Global As-conformal-as-possible Non-rigid Registration Of Multi-view Scans

Vision-based 3D reconstruction is the technology that derives the three-dimensional information of objects in the real scences with the knowledge of computer vision after analyzing and processing the images captured by the sensors such as cameras.3D reconstruction has been always one of the frontier key research topics,which has the wide applicability in life,entertainment and services,such as 3D printing,games and movie production.However,multi-cameras based 3D reconstruction are not readily usable in many real-world applications due to high cost,complex operation and difficult maintenance.Thus single-camera based 3D reconstruction becomes the main research direction in this area.This paper proposes the method that non-simultaneously captures multi-view scans of the deformable objects and reconstructs a 3D model in order to obtain a complete model of deformable objects easily in low-cost way.The main innovations are as follows.(1)This paper contributes a novel method for global non-rigid registration of multiscans,which avoids the drift problem caused by error accumulation in traditional methods,and is robust to noise and outliers.In addition,this method is also suitable for the alignment problem of the sparse viewpoints scans.(2)This paper empolys a joint point-to-point and point-to-plane positional constraint in the optimization frame to reduce the influence of wrong correspondences effectively.(3)This paper incorporates an as-conformal-as-possible constraint into the energy function of global non-rigid registration to avoid mesh distortions and maintain mesh structures by preserving the angles of triangles in the meshes during deformation.(4)This paper designs a reweighting scheme to update weights iteratively on data term and smooth term to better approximate L0-norm in measuring sparseness,which helps reduce final registration errors.In the quantitative and qualitative evaluations of comparative experiment,our approach outperforms state-of-the-art methods on both public datasets and real scanned datasets by RGB-D cameras.