Speaker: Florian Laager (ICGA)
Objects are scanned in 3D by taking overlapping depth images from different perspectives, which then must be registered together to reconstruct the mesh.
Most current work detects a few corresponding points by features in order to find the transform which correctly positions the depth images to each other in 3D. Due to its nature of extracting and matching only a few data, it may not select the best global correspondence, especially if there are several close matches.
It is also difficult to detect complete failure, since the quality of the guess is not measured, so often manual intervention is necessary. Usually there are multiple depth images and several serial pair-wise registrations, so any errors in the determined transform propagate.
Therefore a better approach to reconsider that problem to use all of the available data and determine only overlaps within an error threshold, e.g. based on confidence from the scanner noise model. This also has applications in shape retrieval, for high-quality partial shape matching. A simplified version of that problem would be to find an overlap of two functions in 2D with the difference in their y-value thresholded in some measure.