I am using
vtkIterativeClosestPointTransform in a pipeline that detects radiopaque stereotactic frame fiducials in a CT/MRI. The geometry of the stereotactic frame is know. For a single CT volume there could be ~700-800 markups points, which forms the stereotactic frame localizers. The image below shows all the points that ae found within a CT volume.
When running the ICP registration I convert all the CT markups points to polydata and generate an equal number of points for the frame. The top and bottom of each localizer bar have known coordinates and the bars have a known length (points on the same line in the image above).
My question: Do the source and target point clouds need have the same number of points? Realistically, I could have many more points in the target point cloud as it is generate in real-time based on the frame geometry. I have tried setting both point clouds to the same size as well as setting the target to have 4x as many points. I notice a substantial reduction in RMS error when the target cloud had 4x more points.
Is super-sampling the target point cloud frowned upon? I know it can be an issue if the target cloud has fewer points than the source but I couldn’t find anything about the target having many more points.