First Steps Toward the Automatic Registration of Deformable Scans

Three-dimensional registration is the process of aligning scans of different incomplete views of an object such that their corresponding regions agree in space and time. It is used to build digital complete models from partial acquisitions of real-world objects which typically contain noise, outliers, and acquisition holes due to occlusions and hardware limitations. For non-rigid models, such as humans, shapes of corresponding regions undergo complex deformations in addition to misalignments. Thus, finding explicit correspondences between two dissimilar shapes becomes considerably more challenging. In addition, finding the right deformations is also non-trivial as they depend on the correspondences which are only partially available between consecutive scans. This report presents an analysis of the 3-D registration problem, a general approach for the spatio-temporal registration of deformable scans that are acquired with a fast 3-D scanner, and extensive experiments on synthetic and real acquisition data. The objective is to un-deform the surfaces of all range maps and to fit them into a particular instance of time. We address this problem by finding point correspondences based on initial closest point computations between partial scans and performing deformations according to these matches. Similar to the iterative closest point (ICP) algorithm, the correspondence and deformation steps are tightly coupled and optimized within a loop. Although smooth and small scale deformations between consecutive scans are assumed, we envision allowing large deformations, such as human bodies in completely different poses, during the acquisition. This approach is the main distinction over previous work on non-rigid ICP variants, where the accumulation of registration errors has to be globally minimized. The suggested approach makes extensive use of assumptions on surface and deformation smoothness. Our experiments confirm the success of our method for small deformations where the error accumulation of the correspondences over time is limited. However, this preliminary investigation shows that without additional assumptions about the correspondence flow and object deformation, such as rigidness and near-isometry, registration of scans with large scale deformations remains an unsolved problem. Nevertheless, promising intermediate results open up a new line of sub-problems needed to be solved for the grand challenge of fully automatic reconstruction of surface, motion, and correspondence from time coherent range-scans.