Research map / Point clouds / Point cloud classification
CAD-SIGNet: CAD Language Inference from Point Clouds using Layer-wise Sketch Instance Guided Attention
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.
Proposes CAD-SIGNet, an end-to-end auto-regressive transformer that infers sketch-and-extrusion CAD design histories from point clouds using layer-wise cross-attention and a sketch-instance-guided attention mask.
Problem
Recovering the CAD design history (sketch-and-extrusion sequence) from a 3D point cloud, a reverse engineering task where prior methods used separate-stage or feed-forward inference and gave no interactive choices.
Approach
Jointly learns point cloud and CAD language embeddings via layer-wise cross-attention in multi-modal transformer blocks, auto-regressively predicts extrusion tokens first, and restricts sketch-token attention to the point subset inside the predicted extrusion's sketch bounding box (Sketch instance Guided Attention).
Key results
- On DeepCAD, median CD of 0.283e-3 and invalidity ratio 0.88% versus DeepCAD's 9.640e-3 and 7.14%.
- Arc primitive F1 rises from 60.14 (DeepCAD)... specifically arc F1 improves from 13.82 to 28.65 and extrusion F1 from 86.70 to 92.72.
- Cross-dataset on Fusion360 median CD 1.15e-3 with IR 1.83% versus MultiCAD 42.2e-3 and DeepCAD 89.2e-3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAD-SIGNet instead of DeepCAD for point-cloud-to-CAD inference | DeepCAD dataset · DeepCAD test | median CD (×10−3): 9.640 -> 0.283 | Helped |
| Use CAD-SIGNet instead of DeepCAD for point-cloud-to-CAD inference | DeepCAD dataset · DeepCAD test | IR: 7.14 -> 0.88 | Helped |
| Use CAD-SIGNet instead of DeepCAD for primitive prediction | DeepCAD dataset · DeepCAD test | arc F1: 13.82 -> 28.65 | Helped |
| Add hybrid sampling rather than using top-1 token probabilities (ablation) | CAD-SIGNet · DeepCAD test | mean CD (×10−3): 6.784 -> 3.432 | Helped |
| Add hybrid sampling rather than using top-1 token probabilities (ablation) | CAD-SIGNet · DeepCAD test | IR: 5.02 -> 0.88 | Helped |
| Add Sketch instance Guided Attention (SGA) (ablation) | CAD-SIGNet · DeepCAD test | mean CD (×10−3): 4.995 -> 3.432 | Helped |
| Add Sketch instance Guided Attention (SGA) (ablation) | CAD-SIGNet · DeepCAD test | IR: 2.18 -> 0.88 | Helped |
| Add layer-wise cross-attention (ablation) | CAD-SIGNet · DeepCAD test | mean CD (×10−3): 122.7 -> 3.432 | Helped |
| Add hybrid sampling rather than using top-1 token probabilities (ablation) | CAD-SIGNet · DeepCAD conditional auto-completion | CD ratio Q2 (median): 0.696 -> 0.325 | Helped |
| Add Sketch instance Guided Attention (SGA) (ablation) | CAD-SIGNet · DeepCAD conditional auto-completion | CD ratio Q2 (median): 0.458 -> 0.325 | Helped |
| Use CAD-SIGNet instead of MultiCAD for cross-dataset inference | Fusion360 · Fusion360 cross-dataset test | median CD (×10−3): 42.2 -> 1.15 | Helped |
| Use CAD-SIGNet instead of DeepCAD for realistic scan inference | CC3D · CC3D test | median CD: 263.56 -> 2.398 | Helped |
Surprising findings
- Removing layer-wise cross-attention yields 0% invalidity but drastically worse F1 and CD metrics (e.g., sketch F1 45.89).
Limitations
- Only extrusion operations are considered; other CAD operations are not modeled.
- Selecting points via SGA is motivated partly by high computational cost on large point clouds, which remains an open issue.
- Baselines SkexGen and HNC were adapted and not optimized for the auto-completion task.
Experimental setup
Datasets: DeepCAD Fusion360 CC3D
Benchmarks: DeepCAD test set Fusion360 (cross-dataset) CC3D (cross-dataset)
Metrics: Invalidity Ratio Mean Chamfer Distance Median Chamfer Distance F1 score (line, arc, circle, extrusion, sketch) CD ratio quartiles
Baselines: DeepCAD MultiCAD SkexGen HNC
Builds on: DeepCAD SkexGen LFA (local feature aggregation) Transformer
Closest papers by meaning
- MiCADangelo: Fine-Grained Reconstruction of Constrained CAD Models from 3D Scans (NeurIPS 2025 · similarity 0.90)
- Draw Step by Step: Reconstructing CAD Construction Sequences from Point Clouds via Multimodal Diffusion. (CVPR 2024 · similarity 0.89)
- cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Point2CAD: Reverse Engineering CAD Models from 3D Point Clouds (CVPR 2024 · similarity 0.88)
- CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images (CVPR 2025 · similarity 0.88)
- CAD-GPT: Synthesising CAD Construction Sequence with Spatial Reasoning-Enhanced Multimodal LLMs (AAAI 2025 · similarity 0.87)
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.