PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation

Post any articles or web links that directly relate to laser scanning that you feel other users would deem interesting.
Post Reply
User avatar
Phil Marsh
Administrator
Administrator
Posts: 2490
Joined: Sun Aug 12, 2007 2:28 pm
16
Full Name: Phil Marsh
Company Details: Laser Scanning Forum Ltd
Company Position Title: Director
Country: UK
Skype Name: phil.h.marsh
Linkedin Profile: Yes
Location: Barnsley, UK
Has thanked: 117 times
Been thanked: 36 times
Contact:

PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation

Post by Phil Marsh »

I find this stuff fascinating...
Point cloud is an important type of geometric data structure. Due to its irregular format, most researchers transform such data to regular 3D voxel grids or collections of images. This, however, renders data unnecessarily voluminous and causes issues. In this paper, we design a novel type of neural network that directly consumes point clouds, which well respects the permutation invariance of points in the input. Our network, named PointNet, provides a unified architecture for applications ranging from object classification, part segmentation, to scene semantic parsing. Though simple, PointNet is highly efficient and effective. Empirically, it shows strong performance on par or even better than state of the art. Theoretically, we provide analysis towards understanding of what the network has learnt and why the network is robust with respect to input perturbation and corruption.

youtu.be/Cge-hot0Oc0
LASER SCANNING FORUM
[email protected]
Oatfedgoat
Honorary Member
Honorary Member
Posts: 1047
Joined: Mon Oct 06, 2008 8:31 pm
15
Full Name: Matthew McCarter
Company Details: Costain
Company Position Title: BIM Manager
Country: England
Linkedin Profile: Yes
Has thanked: 5 times
Been thanked: 4 times

Re: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation

Post by Oatfedgoat »

Interesting stuff.
For those that want to skip loads of algorithm and formula, the point cloudy bits start just after 7min 30 secs)

This very much is along the lines of what I saw at the recent RICS evening lecture from UCL.

As amazing as it is to me that there has never been a successful application of point cloud feature recognition (library based) I guess that deep learning and neural networks are what will actually crack this for our industry.

Annoying that such tests always seem to use building based data.
The biggest quick win on this would be for asset information/GIS uses which tend to be very much infrastructure based.

I shall certainly be keeping an eye on this one.
neeravbm
V.I.P Member
V.I.P Member
Posts: 136
Joined: Thu Mar 16, 2017 3:29 pm
7
Full Name: Neerav Mehta
Company Details: Indoor Intelligence
Company Position Title: CTO
Country: USA
Linkedin Profile: No
Has thanked: 1 time
Been thanked: 6 times

Re: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation

Post by neeravbm »

Hi Matthew,

Although we don't use PointNet, we have developed a software, Rep3D, which recognizes objects in a point cloud. Next week, we are releasing a Revit add-in for the same.

You mentioned that a quick win for this type of software will be in asset management / GIS domain. Can you elaborate a little on what the requirements are in asset management / GIS field and how can object recognition be used there?

Regards,
Neerav.
Neerav Mehta
CTO, Indoor Intelligence
Creators of http://scantobim.xyz and http://rep3d.com
dhirota
V.I.P Member
V.I.P Member
Posts: 958
Joined: Sun Nov 01, 2009 11:18 pm
14
Full Name: Dennis Hirota
Company Details: Sam O Hirota Inc
Company Position Title: President
Country: USA
Linkedin Profile: Yes
Location: Hawaii, USA
Has thanked: 87 times
Been thanked: 379 times

Re: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation

Post by dhirota »

This was one of the presentations at CVPR2017 that I reported on on July 25, 2017 on the LSF. LSF folks Jonathan and Eugene were very interested in the topics, but there seemed to be little interest so I quit posting additional information.

viewtopic.php?f=39&t=12054
dhirota wrote: Thu Aug 10, 2017 1:41 am Eugene and Jonathan

I forgot to mention that the DJI Spark UAV is using the Intel Neural Compute Stick AI chip to track the hand motions to control the UAV by recognizing objects.

"The Neural Compute Stick enables rapid prototyping, validation and deployment of Deep Neural Network (DNN) inference applications at the edge. Its low-power VPU architecture enables an entirely new segment of AI applications that aren't reliant on a connection to the cloud. This allows deep learning developers to profile, tune, and deploy Convolutional Neural Network (CNN) on low-power applications that require real-time inferencing."

You need a x86_64 computer running Ubuntu 16.04 | USB 2.0 Type-A port (Recommend USB 3.0) | 1GB RAM | 4GB free storage space.
Fruit_labelled_1000_563_85.jpg

I still have my Neural Computer Stick that I purchased at CVPR2017, so this is a reminder to get it working again.
neeravbm wrote: Fri Dec 15, 2017 2:39 pm Although we don't use PointNet, we have developed a software, Rep3D, which recognizes objects in a point cloud. Next week, we are releasing a Revit add-in for the same.
Neerav

Can you contact me about your Revit plug-in using my email?
You do not have the required permissions to view the files attached to this post.
Dennis Hirota, PhD, PE, LPLS
www.samhirota.com
[email protected]
User avatar
Jason Warren
Administrator
Administrator
Posts: 4224
Joined: Thu Aug 16, 2007 9:21 am
16
Full Name: Jason Warren
Company Details: Laser Scanning Forum Ltd
Company Position Title: Co-Founder
Country: UK
Skype Name: jason_warren
Linkedin Profile: No
Location: Retford, UK
Has thanked: 443 times
Been thanked: 246 times
Contact:

Re: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation

Post by Jason Warren »

Here are a couple of videos of Rep3D


youtu.be/oKVnh_SesWI


youtu.be/aJrzVzlgu64

Deep Learning on Point Sets for 3D Classification and Segmentation, definitely something to be keeping an eye on... ;)
Jason Warren
Co_Founder

Dedicated to 3D Laser Scanning
LaserScanningForum
Post Reply

Return to “Articles, Webinars & Links”