Difference between voxel and plane-based Registration
-
- I have made 100> posts
- Posts: 102
- Joined: Tue Feb 28, 2017 7:49 am
- 7
- Full Name: Julian Weber
- Company Details: Stadt Reutlingen
- Company Position Title: director of engineering surveying
- Country: Germany
- Has thanked: 1 time
- Been thanked: 5 times
Difference between voxel and plane-based Registration
Hello, everyone,
I would be interested in the difference between voxel and plane-based registration.
Can the bundle block adjustment converge in voxel based registration or is it just the ICP procedure with several combined point sets?
What is more accurate?
I would be interested in the difference between voxel and plane-based registration.
Can the bundle block adjustment converge in voxel based registration or is it just the ICP procedure with several combined point sets?
What is more accurate?
- smacl
- Global Moderator
- Posts: 1409
- Joined: Tue Jan 25, 2011 5:12 pm
- 13
- Full Name: Shane MacLaughlin
- Company Details: Atlas Computers Ltd
- Company Position Title: Managing Director
- Country: Ireland
- Linkedin Profile: Yes
- Location: Ireland
- Has thanked: 627 times
- Been thanked: 657 times
- Contact:
Re: Difference between voxel and plane-based Registration
For an excellent overview of the pros and cons of different registration methods, I'd recommend reading Daniel Wujanz' taming errors series of articles. From what I've read, planar methods can be very strong but they obviously require enough planes to work, so not much good inside a circular tunnel or irregular cavern for example. Similarly, long linear jobs such as road or rails will need external control. Basically you pick your most suitable method based on what you're measuring allowing sufficient redundancy to give the method the best chance of working and to enable decent QA.
-
- I have made 100> posts
- Posts: 102
- Joined: Tue Feb 28, 2017 7:49 am
- 7
- Full Name: Julian Weber
- Company Details: Stadt Reutlingen
- Company Position Title: director of engineering surveying
- Country: Germany
- Has thanked: 1 time
- Been thanked: 5 times
Re: Difference between voxel and plane-based Registration
Hey Shane, thanks for the answer .smacl wrote: ↑Sat Jun 27, 2020 2:35 pm For an excellent overview of the pros and cons of different registration methods, I'd recommend reading Daniel Wujanz' taming errors series of articles. From what I've read, planar methods can be very strong but they obviously require enough planes to work, so not much good inside a circular tunnel or irregular cavern for example. Similarly, long linear jobs such as road or rails will need external control. Basically you pick your most suitable method based on what you're measuring allowing sufficient redundancy to give the method the best chance of working and to enable decent QA.
I read this article and i use Scantra for everything now. I did a long tunnel survey using planes and spheres for registration. I Think this combination is still much more accurate than only c2c. so voxel rogistration is simply a c2c-method? What is the benefit of using voxels instead of the full Point cloud?
And one more thing i always think about and struggle with:
Is there a way to get the center of a Target as accurate as possible? because measured with a total station, targets have a pretty good accuracy (<1mm) but by digitizing the center automatically/manually we have an accuracy between 5-8mm depending of the choosen scan-resolution.
To get more accurate results i thought about inventing a cube which scantra can use to calculate the center by plane sectioning.
- smacl
- Global Moderator
- Posts: 1409
- Joined: Tue Jan 25, 2011 5:12 pm
- 13
- Full Name: Shane MacLaughlin
- Company Details: Atlas Computers Ltd
- Company Position Title: Managing Director
- Country: Ireland
- Linkedin Profile: Yes
- Location: Ireland
- Has thanked: 627 times
- Been thanked: 657 times
- Contact:
Re: Difference between voxel and plane-based Registration
For all things Scantra, @Daniel Wujanz is the person to talk to and posts regularly enough on this forum. If you don't get a reply here I'd recommend dropping him a message via LinkedIn, https://www.linkedin.com/in/daniel-wujanz-4100b1150/
For any of the tunnel jobs I've been involved in, scan control has been done externally via total station, sometimes augmented with gyro and/or levelling. The scanner, typically P40 or similar, then uses this control for setup and orientation. C2C can give gross errors on tunnels due to their regularity and long linear geometry with no opportunity to measure sufficient redundancy / overlap from multiple setups.
For any of the tunnel jobs I've been involved in, scan control has been done externally via total station, sometimes augmented with gyro and/or levelling. The scanner, typically P40 or similar, then uses this control for setup and orientation. C2C can give gross errors on tunnels due to their regularity and long linear geometry with no opportunity to measure sufficient redundancy / overlap from multiple setups.
-
- I have made 100> posts
- Posts: 102
- Joined: Tue Feb 28, 2017 7:49 am
- 7
- Full Name: Julian Weber
- Company Details: Stadt Reutlingen
- Company Position Title: director of engineering surveying
- Country: Germany
- Has thanked: 1 time
- Been thanked: 5 times
Re: Difference between voxel and plane-based Registration
Yes I am in contact with him, but i think it would be beneficial for everyone to have those answers here in the Forum and I don't want to ask him so many Questions every day because he has more important things to do I think .
