[pdal] Hello Intro

[Karl Keller]

I came across the point reduction approach on the Matlab site initially, saw it a couple of other places, and put other elements of PDAL to good use slicing and dicing my source and target scan worlds.  It's certainly fast enough - we typically have 50-100 scan worlds to align and those registrations were very fast.  So, hope was in full bloom!  I was planning on getting the transformation from the first, low density registration and then apply it to create a transformed dense target, use that transformed dense scan as source for the next registration, etc.  The transformations surprised me.  In the metadata I was getting a lot of non-converged registrations that were close but not close enough - fitness of ~.5.  Also, the transformations had multiple x,y,z components in the transformation attribute of the meta file.  I made the assumption this was the non-converged aspect of the solution, played around with it a bit to transform the dense clouds but all were bad alignments.  Anyway, I'm willing to admit that perhaps 50% of this is user error and the other 50% wrong expectations.

The vision was to have the daisy-chained registrations feeding into pipelines for producing a joined mesh, using EPT/Potree to visualize, and then get on with the analysis tasks on the merged cloud.  Parallelism seemed like a non-starter in this although I did consider doing a task breakdown with autoscaling similar to the ODM/OpenSFM approach.  For instance, register individual rows of scans in parallel, then join up the head and tail scans. But truthfully, if the whole thing took 7 hrs and was unaided - that would have been a big win.

-----Original Message-----
From: Howard Butler <howard at hobu.co> 
Sent: Wednesday, July 20, 2022 10:19 AM
To: pdal at lists.osgeo.org
Cc: Karl Keller <Karl.Keller at kci.com>
Subject: Re: [pdal] Hello Intro

Karl,

You also don't probably need full density data to compute the transforms. A common approach to this is to divide and conquer by running ICP over smaller patches. Both EPT and COPC would allow you to have a data source that is amenable to controlling how to process across all of the data.

As always, parallelization is not PDAL's prerogative. Running PDAL filters with computational complexities that vary widely based on their settings is tough. PDAL doesn't know your data and cant do that adaptation in parallel. 

Howard

> On Jul 20, 2022, at 7:22 AM, Andrew Bell <andrew.bell.ia at gmail.com> wrote:
> 
> ICP for 20M points is likely to take a really, really long time. You will probably have to do something else. You're only getting out a single transformation anyway. Perhaps you should look at the various methods provided by Open3D:
> 
> http://www.open3d.org/docs/release/tutorial/pipelines/index.html
> 
> On Tue, Jul 19, 2022 at 10:54 AM Karl Keller <Karl.Keller at kci.com> wrote:
> Hello PDAL community members.  Not sure how to get started so I’ll tell you about my interest in PDAL.
> 
>  
> 
> I’m using PDAL with the aim of building pipelines for various LIDAR processing workflows.  As a first step, I’ve been trying to use ICP registration without a lot of luck.  I’ve spent time following PJHartzell’s work on this to try and get a leg up.  If anyone has guidance in fast ways to automatically register relatively dense clouds in the 20M point range, I would be grateful.  Right now, I use sampling before registration to get a transform (quickly) and apply it to the full cloud.  Unfortunately, I’m not achieving < 300 mm accuracy and have yet to get convergence.
> 
>  
> 
> Looking forward to interacting with this group.
> 
>  
> 
> 	• Karl
>  
> 
> Karl S Keller
> 
> Senior Systems Engineer
> 
>  
> 
> <image001.png> 
> 
>  
> 
> KCI TECHNOLOGIES INC.
> 
> 936 Ridgebrook Road | Sparks, MD | 21152
> 
> Karl.Keller at kci.com
> 
> m: 540-827-7786
> 
> http://www.kci.com
> 
>  
> 
> <image002.jpg>
> 
>  
> 
> _______________________________________________
> pdal mailing list
> pdal at lists.osgeo.org
> https://lists.osgeo.org/mailman/listinfo/pdal
> 
> 
> -- 
> Andrew Bell
> andrew.bell.ia at gmail.com
> _______________________________________________
> pdal mailing list
> pdal at lists.osgeo.org
> https://lists.osgeo.org/mailman/listinfo/pdal