Faro: understanding the cloud to cloud registration settings
-
- I have made 80-90 posts
- Posts: 85
- Joined: Wed Apr 29, 2015 8:31 am
- 8
- Full Name: Emanuele Ferrari
- Company Details: Studio FERRARI
- Company Position Title: Titolare
- Country: Italia
- Linkedin Profile: Yes
- Location: Castellamonte (TO)
- Has thanked: 9 times
- Been thanked: 15 times
Re: Faro: understanding the cloud to cloud registration settings
Hello,
I "resume" the post because I can't understand the meaning of the Maximun search distance parameter!
In various courses / webinars it is defined as the radius of the sphere (in the center of the scanner) within which Scene searches for the points to overlap and on which to search for correlation; setting it to 20 m excludes all external points (perhaps placed at 100 m).
If I reduce the distance to 50 cm it would mean that there may be no overlapping points !!
In other posts I have read (and I have understood) that the parameter in question defines the maximum distance between the points to try to find the correlation so the value should be reduced to 2-5 cm; in this case, Scene could try to correlate points 50-100 m away from the scanner.
Sorry but I'm confused !!
I "resume" the post because I can't understand the meaning of the Maximun search distance parameter!
In various courses / webinars it is defined as the radius of the sphere (in the center of the scanner) within which Scene searches for the points to overlap and on which to search for correlation; setting it to 20 m excludes all external points (perhaps placed at 100 m).
If I reduce the distance to 50 cm it would mean that there may be no overlapping points !!
In other posts I have read (and I have understood) that the parameter in question defines the maximum distance between the points to try to find the correlation so the value should be reduced to 2-5 cm; in this case, Scene could try to correlate points 50-100 m away from the scanner.
Sorry but I'm confused !!
- 3DForensics
- Honorary Member
- Posts: 1979
- Joined: Mon Aug 03, 2009 1:52 am
- 14
- Full Name: Eugene Liscio
- Company Details: AI2-3D Forensics
- Company Position Title: Owner
- Skype Name: eliscio
- Location: Toronto, Canada
- Has thanked: 13 times
- Been thanked: 70 times
- Contact:
Re: Faro: understanding the cloud to cloud registration settings
Assume two point clouds have only 30% overlap...Scene will keep looking for points in the non-overlap areas up to that max search distance and then move on. This could take time.
If your points clouds are already roughly registered, there is no reason to have a value like 10m and you can reduce it to something smaller like 30 cm.
At least this was my understanding but hopefully I will be corrected or clarified...
Eugene
If your points clouds are already roughly registered, there is no reason to have a value like 10m and you can reduce it to something smaller like 30 cm.
At least this was my understanding but hopefully I will be corrected or clarified...
Eugene
-
- I have made 80-90 posts
- Posts: 85
- Joined: Wed Apr 29, 2015 8:31 am
- 8
- Full Name: Emanuele Ferrari
- Company Details: Studio FERRARI
- Company Position Title: Titolare
- Country: Italia
- Linkedin Profile: Yes
- Location: Castellamonte (TO)
- Has thanked: 9 times
- Been thanked: 15 times
Re: Faro: understanding the cloud to cloud registration settings
I thank you for the answer.3DForensics wrote: ↑Fri Sep 25, 2020 6:01 pm Assume two point clouds have only 30% overlap...Scene will keep looking for points in the non-overlap areas up to that max search distance and then move on. This could take time.
If your points clouds are already roughly registered, there is no reason to have a value like 10m and you can reduce it to something smaller like 30 cm.
At least this was my understanding but hopefully I will be corrected or clarified...
Eugene
Surely the scans have been pre-aligned.
Following your reasoning, the parameter manages the maximum distance between the points, coming from two different scans, on which to try to search for the overlap.
In the case of external scans at high resolution and at a distance of 15-20 m between them, it would mean that Scene tries to correlate points placed perhaps at 100 m.
I don't understand your statement: "Scene will keep looking for points in the non-overlap areas"
If the parameter defined the maximum distance from the scanner of the points to be used absurdly they could be zero (scanner positioned in the center of the road at 3 m high) then c2c would not make sense
- 3DForensics
- Honorary Member
- Posts: 1979
- Joined: Mon Aug 03, 2009 1:52 am
- 14
- Full Name: Eugene Liscio
- Company Details: AI2-3D Forensics
- Company Position Title: Owner
- Skype Name: eliscio
- Location: Toronto, Canada
- Has thanked: 13 times
- Been thanked: 70 times
- Contact:
Re: Faro: understanding the cloud to cloud registration settings
All I meant was that Scene doesn't know if there is overlap or not. A point on the reference scan is the starting point and Scene begins to look for neighbouring points in the "align" scan. Once it reaches the search limit, without finding a point to align to, it moves on to the next point. If the scans are 10m apart and you have a max search distance of 1m, then you have a problem.
