Wall in TSDF

[sdmiller]

Wall in TSDF

[sdmiller]

Hi @dimtziwnas,

Could you put the example files somewhere online, or e-mail them to me (my github name @ stanford.edu)? It's possible I overlooked something -- though we always work with organized data, which frequently contains NaNs. So it's pretty surprising this would occur, and seems quite arbitrary that they'd appear on a plane like this..

Cheers
-Stephen

[dimtzionas]

Hi all,

after an email discussion with Stephen and some experimentation the source of the problem was found.

There was no bug in the cpu_tsdf project itself, but there was indeed a bug/unexpected_behaviour in the PCL function pcl::FastBilateralFilterOMP that was used for pre-processing. I will report this as an issue to the PCL community (*), but in the meanwhile cpu_tsdf users should be careful of phenomena like this:

The method pcl::FastBilateralFilterOMP changes the value of NaN points in such a way, so that points
like this -> (x=NAN, y=NAN, z=NAN)
become -> (x=NAN, y=NAN, z=const)
In my case const value was const=0.898 for all points and this is why this weird wall appeared in the tsdf and the final mesh reconstruction.

So the issue is closed, but be aware that some pre-processing might result in buggy input to cpu_tsdf methods ;)

Best,
dimitris

PS - (*) - PointCloudLibrary/pcl#1027