Research map / 3D vision / 3D generation
MV2Cyl: Reconstructing 3D Extrusion Cylinders from Multi-View Images
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.
Reconstructs sketch-extrude CAD models (extrusion cylinders) directly from multi-view images by learning 2D U-Net segmentation priors for surfaces and curves, lifting them into TensoRF-based neural fields, and recovering CAD parameters.
Problem
Recovering sketch-extrude CAD structure (extrusion cylinders with 2D sketches and extrusion parameters) from multi-view images rather than raw 3D point clouds, which prior methods require as input.
Approach
Combine 2D U-Net segmentation of extrusion surfaces (instance and start/end/barrel) and base-curve segmentation, integrate both into separate neural fields (TensoRF-based density and semantic fields), and jointly use surface and curve information to estimate axis, sketch, height, and center, since surfaces give robust axes and curves handle occlusion and sketches.
Key results
- On Fusion360, MV2Cyl reaches 1.3939 degree extrusion-axis error vs 9.5228 for Point2Cyl (point-cloud input).
- On DeepCAD, MV2Cyl reaches 0.2202 degree extrusion-axis error vs 7.9156 for Point2Cyl and 0.0121 center error vs 0.0266.
- Ablation on Fusion360 shows surface-only is worse (E.A. 4.1180) than curve-only (1.5612) and joint surface+curve (1.3939).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add curve-field reconstruction to surface-only reconstruction (ablation) | MV2Cyl · Fusion360 | Fit Glob.: 0.0353 -> 0.0212 | Helped |
| Add surface-field reconstruction to curve-only reconstruction (ablation) | MV2Cyl · Fusion360 | Fit Glob.: 0.0232 -> 0.0212 | Helped |
| Use a shared U-Net instead of separate curve and surface U-Nets (ablation) | MV2Cyl · Fusion360 | Fit Glob.: 0.0212 -> 0.0200 | Helped |
| Reduce input views from 50 to fewer views (ablation) | MV2Cyl · Fusion360 | Fit Glob.: 0.0212 -> 0.0218; 0.0259; 0.0364 | Hurt |
| Change curve-map line width from 5 pixels to 2.5 or 7.5 pixels (ablation) | MV2Cyl · Fusion360 | Fit Glob.: 0.0212 -> 0.0216; 0.0225 | Mixed |
| Use both surface and curve representations rather than either alone (ablation) | MV2Cyl · Fusion360 | Fit Cyl.: 0.0984 -> 0.0284; 0.0310 -> 0.0284 | Helped |
| Reduce the number of input images from 50 to 25, 15, or 10 (ablation) | MV2Cyl · Fusion360 | E.A. (°): 1.3939 -> 1.4631; 2.1548; 2.2585 | Hurt |
| Use 2.5-pixel or 7.5-pixel curve-map line width instead of 5 pixels (ablation) | MV2Cyl · Fusion360 | E.A. (°): 1.3939 -> 1.4886; 1.4375 | Mixed |
Surprising findings
- Naively reconstructing surfaces from multi-view images with NeuS2 before Point2Cyl performs far worse (E.A. 35.06 on Fusion360) than the point-cloud baseline.
- Adding ground-truth curve points to Point2Cyl gives no significant improvement over the original Point2Cyl.
Limitations
- Does not predict binary boolean operations between extrusion cylinders explicitly.
- Susceptible to occlusion; fully hidden extrusion sides cannot be reconstructed.
- Does not handle textured CAD models; segmentation trained only on synthetic untextured renderings.
Experimental setup
Datasets: Fusion360 DeepCAD
Benchmarks: Fusion360 DeepCAD
Metrics: extrusion axis error extrusion center error extrusion height error per-cylinder fitting loss global fitting loss
Baselines: Point2Cyl NeuS2+Point2Cyl Curve+Point2Cyl SECAD-Net
Builds on: Point2Cyl NEF TensoRF NeuS2 IGR
Closest papers by meaning
- Point2Cyl: Reverse Engineering 3D Objects From Point Clouds to Extrusion Cylinders (CVPR 2022 · similarity 0.91)
- Learning Partonomic 3D Reconstruction from Image Collections (CVPR 2025 · similarity 0.84)
- From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.84)
- MiCADangelo: Fine-Grained Reconstruction of Constrained CAD Models from 3D Scans (NeurIPS 2025 · similarity 0.84)
- PyTorchGeoNodes: Enabling Differentiable Shape Programs for 3D Shape Reconstruction (CVPR 2025 · similarity 0.83)
- 3DCoMPaT200: Language Grounded Large-Scale 3D Vision Dataset for Compositional Recognition (NeurIPS 2024 · similarity 0.83)
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.