5 ms·

What's the point of point clouds?

by pagekicker 13y ago

mark212 13y ago
my question also. This is useful for what, exactly?
- tinco 13y ago
  Point clouds are not very useful themselves, if we had a choice we would skip them. They are the result of turning any measured data (like pictures or depth samples or whatever) into a 3d visualisation. Ideally you have collected enough data to reconstruct the surface geometry with reasonable accuracy, but realistically often you don't. Our minds are rather good at inferring surface geometry from point clouds, our algorithms not so much yet as far as I know. So sometimes (as in the case of this visualisation) it suffices to show people just the point cloud.
- gavinpc 13y ago
  Because it's there.
dagw 13y ago
Point cloud data is (relatively) cheap and easy to collect. However turning point clouds into 3D surfaces is computationally Very Hard and even the very best tools still require a lot human intervention to give acceptable results. Having good tools that are able to work with point clouds directly dramatically speeds up the time from data collection to data visualization and being able to actually use the data.
- StavrosK 13y ago
  Hmm, why is it very hard? Wouldn't just connecting the points between them yield acceptable results? I can understand if the points are disparate, but this dataset is pretty much a grid, and it looks like you could easily generate the surface from the points. What am I missing?
  - michaelmior 13y ago
    How do you decide which points to connect?
    - phireal 13y ago
      Use an unstructured grid (triangulation) to connect the elements. This is routinely used in numerical modelling (think computational fluid dynamics) to generate linked data sets from scatter data (such as this) in 2 and 3D (i.e. a surface and a volume).
      - dagw 13y ago
        That will work in about 50-70% of cases in the real world depending on the quality of your data and how sensitive you are to errors. For the remaining 30-50% you have to fix things by hand. The further problem is there is no good heuristics to separate the cases where it worked from where it didn't, so you still have to go through all the data to flag the cases which have to be fixed. Trust me, if you can come up with a way to make it work in general, you could build a very successful company around selling that software.
    - StavrosK 13y ago
      What phireal said. If it's a sparse point cloud, sure, it would be troublesome. The one in the example is pretty dense, and it's almost a grid, so just connect neighboring points together, no?
      - michaelmior 13y ago
        How do you decide which points are neighboring?
boyter 13y ago
Its actually fairly interesting technology. A while ago (years?) there was a post about a graphics engine which provided unlimited detail. You could literally zoom from 50,000 feet down to observe a single blade of grass. That technology is still to do anything but the demo's are impressive. https://www.youtube.com/user/EuclideonOfficial https://www.youtube.com/user/EuclideonOfficial

What's the point of point clouds?

mark212 13y ago

my question also. This is useful for what, exactly?

tinco 13y ago
Point clouds are not very useful themselves, if we had a choice we would skip them. They are the result of turning any measured data (like pictures or depth samples or whatever) into a 3d visualisation. Ideally you have collected enough data to reconstruct the surface geometry with reasonable accuracy, but realistically often you don't. Our minds are rather good at inferring surface geometry from point clouds, our algorithms not so much yet as far as I know. So sometimes (as in the case of this visualisation) it suffices to show people just the point cloud.
gavinpc 13y ago
Because it's there.

dagw 13y ago

Point cloud data is (relatively) cheap and easy to collect. However turning point clouds into 3D surfaces is computationally Very Hard and even the very best tools still require a lot human intervention to give acceptable results. Having good tools that are able to work with point clouds directly dramatically speeds up the time from data collection to data visualization and being able to actually use the data.

StavrosK 13y ago
Hmm, why is it very hard? Wouldn't just connecting the points between them yield acceptable results? I can understand if the points are disparate, but this dataset is pretty much a grid, and it looks like you could easily generate the surface from the points. What am I missing?
- michaelmior 13y ago
  How do you decide which points to connect?
  - phireal 13y ago
    Use an unstructured grid (triangulation) to connect the elements. This is routinely used in numerical modelling (think computational fluid dynamics) to generate linked data sets from scatter data (such as this) in 2 and 3D (i.e. a surface and a volume).
    - dagw 13y ago
      That will work in about 50-70% of cases in the real world depending on the quality of your data and how sensitive you are to errors. For the remaining 30-50% you have to fix things by hand. The further problem is there is no good heuristics to separate the cases where it worked from where it didn't, so you still have to go through all the data to flag the cases which have to be fixed. Trust me, if you can come up with a way to make it work in general, you could build a very successful company around selling that software.
  - StavrosK 13y ago
    What phireal said. If it's a sparse point cloud, sure, it would be troublesome. The one in the example is pretty dense, and it's almost a grid, so just connect neighboring points together, no?
    - michaelmior 13y ago
      How do you decide which points are neighboring?

boyter 13y ago

Its actually fairly interesting technology. A while ago (years?) there was a post about a graphics engine which provided unlimited detail. You could literally zoom from 50,000 feet down to observe a single blade of grass. That technology is still to do anything but the demo's are impressive. https://www.youtube.com/user/EuclideonOfficial https://www.youtube.com/user/EuclideonOfficial