Abstract: Registration of non-rigid 3D models has a wide range of applications in different application areas such as animation driving, texture transfer, semantic understanding, etc. and is an important research topic in computer graphics.However, non-rigid objects based on skeletal deformation, such as human bodies and animals, often induce overall shape deformations in accordance with the movements of the joints. The non-rigid deformations and complex interactions of such objects thus pose difficulties and challenges to the 3D alignment task. Although recent studies using deep learning techniques can achieve good results in reconstructing 3D shapes with skeleton structures, these methods tend to rely heavily on large datasets with annotations. In the unsupervised case, they are still prone to generating overstretched and partially crossed model meshes.The paper proposes an unsupervised GAN-based network framework for the registration of 3D models, using a 3D generator to generate model meshes that fits the scanned point clouds, and two novel discriminators to impose regularization constraints on the deformed meshes, which achieves good results on public datasets.

Key words: 3D human reconstruction, 3D registration, adversarial networks, non-rigid objects