FCDNet: A Unified Deformation Network for Non-Rigid 3D Human Shape Correspondence via Feature Consistency

3D shape correspondence is a fundamental task in computer vision, yet it remains difficult for articulated human bodies due to complex non-rigid deformations. In particular, existing methods often struggle to maintain robustness when inputs are limited to single-view partial scans with severe occlusions. To address the challenge of establishing consistent point cloud correspondence across both complete and single-view human body shapes, we propose a unified Deformation Network via Feature Consistency Learning (FCDNet). FCDNet learns deformations from a predefined template to the target shape. Its Space Cover Convolution (SC-Conv) encoder captures robust local geometric structures of point cloud by constructing anisotropic geometries in neighborhoods. Crucially, FCDNet employs feature consistency learning during joint training on both complete and partial (single-view) shapes, enabling the encoder to learn representations invariant to shape incompleteness. With encoded global features, the subsequent decoder is further supervised using complete shapes as ground truth, ensuring accurate template deformation even for single-view inputs. Extensive experiments on the FAUST and SHREC'19 datasets demonstrate that FCDNet outperforms state-of-the-art methods, achieving average correspondence errors of 2.54 cm and 1.8 cm, respectively. FCDNet also performs effectively on single-view data and unclean point clouds captured by depth cameras.