Sampling with Blensor about points2surf

[ErlerPhilipp]

I've seen a setting somewhere in Blensor that automatically centers the created pointcloud in the origin. This could be the reason for the strange transformation. Disabling this is surely worth a try and would be the cleaner solution.

Otherwise, you can modify the object movement and restrict the random rotation to some axes and ranges. You would need to point the roof towards the camera and then apply max 90 degrees of rotation around the side axis.

The global subsample is there exactly for encoding the boundary of a solid. It's not perfect, of course. I did some overfitting on an incomplete point cloud with reconstruction in a full but low-res grid and was able to get all holes right where the (stupid and simple) heuristic with sign propagation would lead to large errors. So, yes, the network can learn to complete such point clouds. You may encounter under-sampling, though.

[ErlerPhilipp]

@chenzhaiyu Yes, you are right that this doesn't match. I wrote a simplification in the paper but the result is the same: an object scanned from multiple directions. The transformations applied to the object are later reverted when assembling the point cloud from the individual scans.

I tried to do it like written in the paper but even after a few days, I still couldn't figure out which transformation(s) Blensor applies to the points. It seems to be dependent on the scanner, but not simply its inverse transformation. So I went with a working simpler solution, which was to move the object instead of the camera.

If you can find a fix for this, please submit a pull-request.

[chenzhaiyu]

Thanks for your quick reply! This workaround somehow hinders what I'd like to do...

I wanted to create the sampling with certain perspective(s) excluded; imagine the MVS point cloud from a building that doesn't have the bottom view. With the way in the paper it seems easy to simulate this scenario, just by restricting the location of the scanner (e.g. at the upper part of the sphere). With the current workaround where the object is both moving and rotating randomly, do you still see an easy way to do that?

BTW, I'm also curious how can P2S handle this kind of incomplete point clouds (whose GT mesh should be solid). Specifically, do you think the global subsample is capable of encoding the coarse shape info despite the missing points?

[chenzhaiyu]

Otherwise, you can modify the object movement and restrict the random rotation to some axes and ranges. You would need to point the roof towards the camera and then apply max 90 degrees of rotation around the side axis.

This worked well for me.

I did some overfitting on an incomplete point cloud with reconstruction in a full but low-res grid and was able to get all holes right where the (stupid and simple) heuristic with sign propagation would lead to large errors. So, yes, the network can learn to complete such point clouds.

I've also trained P2S for a while on a building dataset completely without the bottom points. Indeed it's still working very well to reconstruct solid models, though it might be simply remembering where is the bottom and just appending it.

I'll dig deeper on the cleaner alternative with Blensor when having the time. Thanks again! Closing for now :)