Research map / Label-efficient and robust learning / Recognition and long-tail
SfmCAD: Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations
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.
Unsupervised network that reconstructs voxelized 3D shapes into sketch+path feature-modeling CAD operations (extrude, sweep, loft, revolve) via a two-stage Box+Path then implicit sketch learning.
Problem
Reconstructing 3D shapes from voxels into editable, compact CAD representations based on sketch-based feature modeling, with accuracy and interpretability, without supervision.
Approach
Decouples 3D structure (Bezier sweeping paths with box+path proxies) from local geometric detail (implicit 2D sketch SDFs), learned coarse-to-fine in two stages, with differentiable sketch-extrude/loft/sweep operators that generalize extrude and revolve as sweeps.
Key results
- On ABC, SfmCAD achieves CD 0.395 and NC 0.919 versus 0.506 CD and 0.884 NC for SECAD-Net.
- On PartNet segmented parts, SfmCAD improves CD by 25.3% over the closest competitor SECAD-Net (1.410 vs 1.888).
- On the tree dataset, SfmCAD reaches CD 0.686 and NC 0.722 compared with 2.197 and 0.622 for SECAD-Net.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SfmCAD sketch+path reconstruction instead of SECAD-Net | CAD reconstruction · ABC | CD: 0.506 -> 0.395 | Helped |
| Use SfmCAD sketch+path reconstruction instead of SECAD-Net | CAD reconstruction · ShapeNet | CD: 0.836 -> 0.626 | Helped |
| Use SfmCAD sketch+path reconstruction instead of SECAD-Net | Tree reconstruction · Tree | CD: 2.197 -> 0.686 | Helped |
| Use SfmCAD sketch+path reconstruction instead of SECAD-Net | PartNet reconstruction · PartNet | CD: 1.888 -> 1.410 | Helped |
| Add both sweep twist angles δu and δl (ablation) | SfmCAD · PartNet chair | CD: 4.092 -> 3.142 | Helped |
| Add sweep twist angle δl (ablation) | SfmCAD · PartNet chair | CD: 3.367 -> 3.142 | Helped |
| Add length-width regularization Llw (ablation) | SfmCAD · PartNet chair | CD: 4.804 -> 3.142 | Helped |
| Add smoothness regularization Lsm (ablation) | SfmCAD · PartNet chair | CD: 3.268 -> 3.142 | Helped |
Surprising findings
- On ShapeNet, SfmCAD's ECD (14.096) is slightly worse than ExtrudeNet (14.110) and SECAD-Net (14.818) is worse than SfmCAD, so edge fidelity is not uniformly improved.
Limitations
- Loft restricted to two sketches with a linear path; loft and sweep learned as separate choices per category rather than jointly.
- Requires per-category selection of the best operation (Loft/Sweep) via hyperparameters.
- Evaluation relies on Marching Cubes meshes and voxel inputs; ECD thresholds are heuristic.
Experimental setup
Datasets: ABC ShapeNet PartNet Procedural tree dataset (Palubicki et al.)
Benchmarks: ABC ShapeNet Tree dataset PartNet
Metrics: Chamfer Distance Edge Chamfer Distance Normal Consistency
Baselines: UCSG-Net CSG-Stump ExtrudeNet SECAD-Net
Builds on: SECAD-Net ExtrudeNet CSG-Stump Bezier curves
Closest papers by meaning
- MiCADangelo: Fine-Grained Reconstruction of Constrained CAD Models from 3D Scans (NeurIPS 2025 · similarity 0.85)
- Revisiting CAD Model Generation by Learning Raster Sketch (AAAI 2025 · similarity 0.85)
- CAD-SIGNet: CAD Language Inference from Point Clouds using Layer-wise Sketch Instance Guided Attention (CVPR 2024 · similarity 0.84)
- SketchDNN: Joint Continuous-Discrete Diffusion for CAD Sketch Generation (ICML 2025 · similarity 0.84)
- From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.82)
- cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 2026 · similarity 0.82)
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.