We dream a magic button for 3-D point cloud processing
- Joon
- V.I.P Member
- Posts: 335
- Joined: Wed Aug 21, 2013 8:01 pm
- 10
- Full Name: Sung Joon Ahn
- Company Details: CurvSurf
- Company Position Title: Founder + CEO
- Country: Republic of Korea
- Linkedin Profile: Yes
- Location: Seongnam-si, Korea
- Has thanked: 7 times
- Been thanked: 35 times
- Contact:
Re: We dream a magic button for 3-D point cloud processing
Region Growing
The shape and the initial parameter values of object surface are determined by the local curvature analysis of a small patch (i.e., initial inlier points) of point cloud. Then, the initial parameter values are refined by ODF (orthogonal distance fitting) to the small patch of point cloud.
Next, based on the current shape parameter values determined by ODF, the region growing method gathers the inlier points. The activity level of region growing is controlled in two orthogonal directions.
The radial expansion controls the thickness of the ROI (region of interest) and the lateral extension the length/width of the ROI.
Then, the model fitting (ODF) to the current inlier points updates the shape parameter values.
The region growing and the model fitting are being repeated alternately, as long as no break condition is fulfilled.
The break conditions include:
- Rms error of model fitting
- Parameter reliabilities (covariances, correlations)
- Density of inlier points, etc.
The program control parameters include:
- Point measuring accuracy of the input point cloud
- Mean distance between adjacent points
- Size of the point cloud patch
- Degree of the region growing.
You can test the 3-D point cloud processing of CurvSurf by using a web-browser.
FindSurface for Web - SignIn
FindSurface for Web - Manual
If you like to test the FindSurface for Web yourself, please email to 'info at curvsurf.com' by using your company/university email address.
youtu.be/oKlxI2r2oWU
Joon
The shape and the initial parameter values of object surface are determined by the local curvature analysis of a small patch (i.e., initial inlier points) of point cloud. Then, the initial parameter values are refined by ODF (orthogonal distance fitting) to the small patch of point cloud.
Next, based on the current shape parameter values determined by ODF, the region growing method gathers the inlier points. The activity level of region growing is controlled in two orthogonal directions.
The radial expansion controls the thickness of the ROI (region of interest) and the lateral extension the length/width of the ROI.
Then, the model fitting (ODF) to the current inlier points updates the shape parameter values.
The region growing and the model fitting are being repeated alternately, as long as no break condition is fulfilled.
The break conditions include:
- Rms error of model fitting
- Parameter reliabilities (covariances, correlations)
- Density of inlier points, etc.
The program control parameters include:
- Point measuring accuracy of the input point cloud
- Mean distance between adjacent points
- Size of the point cloud patch
- Degree of the region growing.
You can test the 3-D point cloud processing of CurvSurf by using a web-browser.
FindSurface for Web - SignIn
FindSurface for Web - Manual
If you like to test the FindSurface for Web yourself, please email to 'info at curvsurf.com' by using your company/university email address.
youtu.be/oKlxI2r2oWU
Joon
You do not have the required permissions to view the files attached to this post.
- Joon
- V.I.P Member
- Posts: 335
- Joined: Wed Aug 21, 2013 8:01 pm
- 10
- Full Name: Sung Joon Ahn
- Company Details: CurvSurf
- Company Position Title: Founder + CEO
- Country: Republic of Korea
- Linkedin Profile: Yes
- Location: Seongnam-si, Korea
- Has thanked: 7 times
- Been thanked: 35 times
- Contact:
Re: We dream a magic button for 3-D point cloud processing
Future Developments (1)
RGB image + point cloud processing
As you may have noticed, in order to extract a geometric object primitive from a point cloud, an appropriate gazing point (interest point) must be selected among the input point cloud. A gazing point can be selected manually by mouse-clicking a screen point or by directing the measuring device onto the object surface.
Now, we will be extracting automatically all probable geometric object primitives from a scene. We utilize that the information of 2-D image and 3-D point cloud are inherently mutually orthogonal, delivering complementary information. 3-D surfaces can be determined by 3-D point cloud processing, of which boundaries can be determined by 2-D image processing.