Not sure if that is clearer....
Eugene
Not sure if that is clearer....
Eugene
-
- V.I.P Member
- Posts: 1237
- Joined: Mon Jan 04, 2010 7:51 pm
- 14
- Full Name: Jed Frechette
- Company Details: Lidar Guys
- Company Position Title: CEO and Lidar Supervisor
- Country: USA
- Linkedin Profile: Yes
- Location: Albuquerque, NM
- Has thanked: 62 times
- Been thanked: 220 times
- Contact:
Re: Faro: understanding the cloud to cloud registration settings
Cloud to cloud in Scene is some variation of an Iterative Closest Point algorithm:
https://en.wikipedia.org/wiki/Iterative_closest_point
Max search distance is an additional constraint on Step 1 as described in that article: If no point in the reference cloud is found within max_distance of the current point in the source cloud, stop searching for a match and move on to the next point in the source cloud.
In practice, I would never use a max distance of 10 meters. For this type of algorithm to work reliably you really need to prealign the scans pretty accurately to start with. 1 m is about the largest max distance I ever use and final alignments are typically done with a max distance of 0.1 m.
https://en.wikipedia.org/wiki/Iterative_closest_point
Max search distance is an additional constraint on Step 1 as described in that article: If no point in the reference cloud is found within max_distance of the current point in the source cloud, stop searching for a match and move on to the next point in the source cloud.
In practice, I would never use a max distance of 10 meters. For this type of algorithm to work reliably you really need to prealign the scans pretty accurately to start with. 1 m is about the largest max distance I ever use and final alignments are typically done with a max distance of 0.1 m.
Jed
- Leandre Robitaille
- V.I.P Member
- Posts: 542
- Joined: Sat Aug 03, 2019 1:53 am
- 4
- Full Name: Leandre Robitaille
- Company Details: Cima+
- Company Position Title: Civil Technician - Surveyor
- Country: Canada
- Has thanked: 61 times
- Been thanked: 246 times
Re: Faro: understanding the cloud to cloud registration settings
Wow I am extremely gratefull this post was brought back to the front page. I have been using cloud2cloud wrong for the past 4 years. I always tought search distance meant "use points in this radius around the scanner". For exemple if I had scans 25m appart I could use a 25m search radius for good overlap. It makes a lot more sens to have the algorythm work in this way and I am convinced knowing this will make my life easier as I was finding c2c registration to be like gambling...especially with the search radius I was using.
But now, does this means it uses all the points of the scanner? There is lost of accuracy at long range, so If I was to survey with a s350 and would use it in a c2c, I could potentially get worst cloudtocloud results than using a s70? Lets say I would have scans 350ms away and would do a c2c,the points from that other scan 350away would weight in on the registration.
Anyhow this is good to know and I will work with this from now on.
But now, does this means it uses all the points of the scanner? There is lost of accuracy at long range, so If I was to survey with a s350 and would use it in a c2c, I could potentially get worst cloudtocloud results than using a s70? Lets say I would have scans 350ms away and would do a c2c,the points from that other scan 350away would weight in on the registration.
Anyhow this is good to know and I will work with this from now on.
- landmeterbeuckx
- V.I.P Member
- Posts: 1615
- 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: Faro: understanding the cloud to cloud registration settings
This is an excellent thread, i'll try it on some older projects.
-
- V.I.P Member
- Posts: 1237
- Joined: Mon Jan 04, 2010 7:51 pm
- 14
- Full Name: Jed Frechette
- Company Details: Lidar Guys
- Company Position Title: CEO and Lidar Supervisor
- Country: USA
- Linkedin Profile: Yes
- Location: Albuquerque, NM
- Has thanked: 62 times
- Been thanked: 220 times
- Contact:
Re: Faro: understanding the cloud to cloud registration settings
Even with the "right" search radius it still kind of is. I've found that the most important factor for gettng really good c2c results is having good ways to visualize them so you can quickly identify and correct problems. Ultimately, it's just an optimization problem and one that is susceptible to getting stuck in local minimum. Just dumping 100 scans in to a c2c alignment and hoping for the best often doesn't provide optimal results.Leandre Robitaille wrote: ↑Sat Sep 26, 2020 4:01 amI was finding c2c registration to be like gambling
Without knowing exactly how Faro has implemented their algorithm that's a little hard to predict a priori. However, I tend to be more worried about the points close to the scanner. Even though slightly less accurate the distant points provide a much longer baseline. If you just rely on nearby points you might get a good fit close to the scanner but when you extrapolate that out it is likely that the far points will be badly misaligned due to magnification of small angular misalignments. If 100% of the points in the scan are used the same type of effect can also happen because there are so many more points close to the scanner than there are far away. The point density essentially ends up weighting the nearby geometry much higher than the far away geometry.Leandre Robitaille wrote: ↑Sat Sep 26, 2020 4:01 amBut now, does this means it uses all the points of the scanner? There is lost of accuracy at long range, so If I was to survey with a s350 and would use it in a c2c, I could potentially get worst cloudtocloud results than using a s70? Lets say I would have scans 350ms away and would do a c2c,the points from that other scan 350away would weight in on the registration.