Why is there no Section for Scantra in LSF?
You are completely right with the P40 or also the SX10 would be good for those purposes. Problem here is, that both are soooo slow because these are impulse-based scanners.
Julian
Why is there no Section for Scantra in LSF?
You are completely right with the P40 or also the SX10 would be good for those purposes. Problem here is, that both are soooo slow because these are impulse-based scanners.
Julian
- Daniel Wujanz
- V.I.P Member
- Posts: 227
- Joined: Fri Aug 24, 2018 11:26 am
- 5
- Full Name: Daniel Wujanz
- Company Details: technet GmbH
- Company Position Title: 3D Laser Scanning Specialist
- Country: Germany
- Linkedin Profile: Yes
- Location: Berlin
- Has thanked: 206 times
- Been thanked: 188 times
- Contact:
Re: Difference between voxel and plane-based Registration
Dear Julian,
I'm not aware of "the" voxel-based registration approach (yet, I don't know everything) even though I've used octrees myself to sample corresponding points in the context of cloud2cloud algorithms (see reference below).
Can you tell me where you've found the phrase "voxel-based registration"?
Regarding tunneling: EVERY registration algorithm will have problems (as Shane mentioned) because some degrees of freedom are typically hard to determine (usually the longitudinal one). Consequently, the object is to blame and not the algorithm ; ) We're currently working on this problem in R&D yet it is not certain if this approach will work - that's why it's called research!
All the best
Daniel
Reference:
Wujanz, D., Krueger, D., & Neitzel, F. (2016). Identification of stable areas in unreferenced laser scans for deformation measurement. The Photogrammetric Record, 31(155), 261-280.
I'm not aware of "the" voxel-based registration approach (yet, I don't know everything) even though I've used octrees myself to sample corresponding points in the context of cloud2cloud algorithms (see reference below).
Can you tell me where you've found the phrase "voxel-based registration"?
Regarding tunneling: EVERY registration algorithm will have problems (as Shane mentioned) because some degrees of freedom are typically hard to determine (usually the longitudinal one). Consequently, the object is to blame and not the algorithm ; ) We're currently working on this problem in R&D yet it is not certain if this approach will work - that's why it's called research!
All the best
Daniel
Reference:
Wujanz, D., Krueger, D., & Neitzel, F. (2016). Identification of stable areas in unreferenced laser scans for deformation measurement. The Photogrammetric Record, 31(155), 261-280.
- smacl
- Global Moderator
- Posts: 1409
- Joined: Tue Jan 25, 2011 5:12 pm
- 13
- Full Name: Shane MacLaughlin
- Company Details: Atlas Computers Ltd
- Company Position Title: Managing Director
- Country: Ireland
- Linkedin Profile: Yes
- Location: Ireland
- Has thanked: 627 times
- Been thanked: 657 times
- Contact:
Re: Difference between voxel and plane-based Registration
The P40 is a very different beast to the SX10 here, the P40 captures 1 million points per second where the SX10 only captures 26 thousand points per second. For scanning total stations you've also got the MS60 (30 thousand points per second) and Topcon GTL-1000 (100 thousand points per second).
- landmeterbeuckx
- V.I.P Member
- Posts: 1616
- Joined: Tue May 01, 2012 5:19 pm
- 11
- Full Name: Lieven Beuckx
- Company Details: Studiebureau Beuckx
- Company Position Title: Owner
- Country: Belgium
- Linkedin Profile: Yes
- Has thanked: 183 times
- Been thanked: 548 times
Re: Difference between voxel and plane-based Registration
voxel based c2c is used by rieglDaniel Wujanz wrote: ↑Mon Jun 29, 2020 11:08 am Dear Julian,
I'm not aware of "the" voxel-based registration approach (yet, I don't know everything) even though I've used octrees myself to sample corresponding points in the context of cloud2cloud algorithms (see reference below).
Can you tell me where you've found the phrase "voxel-based registration"?
Regarding tunneling: EVERY registration algorithm will have problems (as Shane mentioned) because some degrees of freedom are typically hard to determine (usually the longitudinal one). Consequently, the object is to blame and not the algorithm ; ) We're currently working on this problem in R&D yet it is not certain if this approach will work - that's why it's called research!
All the best
Daniel
Reference:
Wujanz, D., Krueger, D., & Neitzel, F. (2016). Identification of stable areas in unreferenced laser scans for deformation measurement. The Photogrammetric Record, 31(155), 261-280.
- Daniel Wujanz
- V.I.P Member
- Posts: 227
- Joined: Fri Aug 24, 2018 11:26 am
- 5
- Full Name: Daniel Wujanz
- Company Details: technet GmbH
- Company Position Title: 3D Laser Scanning Specialist
- Country: Germany
- Linkedin Profile: Yes
- Location: Berlin
- Has thanked: 206 times
- Been thanked: 188 times
- Contact:
Re: Difference between voxel and plane-based Registration
Thanks for clarifying, Lieven!