The RGB image segmentation finds all probable object locations (i.e. gazing points) in a scene. At the same time, the approximate initial sizes of object (i.e. the sizes of individual patch of point cloud) can be determined from the image blobs.
Joon
RGB image + point cloud processing
As you may have noticed, in order to extract a geometric object primitive from a point cloud, an appropriate gazing point (interest point) must be selected among the input point cloud. A gazing point can be selected manually by mouse-clicking a screen point or by directing the measuring device onto the object surface.
Now, we will be extracting automatically all probable geometric object primitives from a scene. We utilize that the information of 2-D image and 3-D point cloud are inherently mutually orthogonal, delivering complementary information. 3-D surfaces can be determined by 3-D point cloud processing, of which boundaries can be determined by 2-D image processing.
The RGB image segmentation finds all probable object locations (i.e. gazing points) in a scene. At the same time, the approximate initial sizes of object (i.e. the sizes of individual patch of point cloud) can be determined from the image blobs.
Joon
You do not have the required permissions to view the files attached to this post.
- Joon
- V.I.P Member
- Posts: 335
- Joined: Wed Aug 21, 2013 8:01 pm
- 10
- Full Name: Sung Joon Ahn
- Company Details: CurvSurf
- Company Position Title: Founder + CEO
- Country: Republic of Korea
- Linkedin Profile: Yes
- Location: Seongnam-si, Korea
- Has thanked: 7 times
- Been thanked: 35 times
- Contact:
Re: We dream a magic button for 3-D point cloud processing
Future Developments (2)
On-line data processing service (SaaS)
There are a variety of 3-D measurement methods and devices including tactile, ultra-sonic, CT, stereo vision, structured light, ToF, phase-shift, SLAM, etc.
The common data format of 3-D measurement is the ‘point cloud’, an array of xyz-coordinates.
By using FindSurface for Web, we can determine the shape, size, position & orientation of object surfaces on site.
3-D Measurement with ARKit
youtu.be/DPBFOPdQNBk
Joon
On-line data processing service (SaaS)
There are a variety of 3-D measurement methods and devices including tactile, ultra-sonic, CT, stereo vision, structured light, ToF, phase-shift, SLAM, etc.
The common data format of 3-D measurement is the ‘point cloud’, an array of xyz-coordinates.
By using FindSurface for Web, we can determine the shape, size, position & orientation of object surfaces on site.
3-D Measurement with ARKit
youtu.be/DPBFOPdQNBk
Joon
You do not have the required permissions to view the files attached to this post.
- Joon
- V.I.P Member
- Posts: 335
- Joined: Wed Aug 21, 2013 8:01 pm
- 10
- Full Name: Sung Joon Ahn
- Company Details: CurvSurf
- Company Position Title: Founder + CEO
- Country: Republic of Korea
- Linkedin Profile: Yes
- Location: Seongnam-si, Korea
- Has thanked: 7 times
- Been thanked: 35 times
- Contact:
Re: We dream a magic button for 3-D point cloud processing
3-D Measurement with ARKit - Large dome
In few months, we will put a virtual airplane flying around the real dome as an example of augmented reality with occlusion.
ARKit/ARCore provide applications with the device motion and the point cloud.
CurvSurf FindSurface determines the size & position of dome by sphere fitting to point cloud.
A variety of AR applications/games could be made by utilizing the device motion and the size & position of ball.
youtu.be/vWI-dO9fsBk
In few months, we will put a virtual airplane flying around the real dome as an example of augmented reality with occlusion.
ARKit/ARCore provide applications with the device motion and the point cloud.
CurvSurf FindSurface determines the size & position of dome by sphere fitting to point cloud.
A variety of AR applications/games could be made by utilizing the device motion and the size & position of ball.
youtu.be/vWI-dO9fsBk
-
- V.I.P Member
- Posts: 543
- Joined: Wed Oct 05, 2011 10:14 am
- 12
- Full Name: Scan ku kaï
- Company Details: Scankukai
- Company Position Title: CEO
- Country: France
- Linkedin Profile: Yes
- Has thanked: 58 times
- Been thanked: 207 times
Re: We dream a magic button for 3-D point cloud processing
Hello,
I'm really impressed by what you manage to achieve and I think you're going the right way. Keep on searching!