The software we use to do c2c combats this by normalizing point density within range bins. So for example, from 1-10 m from the scanner points are 1 cm apart, from 10-30 m 2 cm, etc.
Jed
-
- I have made 80-90 posts
- Posts: 85
- Joined: Wed Apr 29, 2015 8:31 am
- 8
- Full Name: Emanuele Ferrari
- Company Details: Studio FERRARI
- Company Position Title: Titolare
- Country: Italia
- Linkedin Profile: Yes
- Location: Castellamonte (TO)
- Has thanked: 9 times
- Been thanked: 15 times
Re: Faro: understanding the cloud to cloud registration settings
Thanks everyone for the answers!
The pre-alignment (top view or even with "manual" targets) leads to deviations generally lower than 2-10 cm.
Following Jed's reasoning, the Maximun search distance is therefore the distance within which there is the probability of having the homologous point between the reference scan and the one to be aligned; this would explain the need to iterate the c2c starting from the value of 100 cm and going down to 5-10 cm also in relation to the subsampling (it might be useless to use a radius of 2 cm if the subsampling is 10 cm).
Leandre Robitaille's observation is correct: in theory distant points (above 50 m) have a greater angular error than neighboring points but for very distant points I think subsampling and the density of points are preponderant, which decrease the probability of correlating wrong points .
The usefulness of using distant points is precisely the possibility of limiting the curvature of the scan chain due only to nearby points.
I believe it is the default setting of Scene (10m) and the lack of instructions has led to incorrect interpretation of the parameter.
The doubt arose again because in the last road work I aligned 72 scans (divided into clusters of 10-12 scans) using the search parameter at 30 m and the results were very good! Both the tensions given by Scene and the visual inspection led to differences of less than 5 mm on the overlaps!
It would be the case that the developers of Scene would give us some more information !!
The pre-alignment (top view or even with "manual" targets) leads to deviations generally lower than 2-10 cm.
Following Jed's reasoning, the Maximun search distance is therefore the distance within which there is the probability of having the homologous point between the reference scan and the one to be aligned; this would explain the need to iterate the c2c starting from the value of 100 cm and going down to 5-10 cm also in relation to the subsampling (it might be useless to use a radius of 2 cm if the subsampling is 10 cm).
Leandre Robitaille's observation is correct: in theory distant points (above 50 m) have a greater angular error than neighboring points but for very distant points I think subsampling and the density of points are preponderant, which decrease the probability of correlating wrong points .
The usefulness of using distant points is precisely the possibility of limiting the curvature of the scan chain due only to nearby points.
I believe it is the default setting of Scene (10m) and the lack of instructions has led to incorrect interpretation of the parameter.
The doubt arose again because in the last road work I aligned 72 scans (divided into clusters of 10-12 scans) using the search parameter at 30 m and the results were very good! Both the tensions given by Scene and the visual inspection led to differences of less than 5 mm on the overlaps!
It would be the case that the developers of Scene would give us some more information !!
-
- I have made 20-30 posts
- Posts: 26
- Joined: Tue Oct 02, 2018 1:08 pm
- 5
- Full Name: Konstantin
- Company Details: No longer at FARO
- Company Position Title: Former Software Engineer at FARO
- Country: Germany
- Linkedin Profile: Yes
- Has thanked: 4 times
- Been thanked: 6 times
Re: Faro: understanding the cloud to cloud registration settings
Hello,
for your information, the statement
Kind regards
Konstantin
for your information, the statement
is wrong. And the advice given by the userjedfrechette wrote: ↑Fri Sep 25, 2020 8:04 pm Cloud to cloud in Scene is some variation of an Iterative Closest Point algorithm:
https://en.wikipedia.org/wiki/Iterative_closest_point
is not necessary correct in all situations either. While a good pre-alignment is necessary for SCENE's Cloud-2-Cloud algorithm to produce reliable and accurate results, it does not mean, that a high search radius cannot be used to create a pre-alignment itself. (Although, Top-View-Based Registration should be considered first.)jedfrechette wrote: ↑Fri Sep 25, 2020 8:04 pm In practice, I would never use a max distance of 10 meters. For this type of algorithm to work reliably you really need to prealign the scans pretty accurately to start with. 1 m is about the largest max distance I ever use and final alignments are typically done with a max distance of 0.1 m.
Kind regards
Konstantin