I'm guessing it's the described approach where points are regularly sampled by an octree before correspondences are determined via cloud2cloud. As far as I know Riegl also / optionally makes use of detected planes to improve the outcome but these are just hazy memories. Maybe somebody from Riegl is around to give a qualified answer. I can only comment how our plane-based approach in Scantra works:
- Step 1: Plane detection in the original point clouds - every single plane receives stochastics that characterize its precision. After this step we (typically) don't touch the original point cloud anymore and enter the parametric world of planes. This is why you can register thousands of scans on rubbish computers.
- Step 2: Pairwise plane matching - plane correspondences are established where the aforementioned stochastics are used to weigh individual correspondences. This helps to rule out self-betrayal where people think they only use the "best" piece of information and end up with great numbers but poor results due to extrapolative effects. In addition every registration receives individual stochastics
- Step 3: (The most important step) Block adjustment - all registrations are thrown into the big bowl of wisdom in order to spot tensions and false registrations. The block adjustment doesn't care if you chuck plane2plane, point2point, cloud2cloud or results from some crazy registration algorithm of the future at it, since all registration algorithms yield in registration parameters. Again, the stochastics are used to weigh individual registrations and help to identify outliers.
Please note: If you have no stochastics, you can not find outliers based on numbers! All you can do is guessing and staring...
Cheers
Daniel
I'm guessing it's the described approach where points are regularly sampled by an octree before correspondences are determined via cloud2cloud. As far as I know Riegl also / optionally makes use of detected planes to improve the outcome but these are just hazy memories. Maybe somebody from Riegl is around to give a qualified answer. I can only comment how our plane-based approach in Scantra works:
- Step 1: Plane detection in the original point clouds - every single plane receives stochastics that characterize its precision. After this step we (typically) don't touch the original point cloud anymore and enter the parametric world of planes. This is why you can register thousands of scans on rubbish computers.
- Step 2: Pairwise plane matching - plane correspondences are established where the aforementioned stochastics are used to weigh individual correspondences. This helps to rule out self-betrayal where people think they only use the "best" piece of information and end up with great numbers but poor results due to extrapolative effects. In addition every registration receives individual stochastics
- Step 3: (The most important step) Block adjustment - all registrations are thrown into the big bowl of wisdom in order to spot tensions and false registrations. The block adjustment doesn't care if you chuck plane2plane, point2point, cloud2cloud or results from some crazy registration algorithm of the future at it, since all registration algorithms yield in registration parameters. Again, the stochastics are used to weigh individual registrations and help to identify outliers.
Please note: If you have no stochastics, you can not find outliers based on numbers! All you can do is guessing and staring...
Cheers
Daniel
-
- I have made 10-20 posts
- Posts: 18
- Joined: Fri Apr 21, 2017 1:28 pm
- 6
- Full Name: Ananda Fowler
- Company Details: GEOHAWK GmbH
- Company Position Title: Owner
- Country: Austria
- Linkedin Profile: Yes
- Has thanked: 1 time
- Been thanked: 7 times
Re: Difference between voxel and plane-based Registration
Hi Daniel,
Late reply, and I don't work for Riegl, but I've seen a thing or two
You're close. The Voxel method simply computes the standard covariance data from voxelized samples of a pointcloud.
This provides more information than planes do, although it all comes down to implementation.
As for the registration algorithm, it is more sophisticated than ICP, but I won't give the details on this, as I believe the implementation by Riegl is still too far ahead of the competing manufacturers' implementations and therefore saying anything in public would be unprofessional.
It would indeed be interesting to see a 'scan off' of registration algorithms...
I REALLY enjoyed Neuss meetings for this. They had no concern, whatsoever, for manufacturers' feelings when doing comparisons, and openly reported results. I have to say, it was the only event I experienced that level of transparency. I miss Neuss more than all the other gatherings...
As with anything, though; garbage in, many hours spent trying to turn garbage into mozart... (yeah, it doesn't rhyme... )
Late reply, and I don't work for Riegl, but I've seen a thing or two
You're close. The Voxel method simply computes the standard covariance data from voxelized samples of a pointcloud.
This provides more information than planes do, although it all comes down to implementation.
As for the registration algorithm, it is more sophisticated than ICP, but I won't give the details on this, as I believe the implementation by Riegl is still too far ahead of the competing manufacturers' implementations and therefore saying anything in public would be unprofessional.
It would indeed be interesting to see a 'scan off' of registration algorithms...
I REALLY enjoyed Neuss meetings for this. They had no concern, whatsoever, for manufacturers' feelings when doing comparisons, and openly reported results. I have to say, it was the only event I experienced that level of transparency. I miss Neuss more than all the other gatherings...
As with anything, though; garbage in, many hours spent trying to turn garbage into mozart... (yeah, it doesn't rhyme... )