What do you think about meshing the pointclouds, especially in buildings that are full of occlusions and sharp edges ?(industrial facilites for example)
I see many companies that are trying to go this way.
I'm really impressed by what you manage to achieve and I think you're going the right way. Keep on searching!
What do you think about meshing the pointclouds, especially in buildings that are full of occlusions and sharp edges ?(industrial facilites for example)
I see many companies that are trying to go this way.
- Joon
- V.I.P Member
- Posts: 335
- Joined: Wed Aug 21, 2013 8:01 pm
- 10
- Full Name: Sung Joon Ahn
- Company Details: CurvSurf
- Company Position Title: Founder + CEO
- Country: Republic of Korea
- Linkedin Profile: Yes
- Location: Seongnam-si, Korea
- Has thanked: 7 times
- Been thanked: 35 times
- Contact:
Re: We dream a magic button for 3-D point cloud processing
Hi Yan,Tgl_Nan wrote: ↑Sat Feb 02, 2019 2:15 pm Hello,
I'm really impressed by what you manage to achieve and I think you're going the right way. Keep on searching!
What do you think about meshing the pointclouds, especially in buildings that are full of occlusions and sharp edges ?(industrial facilites for example)
I see many companies that are trying to go this way.
Much thanks for your encouraging praises.
As I mentioned early of this posting, meshes are arbitrary artifacts that do not exist really but for the visualization of surfaces for human vision.
Computer graphics had invented the triangular meshes for the visualization of surfaces because we human recognize visually.
Machines measure the coordinates of object surface points. They do not need to visualize for themselves the points they have measured. Mathematics can process the object coordinates without visualization. The visualization of how mathematics works is only for human vision.
3-D information shuld be processed in 3-D space without projecting (i.e., visualizing) to 2-D plane.
The meshing of 3-D point cloud (i.e., connecting/differentiating 3-D point coordinate values) is inherently a process of noise amplification. Unfortunately, because the projection (visualization) of 3-D meshes onto 2-D planes is a process of integration, there may be apparently no problem (a process of differentiation followed by integration). The image of meshes may look apparently well on the 2-D plane.
Meshing is for human vision but not for mathematics.
Once, the point cloud is converted to meshes, there will have been much information loss.
Joon
You do not have the required permissions to view the files attached to this post.
- 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: We dream a magic button for 3-D point cloud processing
Hi Joon,
Firstly, let me say I'm really enjoying your posts.
I think the efficacy of meshing depends to a large degree to the context of what we're measuring. Direct extraction of geometry from observations is going to work very well where the objects we're measuring are made up of those geometries. Meshing works better where those geometries are complex and display a high degree of entropy (e.g. an old cathedral with irregular stonework and gargoyles on the walls). Also worth remembering meshes may be tetrahedral rather than triangular.
I think a hybrid approach is the best of both worlds. Extract understandable geometry and mesh what we can't extract as known geometric primitives.
Firstly, let me say I'm really enjoying your posts.
I'd disagree with this. In addition to visualisation, meshing is also used for fast linear interpolation of irregular surfaces, density analysis and adaptive meshing for data size reduction. For example, say I'm measuring the deformation of an old building over time. I scan the building facade, and mesh based on the mean plane of that facade. I then scan again at a later date and compare the distance of every scanned point to the mesh to get movement w.r.t the reference plane. Difficult to do this without meshing if my original building does not comprise of readily identifiable geometric shapes.
I think the efficacy of meshing depends to a large degree to the context of what we're measuring. Direct extraction of geometry from observations is going to work very well where the objects we're measuring are made up of those geometries. Meshing works better where those geometries are complex and display a high degree of entropy (e.g. an old cathedral with irregular stonework and gargoyles on the walls). Also worth remembering meshes may be tetrahedral rather than triangular.
I think a hybrid approach is the best of both worlds. Extract understandable geometry and mesh what we can't extract as known geometric primitives.