Flow-deformation-aware point cloud completion network for 3D metal bent tube

The task of completing shapes of complex curved surface components is a challenging yet crucial problem in industrial quality inspection. The hollow structures, irregular sectional profiles, and flexible trajectories of 3D bent tubes impose difficulties for precise shape completion. A deeper understanding of the material plastic flow mechanism enhances learning point clouds. This paper proposes a flow-deformation-aware point cloud completion network, FDANet, for industrial 3D metal plastic-formed bent tube. We exploit knowledge of plastic deformation during physical manufacturing processes to guide learning and preservation of geometric details. FDANet introduces an attention mechanism to dynamic graph convolution to learn the interactions between neighborhood nodes. An approximate prediction model of tube cross-sectional distortion is proposed, which injects section-dimensional features into FDANet and facilitates the learning of realistic defect shapes. We have further designed a global attention module in the transformer encoder to adaptively integrate sectional points into high-level features and summarize hierarchical geometrical information, which enables fine shape completion. FDANet shows superior completion performance for overall shape and profile details on our own multi-scale tube datasets as well as widely used benchmark datasets. Our code is available at https://github.com/wangle0816/FDANet.