Research map / 3D vision / Face and shape
MiCADangelo: Fine-Grained Reconstruction of Constrained CAD Models from 3D Scans
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 MiCADangelo, a CAD reverse-engineering pipeline that detects key multi-plane cross-sections of a 3D scan, predicts constrained parametric sketches from rasterized loops, and optimizes extrusion lengths differentiably.
Problem
Converting 3D scans into editable, fully parametric CAD sketch-extrude models is hard; prior top-down methods miss fine geometric detail, bottom-up methods lack full parametrization, and none infer sketch constraints.
Approach
Mimics human designers by slicing the scan along x/y/z axes, selecting key sketch planes with a transformer over slice embeddings, parameterizing each closed loop's sketch primitives and constraints from raster images, and fitting extrusion lengths by differentiable point-to-vector distance to the mesh.
Key results
- On the DeepCAD test set, median CD 0.20 and IoU 80.6 vs CAD-SIGNet's 0.28 and 77.6; on Fusion360, median CD 0.48 and IoU 68.7 vs 0.56 and 65.6.
- On DeepCAD models with at least 4 loops, median CD 0.37 and IoU 68.3 vs CAD-SIGNet's 1.34 and 49.2.
- Invalidity ratio is worse than CAD-SIGNet on DeepCAD (2.6 vs 0.9) and Fusion360 (3.2 vs 1.6) in the main table; the paper notes CAD-SIGNet's low IR relies on test-time sampling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add contextual embeddings to plane detection (ablation) | MiCADangelo · DeepCAD reconstruction | Median CD: 0.39 -> 0.20 | Helped |
| add contextual embeddings to plane detection (ablation) | MiCADangelo · DeepCAD reconstruction | IoU: 69.4 -> 80.6 | Helped |
| increase cross-section sampling from N=10 to N=40 per axis (ablation) | MiCADangelo · DeepCAD reconstruction | Median CD: 0.26 -> 0.20 | Helped |
| increase cross-section sampling from N=10 to N=40 per axis (ablation) | MiCADangelo · DeepCAD reconstruction | IoU: 78.2 -> 80.6 | Helped |
| use MiCADangelo instead of CAD-SIGNet | CAD reconstruction · DeepCAD test set | Median CD: 0.28 -> 0.20 | Helped |
| use MiCADangelo instead of CAD-SIGNet | CAD reconstruction · DeepCAD test set | IR: 0.9 -> 2.6 | Hurt |
| use MiCADangelo instead of CAD-SIGNet | CAD reconstruction · Fusion360 test set | ECD: 4.14 -> 2.66 | Helped |
| fine-tune sketch parameterization on synthetic noisy closed loops | MiCADangelo · Single-extrusion DeepCAD test models | Avg. SCD | Helped |
| increase extrusion vector count up to 8 | MiCADangelo extrusion optimization · Extrusion optimization evaluation | performance | Helped |
Surprising findings
- CAD-SIGNet's low invalidity ratio is attained via test-time sampling; without it CAD-SIGNet has IR 4.4 on DeepCAD and 9.3 on Fusion360.
- Removing contextual embeddings in plane detection gives much worse mean CD (8.80 vs 2.27) yet still reasonable median CD, attributed to robustness of downstream stages.
Limitations
- Supports only extrusion operations; extrusion direction tied to sketch plane normals, failing on non-axis-aligned extrusions.
- Does not support B-spline sketch primitives.
- Invisible cut extrusions are not directly handled.
Experimental setup
Datasets: DeepCAD Fusion360 SketchGraphs CC3D Extruded SketchGraphs
Benchmarks: DeepCAD test set Fusion360 test set CC3D DeepCAD complex subset
Metrics: Median Chamfer Distance IoU Invalidity Ratio Edge Chamfer Distance Sketch Chamfer Distance Precision Recall F1
Baselines: DeepCAD Point2Cyl CAD-Diffuser CAD-SIGNet Vitruvion Davinci
Builds on: Vitruvion DeepCAD CAD-SIGNet SketchGraphs
Closest papers by meaning
- CAD-SIGNet: CAD Language Inference from Point Clouds using Layer-wise Sketch Instance Guided Attention (CVPR 2024 · similarity 0.90)
- CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images (CVPR 2025 · similarity 0.89)
- cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 2026 · similarity 0.89)
- From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.88)
- Point2CAD: Reverse Engineering CAD Models from 3D Point Clouds (CVPR 2024 · similarity 0.88)
- Draw Step by Step: Reconstructing CAD Construction Sequences from Point Clouds via Multimodal Diffusion. (CVPR 2024 · 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.