Existing methods for CAD reconstruction from 3D scans achieve strong performance on standard benchmarks, yet their effectiveness often degrades in real-world scenarios where objects consist of multiple parts and scan artifacts such as noise. A key reason is that current CAD reconstruction datasets predominantly contain single-part models and clean geometric inputs, limiting the ability of learned models to generalize to realistic settings. In this work, we introduce CADonatello, a part-aware framework for reconstructing parametric CAD models from real-world 3D scans. Our approach first decomposes the input geometry into structurally meaningful components using a foundation-scale 3D part segmentation model, and then reconstructs each component independently through a cross-section-driven pipeline that performs sketch parameterization followed by an extrusion. To improve the robustness of sketch parameterization to scan-derived cross-sections, we further introduce CC3D Cross-Sections, a dataset of paired rasterized cross-section images from realistic scans and their corresponding parametric sketches derived from aligned CAD models. This dataset enables effective domain adaptation from clean CAD sketch images to scan-derived cross-section images. Extensive experiments demonstrate that the proposed approach improves generalization to noisy scans and enables reliable reconstruction of complex multi-part objects across multiple real-world CAD reverse engineering benchmarks.
Figures & tables
Figure 1: CADonatello is a part-aware CAD reverse engineering method that decomposes 3D scans using a foundation model for 3D part segmentation and reconstructs editable CAD models through cross-section–driven parametric part reconstruction.
Figure 2: Overview of CADonatello. The input mesh is decomposed into parts, each part is reconstructed through cross-section parameterization and extrusion inference, and the recovered parts are assembled into the final CAD model.
Method
CC3D
CC3D-Complex
CD ↓
IR ↓
ECD ↓
CD ↓
IR ↓
ECD ↓
#Params
CAD-SIGNet ( Khan et al. 2024 )
2.90
4.4
8.68
5.39
5.03
15.36
6.1M
MiCADangelo ( Karadeniz et al. 2025b )
1.69
2.2
5.93
3.37
3.2
8.18
59M
CAD-Recode ( Rukhovich et al. 2025 )
0.56
8.6
3.41
0.93
10.0
4.66
1.5B
Cadrille ( Kolodiazhnyi et al. 2025 )
0.44
0.5
3.65
0.74
0.5
4.91
2B
Ours
0.48
1.6
2.78
0.71
0.7
3.57
171M
Table 1: Quantitative Results on CC3D and CC3D-Complex.
Figure 3: Qualitative comparison of our method with CAD-SIGNet ( Khan et al. 2024 ) , MiCADangelo ( Karadeniz et al. 2025b ) , CAD-Recode ( Rukhovich et al. 2025 ) , and Cadrille ( Kolodiazhnyi et al. 2025 ) on CC3D dataset.
Method
F360
F360-Complex
CD ↓
IR ↓
ECD ↓
CD ↓
IR ↓
ECD ↓
CAD-SIGNet ( Khan et al. 2024 )
19.32
26.5
49.89
19.67
29.3
67.78
MiCADangelo ( Karadeniz et al. 2025b )
12.07
9.0
23.23
14.89
11.1
33.34
CAD-Recode ( Rukhovich et al. 2025 )
2.52
15.5
8.15
2.94
17.4
8.68
Cadrille ( Kolodiazhnyi et al. 2025 )
1.81
0.0
8.39
2.14
0.0
11.44
Ours
1.57
0.0
5.28
1.87
0.0
5.09
Table 2: Quantitative results on Fusion 360 Assembly and its complex subset.
Figure 4: Qualitative comparison of our method with CAD-SIGNet ( Khan et al. 2024 ) , MiCADangelo ( Karadeniz et al. 2025b ) , CAD-Recode ( Rukhovich et al. 2025 ) , and Cadrille ( Kolodiazhnyi et al. 2025 ) on Fusion 360 Assembly dataset.
Method
CD ↓
ECD ↓
IR ↓
CAD-SIGNet ( Khan et al. 2024 )
13.82
49.5
29.5
MiCADangelo ( Karadeniz et al. 2025b )
10.5
36.16
2.5
CAD-Recode ( Rukhovich et al. 2025 )
2.1
4.87
12.0
Cadrille ( Kolodiazhnyi et al. 2025 )
1.37
4.31
0.0
Ours
0.08
0.24
0.0
Table 3: Quantitative results on extrusions non-aligned with canonical axes.
CAD Sketch Parameterization
3D Reconstruction
Training Dataset
Acc ↑
PF1 ↑
SCD ↓
NCC ↓
CD ↓
IR ↓
ECD ↓
SG ( Seff et al. 2020 )
0.851
0.427
2.24
1.95
0.59
2.8
3.17
SG + Synt ( Karadeniz et al. 2025b )
0.880
0.550
0.62
1.31
0.52
1.3
3.01
SG + CC3D-CS (Ours)
0.928
0.844
0.251
1.10
0.48
1.6
2.78
Table 4: Effect of different training datasets on sketch parameterization and their impact on 3D reconstruction performance on CC3D.
Method
Seg. IoU ↑
CD ↓
IR ↓
ECD ↓
Part segmentation on Fusion360 Assembly
PointNeXt ( Qian et al. 2022 )
43.1
3.61
10.0
9.54
SAMPart3D ( Yang et al. 2024 )
52.5
2.90
6.5
9.47
PartSAM ( Zhu et al. 2025 )
72.5
2.08
6.0
5.51
P3-SAM ( Ma et al. 2025 )
77.6
1.57
0.0
5.28
Impact of Extrusion strategy on CC3D
Table 5: Ablation studies on part segmentation and extrusion strategy.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Example cross-section image and parametric sketch pairs from the CC3D Cross-Section dataset.
Figure 6: Relationship between intermediate pipeline errors and final reconstruction CD on Fusion 360 Assembly.
CAD Sketch Parameterization
3D Reconstruction
Training Dataset
Acc ↑
PF1 ↑
SCD ↓
NCC ↓
CD ↓
IR ↓
ECD ↓
SG
0.907
0.697
0.746
2.32
1.64
1.0
6.06
SG + Synt
0.933
0.764
0.262
1.39
1.58
0.0
5.42
SG + CC3D-CS
0.935
0.852
0.217
1.70
1.57
0.0
5.37
Appendix
Table 6: Effect of sketch training data on Fusion 360 Assembly.
Figure 7: Representative failure cases of CADonatello .
Figure 8: Example part segmentations from the CC3D dataset using P3-SAM ( Ma et al. 2025 ) .
Figure 9: Example part segmentations from the F360 Assembly dataset using P3-SAM ( Ma et al. 2025 ) .
Figure 10: More qualitative comparisons with CAD-SIGNet ( Khan et al. 2024 ) , MiCADangelo ( Karadeniz et al. 2025b ) , CAD-Recode ( Rukhovich et al. 2025 ) , and Cadrille ( Kolodiazhnyi et al. 2025 ) on CC3D dataset.
Figure 11: More qualitative comparisons with CAD-SIGNet ( Khan et al. 2024 ) , MiCADangelo ( Karadeniz et al. 2025b ) , CAD-Recode ( Rukhovich et al. 2025 ) , and Cadrille ( Kolodiazhnyi et al. 2025 ) on Fusion 360 Assembly dataset.
Figure 12: Qualitative comparison of sketch parameterization between our model (SketchGraphs pre-training + CC3D Cross-Section fine-tuning) and ( Karadeniz et al. 2025b ) , which is trained on SketchGraphs with synthetic noise (SG+Synt).
Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Siemens Digital Industries Software, Plano, TX 75024, USA