[question] implement ICP registration

Hi Martin,

I'm working on a similar pose estimation project using autoencoder, and I would like to enquire about how to implement the ICP refinement.

The situation is: I have the source point cloud A, which is generated from the object 3D model at the estimated pose. The target point cloud B, is calculated from the depth image, for example the 480*640 depth image from the Linemod dataset. My question is, the point cloud B is much bigger than A, how can I extract the points of the target object from B, and make sure B==A?

I tried to search on your code here https://github.com/DLR-RM/AugmentedAutoencoder/blob/master/auto_pose/icp/icp.py, I could only find something like "assert A.shape == B.shape", but could not get how you got A and B. Could you please help me with it?

Best,
Jianyu