How does unsupervised featuring compare against the best-performing deep learning methods ?

Post any articles or web links that directly relate to laser scanning that you feel other users would deem interesting.
Post Reply
User avatar
Poux
I have made 30-40 posts
I have made 30-40 posts
Posts: 30
Joined: Wed Nov 05, 2014 9:24 pm
Full Name: Florent POUX
Company Details: Geovast 3D - University of Liege
Company Position Title: CEO - TA
Country: France
Linkedin Profile: Yes
Location: France
Contact:

How does unsupervised featuring compare against the best-performing deep learning methods ?

Post by Poux » Mon May 13, 2019 12:33 pm

The task of semantic segmentation, and by extension classification is a very challeging research, especially for point clouds. Specifically, deep learning techniques are interesting for their potential high level of performance and fully-fledge applicability (learning features that work best as part of the system). However, they are very constrained by the availability of high quality training datasets, even more challenging due to the various domain specialization thus labels.

Therefore, it is interesting to know how well unsupervised - in the sense that their is no need of training data - method compare to the best-performing state of the art 3D deep learning approaches ? The recent open access article provided holds some answers...

https://www.mdpi.com/2220-9964/8/5/213 :ugeek:
Florent POUX
3D Data Scientist

tbwester
V.I.P Member
V.I.P Member
Posts: 879
Joined: Fri Apr 20, 2012 1:49 pm
Full Name: Thad Wester
Company Details: WeWork
Company Position Title: Head of Reality Capture
Country: USA
Linkedin Profile: Yes
Location: SC/NYC

Re: How does unsupervised featuring compare against the best-performing deep learning methods ?

Post by tbwester » Mon May 13, 2019 5:28 pm

Its interesting thanks for sharing.

We are scanning so much, I know our data would be valuable to these researchers. At the end of this year we will have over 30,000 floors (entire FLOORS), of point clouds and 3D revit models. Seems this is the training dataset they need.

User avatar
Poux
I have made 30-40 posts
I have made 30-40 posts
Posts: 30
Joined: Wed Nov 05, 2014 9:24 pm
Full Name: Florent POUX
Company Details: Geovast 3D - University of Liege
Company Position Title: CEO - TA
Country: France
Linkedin Profile: Yes
Location: France
Contact:

Re: How does unsupervised featuring compare against the best-performing deep learning methods ?

Post by Poux » Wed May 15, 2019 9:35 am

Hey Thad,

That indeed would be very valuable !! Especially if one wants to directly have a mapping between point cloud and BIM elements !
We could set up a POC over your data to see how good it can perform once you have these high quality datasets.

Especially, what is important is your "expertise" and interpretation of the point cloud in correct BIM elements, to insure the data quality translate into the learning model quality !

Cheers,
Florent POUX
3D Data Scientist

Scott
V.I.P Member
V.I.P Member
Posts: 754
Joined: Tue Mar 29, 2011 7:39 pm
Full Name: Scott Page
Company Details: Scott Page Design- Architectural service
Company Position Title: Owner
Country: USA
Linkedin Profile: No
Contact:

Re: How does unsupervised featuring compare against the best-performing deep learning methods ?

Post by Scott » Wed May 15, 2019 3:22 pm

Sounds like a marriage in heaven --keep us posted.

Post Reply

Return to “Articles, Webinars & Links”