.png){kind=icon}
.png){kind=icon}
Return
Topology-Constrained NeuroB-Rep: Contact-Aware CAD Reconstruction From 3-D Point Clouds
.png){kind=icon}
OA
AI
J
C
Y
S
DOI：10.1109/access.2026.3719250.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Automating the conversion of raw 3D scans into watertight and manufacturable computer-aided design (CAD) models remains challenging, as pipelines based on learning rarely enforce explicit topology such as inter-surface contacts, G1 continuity, and self-intersection avoidance. To address this, we present Topology-Constrained Neuro Boundary Representation (NeuroB-Rep), a solver that couples a hybrid surface model with a topology-aware objective and a stable contact gating schedule. The surface model combines analytic primitives for simple patches and neural implicit signed distance functions for freeform regions. The objective jointly penalizes gaps between surfaces, G1/G2 discontinuities, and self-intersections while regularizing the coaxiality and angular consistency of primitives. We employ a temperature-based gating mechanism with exponentially smoothed tolerance and a small persistent anchor set to activate candidate contacts gradually, thereby mitigating spurious snaps. Scans are normalized to a canonical pose, and the optimized patch graph is converted to a valid B-rep. The final result is exported as Standard for the Exchange of Product Model Data (STEP) via trimmed B-spline fitting and robust intersection curve construction. In a standardized study on industrial components (impeller, shaft, and casing), our method achieves geometric errors in the range of 0.01–0.03 mm relative to the scan while consistently producing watertight STEP files with closed topology. Ablation studies demonstrate that the proposed gating maintains geometric fidelity while improving contact closure and G1 continuity compared to index-based selection.
Keywords:
Design automation
solid modeling
geometric modeling
surface reconstruction
optimization methods
computational geometry
Journal
IF:
3.6
Papers:
9.7W
Citations:
29.4W
Organization
Cited Papers