0
$\begingroup$

I'm working (as a Post-doc) on an archaeological excavation point cloud dataset with over 2.5 Billion points. This points come from a trench, a cuboid 10 x 10 x 3 m. Each point cloud is a layer, the gaps between are the excavated volumes. There are 444 volumes from this trench, 700 individual point clouds.

Can anyone give me some direction to any deep learning algorithms which can mesh these empty spaces? I'm already doing this semi-automatically using Open3D and other python libraries, but if we could train the program to assess all the point clouds and deduce the volumes it would save us a lot of time and hopefully get better results.

We excavate more trenches every year so the perfect place to test these out.

reposted from here https://stackoverflow.com/questions/58757853/deep-learning-for-3d-point-clouds-volume-detection-and-meshing

To clarify the nature of the data, as requested.

We gather a point cloud of the ground surface using Structure from Motion - photogrammetry. We then excavate a layer (e.g remove 10cm of soil across the trench) and record again. This process repeats until we hit bedrock, in this trench it is 3m deep. This is a vastly simplified example since we have fire pits, walls, and irregularly shaped features. each point cloud is between 1 - 20 million points, we are currently using a 1 - 10 % sample of this data and getting good results, however we would like to employ deep learning to eventually save time and produce more accurate results.

Improve this question
$\endgroup$
13
  • 1
    $\begingroup$ Are you sure this is the right place to ask about "AI" but then i suspect that you need no "AI" $\endgroup$
    – joojaa
    Commented Nov 8, 2019 at 15:55
  • $\begingroup$ In another post I was recommended this group, apologies if this is the wrong place I'm not after "AI" just machine learning / deep learning direction, I don't want to reinvent the wheel. $\endgroup$
    – Spatial Digger
    Commented Nov 8, 2019 at 22:08
  • 1
    $\begingroup$ Thats fine, however you may consider explsining the nature of the data better. $\endgroup$
    – joojaa
    Commented Nov 9, 2019 at 6:21
  • 1
    $\begingroup$ sounds like a XY problem. But your not describing yourself very well. In either case if you persist in asking deep learning solutions then you are in the wrong forum. Dont specify the solution just describe the problem, youll get better answers. Why would there be overlapping points? I can guess but . Volume classifying should be trivially just the distance to a datum of some kind or just each pointcloud. Outlying datapoints should be handled by the poissob disk surface reconstruction. Anyway i am afraid you have consumed any interest I have in your problem. $\endgroup$
    – joojaa
    Commented Nov 11, 2019 at 6:15
  • 1
    $\begingroup$ Is there any specific reason you need to do this via deep learning other than it being da next big thing or you being academically pressured into doing something with deep learning? $\endgroup$
    – Christian Rau
    Commented Nov 12, 2019 at 10:25

1 Answer 1

Reset to default
1
$\begingroup$

This is just food for thought:

One potential way to analyze the data is by converting the point clouds to soil density clouds by looking at the distances between points at each excavation level. Potentially the soil density could be inferred by comparing nearest points, points that have a lot of distance to the next level have lower density, points that are close to the next level have higher density.

Using a density cloud would make it possible to infer densities between layers and extrapolate data in unexcavated regions, it might even be possible to compute an error metric by having the algorithm extrapolate data, then comparing that to measured data.

The reason I am going on and on about density is that is how MRI scans allow us to view different structures. Blood vessels have a different density then other tissues, so we can examine only blood vessels by isolating only that density in the density cloud generated by a scan. (think marching cubes algorithm here) So by isolating density ranges in a generated density map, different structures could potentially be visualized.

Improve this answer
$\endgroup$
1

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.