SheepRPMNet: Towards low-overlap-aware multi-view 3D point cloud registration via global normalization in dairy sheep

Multi-view low-overlap-aware point cloud registration remains a major challenge in 3D vision, especially for agricultural visualization and livestock modeling. To address the low overlap and structural deficiencies in multi-view dairy sheep scans, we propose SheepRPMNet, an enhanced RPMNet-based registration model. The method first establishes multi-view consistency through a centralization and global normalization pipeline, which aligns all views into a unified coordinate system; subsequent pairwise registration between side (left/right) and top views is then performed within this globally consistent framework. This is further supported by three key enhancements: a local reference frame-based Histograms of Projected Normal Vector Distribution (HPNVD) feature descriptor is proposed, which integrates hierarchical graph convolution with a Transformer Attention mechanisms, thereby enhancing the representation capabilities of local geometric details and global structures; a multi-scale parameter prediction module that dynamically adjusts annealing parameters; and a dynamic-weight loss function that adaptively balances rotation and translation errors. Experiments show that SheepRPMNet reduces the improved Chamfer distance by 64% compared with RPMNet and delivers accurate body-size estimations, achieving mean absolute percentage errors of 3.10%, 3.15%, 3.06%, 3.47%, and 3.05% for key livestock measurements. These results demonstrate the potential of SheepRPMNet for precise multi-view registration and intelligent body measurement of dairy sheep. Code will be released upon acceptance.