Weird flying pixels in SupervoxelClustering

[taketwo]

I encountered strange artifacts when running SupervoxelClustering on some of the scenes in OSD dataset (namely: 'learn5', 'learn25', 'learn26'):

Haven't had time to look closer yet.

@jpapon Any ideas?

[jpapon]

I get the same behavior - it seems to be due to the depth transform. I'm looking into it...

[jpapon]

So this is because these clouds have lots of points at 0,0,0 - and the transform used by supervoxels is x/=z,y/=z, and log z. As you might imagine, this results in some wonky behavior. It seems that at some point the sanitizing of input got removed... I'll add it back in.

[jpapon]

This is fixed by commit c8fa843.

It was due to the bounding box getting screwed up by non-finite points in octree_adjacency after the transform. I'm not sure why they replaced nan points by points at 0,0,0 in the OSD dataset... but I've added a check to ensure that points are finite after the transform (which I should have been doing anyway).

[taketwo]

Wow, thanks for taking care of this! I noticed the (0,0,0) thing in the OSD dataset before and also have no clue why they decided to do this.

I'll leave this issue open until the fix is merged.

[jspricke]

@jpapon can you please send a pull request with the fix? Thanks!