Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds

Representing complex 3D objects as simple geometric primitives, known shape abstraction, is important for modeling, structural analysis, and synthesis. In this paper, we propose an unsupervised abstraction method to map a point cloud into compact cuboid representation. We jointly predict allocation part segmentation shapes enforce the consistency between self-learning. For task, transform input set of parametric cuboids using variational auto-encoder network. The network allocates each considering point-cuboid affinity. Without manual annotations parts in clouds, design four novel losses supervise two branches terms similarity compactness. evaluate our on multiple collections demonstrate its superiority over existing methods. Moreover, based architecture learned representations, approach supports various applications including structured generation, interpolation, clustering.