I have a 3D point cloud obtained by a image (using depth and disparity maps). This cloud is noisy and its "curves" aren't smooth. I need to match it with a synthetically generated cloud of the same object. I did some research about Registration and tested a few of its algorithms and I got good results with some clouds and very bad results with others. That's what got me wondering... Is this considered a Registration problem? If not, what kind of problem is this? Have anyone worked with anything like this before? Could you recommend me some material to study?
PS.: a sample of the first cloud is attached. the second one is the stanford bunny. Attachment: Generated Cloud
This is also closely related to object recognition, which in the world of 3D model-based perception is closely tied to pose estimation. One issue that frequently pops up in object recognition is what scales at which to ignore features; calculating features at a larger scale is a way to get rid of such noise.
Here is some reading, related to feature calculation of normals, from the PCL tutorials. If I remember rightly, these have some interesting notes about scale:
http://pointclouds.org/documentation/tutorials/normal_estimation.php#normal-estimation