Research map / Point clouds / Shape and surface completion
Point2CAD: Reverse Engineering CAD Models from 3D Point Clouds
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Point2CAD reconstructs complete B-rep CAD models from point clouds by combining pretrained segmentation, closed-form primitive fitting, a novel test-time-optimized implicit neural surface representation for freeform surfaces, and geometric intersection for topology.
Problem
Reverse engineering CAD models (surfaces, edges, corners with topology) from 3D point clouds, where prior work handles either freeform surfaces or topology but rarely both, and learned end-to-end methods produce topologically inconsistent models.
Approach
A hybrid analytic-neural pipeline: a pretrained segmentation backbone clusters points into faces, each cluster is fit with closed-form primitives or a per-surface 1-hidden-layer autoencoder with mixed SiLU and sinusoidal neurons (2D uv latent), and edges and corners come from pairwise mesh intersections rather than learned prediction.
Key results
- Freeform surface fitting residual error on open surfaces drops to 0.002 vs 0.006 for ParSeNet and 0.021 for ComplexGen.
- With HPNet segmentation, Point2CAD reaches surface F-score 0.788 at threshold 0.08 vs 0.732 precision-level baseline for ComplexGen (ComplexGen F-score 0.731).
- Point2CAD with ground-truth segmentation achieves Chamfer distance 0.016 vs 0.042 for ComplexGen and corner F-score 0.736 at the loosest threshold.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed mixed SiLU and sinusoidal activations instead of ReLU (ablation) | Point2CAD INR · ABC freeform surfaces with σ = 0.05 noise | P-coverage: 0.519 -> 0.585 (open); 0.509 -> 0.581 (closed) | Helped |
| Use the proposed mixed SiLU and sinusoidal activations instead of SiLU (ablation) | Point2CAD INR · ABC freeform surfaces with σ = 0.05 noise | residual error: 0.024 -> 0.011 (open); 0.030 -> 0.013 (closed) | Helped |
| Use the proposed mixed SiLU and sinusoidal activations instead of SiLU+PosEnc (ablation) | Point2CAD INR · ABC freeform surfaces, no noise | residual error: 0.004 -> 0.002 (open); 0.007 -> 0.003 (closed) | Mixed |
| Use the proposed mixed SiLU and sinusoidal activations instead of SIREN (ablation) | Point2CAD INR · ABC freeform surfaces, no noise | P-coverage: 0.996 -> 0.999 (open); 0.990 -> 0.998 (closed) | Helped |
| Increase extrapolation margin from c = 0% to c = 10% (ablation) | Point2CAD INR · ABC surface extrapolation | residual error: 0.0052 -> 0.0156 | Hurt |
| Increase extrapolation margin from c = 10% to c = 20% (ablation) | Point2CAD INR · ABC surface extrapolation | residual error: 0.0156 -> 0.0474 | Hurt |
| Use the proposed INR instead of ParseNet surface fitting | Freeform surface fitting · ABC open and closed freeform surfaces | residual error: 0.006 -> 0.002 (open); 0.008 -> 0.003 (closed) | Helped |
| Use Point2CAD rather than ComplexGen for full CAD reconstruction | ABC reconstructed CAD · ABC | Chamfer distance: 0.042 -> 0.016 | Helped |
| Use Point2CAD rather than ComplexGen for full CAD reconstruction | ABC reconstructed CAD · ABC | P-coverage: 0.950 -> 0.937 | Hurt |
Surprising findings
- Modular pipeline with learning-free geometric fitting and intersection outperforms end-to-end learned ComplexGen, suggesting learning is mainly needed for segmentation.
- Reported P-coverage is slightly lower than ComplexGen because clipping surfaces reduces coverage, while ComplexGen's spurious extra surfaces inflate that metric.
Experimental setup
Datasets: ABC dataset
Benchmarks: ABC dataset
Metrics: residual error P-coverage surface precision surface recall surface F-score edge F-score corner F-score chamfer distance
Baselines: ComplexGen ParSeNet HPNet Point2Cyl SPFN
Builds on: SIREN ParSeNet HPNet ComplexGen UMAP
Closest papers by meaning
- Draw Step by Step: Reconstructing CAD Construction Sequences from Point Clouds via Multimodal Diffusion. (CVPR 2024 · similarity 0.88)
- MiCADangelo: Fine-Grained Reconstruction of Constrained CAD Models from 3D Scans (NeurIPS 2025 · similarity 0.88)
- CAD-SIGNet: CAD Language Inference from Point Clouds using Layer-wise Sketch Instance Guided Attention (CVPR 2024 · similarity 0.88)
- EdgeMovingNet: Edge-preserving Point Cloud Reconstruction via Joint Geometry Features (CVPR 2025 · similarity 0.87)
- cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 2026 · similarity 0.87)
- From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.