applications.
Right now object detection is reaching 90% accuracy. Maybe a few more conferences and some new improvements will come out to reach above 95%.
Part segmentation accuracy is still below 88% on benchmark dataset.

I think it will happen in a year or so though. At least some proof of concepts tools, maybe integrated on the web (these ML methods are quite costly, computation-wise).

[selvakarna]

Hi @ Arthurg27 ,

How to train 3d Volume component ?
How to extracting 3d volume features and how to matching this features ?
How to segment 3d volume component based on 3d features ?
How to calculate scaling of 3d volume components?
How to classify the 3d volume components ?
without GPU, how to train those above question ?

can you share steps of above question ? using dicom data sets?

[Arthurg27]

Hi@selvakarna
This article on medical 3D Image Analysis will give you an insight on what can be cone with brain MRI implementing 3D CNN

(https://resolver.ebscohost.com/openurl?sid=google&auinit=K&aulast=Kamnitsas&atitle=Efficient+multi-scale+3D+CNN+with+fully+connected+CRF+for+accurate+brain+lesion+segmentation&id=doi%3a10.1016%2fj.media.2016.10.004&title=Medical+Image+Analysis&volume=36&date=2017&spage=61&site=ftf-live)

Try this one as well:
https://github.com/deepmedic/deepmedic

[selvakarna]

Hi @Arthurg27 , how to sign up above site ?

can you share ? dicom volume image training and 2d image based 3d volume segment ?
CPU training link ?

[jousefm]

It's possible that the accuracy is not high enough for commercial CAD applications.
Right now object detection is reaching 90% accuracy. Maybe a few more conferences and some new improvements will come out to reach above 95%.
Part segmentation accuracy is still below 88% on benchmark dataset.

I think it will happen in a year or so though. At least some proof of concepts tools, maybe integrated on the web (these ML methods are quite costly, computation-wise).

Any repos in particular that you know of tackling DL in the 3D/CAD space?

[itachimoonshot]

Guys, with all the magnificent work done here is it possible to start bringing together the 3D ML and the traditional CAD programs (Creo, SolidWorks, Catia, etc...)
This would be a real breakthrough in machine design world, reinforced with ML capabilities

Hey, Arthur. Found your comment from three years ago. Today i think we are at a point where this technology can be developed. Infact I am investigating few of them.