Slooooooow ICP fine registration?

[daveg]

I'm working on the same datasets as a few weeks/months ago, and using the same workflow for registration and distances. I've always been amazed at how quickly the ICP worked even on huge data (many millions of faces in each model). I could watch the process get to 95% really quickly, then in a few minutes at most it would reach whatever improvement level I set (up to 10^-8 at times) and the results would be great. However, the last few days the process seems to be REALLY slow - the last one I tried didn't finish after 2 hrs today, but was complete in only a few minutes back in Sept.

Has anyone else noticed this? I noticed it in 2.6.2 and also 2.6.3 beta. I have also recently plugged in a second monitor, so maybe the difference is in my hardware somehow?

Thanks for any ideas!
Dave

[daniel]

Hi,

It's weird indeed. And I'm not sure why the additional screen would slow the process down... Just to be sure, you could simply move the progress dialog away from CC's main window (e.g. on the other screen for instance). And then see if it helps (but I really don't think it will!)

Have you changed the default random sampling limit maybe without noticing it?

[daveg]

Hi Daniel
Thanks for your reply. Since my first post I reverted to 2.6.1 and it is back up to top speed again, with or without the dual monitors. I haven't gone back to check with a 2.6.3 to see if a totally new install would fix it. Were there any changes between versions that could have caused this? I am fairly sure that none of the parameters were default/set differently.

[daniel]

Now that I think about it there has been a big modification but only when registering a cloud with a mesh: instead of (silently) sampling points on the mesh and do standard cloud-to-cloud registration, CC now computes real cloud-to-mesh distances.

This increases the registration accuracy but of course if the mesh is dense it can be slower. Is this your case? If yes, then you can still revert to the old way by sampling points on the mesh first and then use the ICP registration between two clouds.

[daveg]

Yes, I think this is the issue! My meshes are normally 50-100 million faces. I did notice the shortest distance computation dialog flashing on the screen in 2.6.3. The cloud-cloud work around would involve registering the cloud, then applying the matching transformation to the mesh? Could I use the vertices of the mesh as the cloud?

[daniel]

Indeed you would have to make this in two steps if you sample the points.

But if your vertices are dense enough then yes you could use the vertices I guess.