Source-linked AI summary
CADonatello: Towards Real-World Multi-Part 3D CAD Reconstruction
Ahmet Serdar Karadeniz, Dimitrios Mallis, Parsa Akbari Chianeh, Kseniya Cherenkova, Anis Kacem, Djamila Aouada
TL;DR
Recovering editable parametric CAD from scans remains challenging for complex, multi-part objects. CADonatello decomposes scans into parts and reconstructs them through cross-section-driven modeling, reporting state-of-the-art performance on CC3D-Complex and Fusion360-Assembly.
Problem
Automating recovery of editable parametric CAD from scans remains challenging, particularly for geometrically complex or multi-part assemblies.
Source: Introduction · p. 1Method
CADonatello segments scans into structurally meaningful parts and reconstructs each through cross-section-driven sketch parameterization and extrusion, alongside paired scan-derived cross-sections and parametric sketches.
Source: Introduction · p. 2, Part Decomposition · p. 3, CC3D Cross-Section Dataset · p. 5Results
The paper reports state-of-the-art performance on CC3D-Complex and Fusion360-Assembly, including complex multi-part assemblies in the latter benchmark.
Source: Introduction · p. 2, Conclusion · p. 7Takeaways & Limitations
The supported conclusion is that part-aware reconstruction generalizes to realistic scanning artifacts and scales to complex multi-part objects.
Source: Conclusion · p. 7
Abstract
from arXiv · showhide
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.
Introduction
CADonatello targets editable parametric CAD reconstruction from realistic scans by decomposing objects into parts and reconstructing them parametrically. It addresses the limits of datasets dominated by single-part models and clean inputs.
Introduction
- CAD reverse engineering recovers editable parametric models from scans because manufactured objects may lack available original CAD models, while point clouds and meshes are not directly editable.The target is a parametric construction sequence of sketches and solid-modeling operations that reproduces the observed geometry.Source: Introduction · p. 1
- Existing CAD reconstruction datasets predominantly contain single-part models and clean synthetic geometry, limiting generalization to complex assemblies and scans with noise, occlusions, or incomplete surfaces.These limitations motivate moving beyond direct mapping from input geometry to CAD sequences.Source: Introduction · p. 2
- The framework segments scans into semantically meaningful parts, reconstructs each part parametrically, and assembles the recovered components.Its pipeline uses foundation-scale 3D segmentation, orientation-aware cross-section plane detection, image-level sketch parameterization, and adaptive extrusion optimization.Source: Introduction · p. 2
- CC3D Cross-Sections contains 44k paired rasterized scan-derived cross-sections and parametric sketches from aligned BReps, capturing scan artifacts and geometric distortions.The dataset is intended to support learning and evaluation of robust cross-section parameterization under realistic noise.Source: Introduction · p. 2
- CADonatello reports state-of-the-art performance on noisy CC3D scans and results on complex multi-part assemblies using the Fusion 360 Assembly dataset.The paper describes this as the first demonstration of reconstruction on complex multi-part assemblies.Source: Introduction · p. 2
Related Work
Prior CAD reconstruction methods largely model whole objects, while CADonatello uses part segmentation and scan-paired sketch training to support reconstruction of multi-part objects from real-world scans.
3D CAD Reverse Engineering
- Earlier CSG and BRep methods gave way to learned sketch–extrude sequences, but whole-object reconstruction can be fragile on real-world scans with multiple components; CADonatello reconstructs structural parts.The cited approaches report effectiveness on existing CAD benchmarks, but typically reconstruct the entire object directly.Source: Related Work · p. 2
3D Part Segmentation
- Foundation-scale native 3D models enable category-agnostic part decompositions and generalization to unseen objects, but their use for CAD reverse engineering remains largely unexplored; CADonatello applies them to scans.Earlier supervised point-cloud or mesh methods generalize poorly to unseen categories, while 2D-model-based approaches can yield viewpoint ambiguities and inconsistent 3D boundaries.Source: Related Work · p. 2
Parametric CAD Sketches
- Image-conditioned sketch methods predict geometric primitives and constraints, while Karadeniz et al. (2025b) infer parametric sketches from scan-derived cross-sections.Training that method with synthetic noise on clean sketch images may not reflect irregularities in real cross-sections; CADonatello fine-tunes on CC3D Cross-Sections, pairing real cross-sections with parametric sketches to improve robustness to real-world scans.Source: Related Work · pp. 2–3
Problem Statement
The problem is to recover a sequence of sketch–extrude steps whose composition reconstructs an input 3D mesh. The representation includes sketch planes, constrained parametric sketches, and additive or subtractive extrusions.
Problem Statement
- The objective is to recover a sequence of sketch–extrude steps whose composition reconstructs the input mesh.The target CAD representation is C = {(K_j, e_j)} for j = 1 through n_s.Source: Problem Statement · p. 3
- Each sketch lies on a plane and encodes 2D line, circle, or arc primitives with geometric parameters and constraints expressing relationships between primitives.Constraints include coincident, concentric, equal, horizontal, midpoint, normal, parallel, perpendicular, tangent, and vertical relationships.Source: Problem Statement · p. 3
- Each extrusion extends its sketch along a direction vector by a specified height, with its type indicating whether material is added or removed.An extrusion is represented as e_j = (π_j, t_j, v_j, h_j).Source: Problem Statement · p. 3
Proposed Method
CADonatello reconstructs real-world multi-part scans by segmenting meshes into parts, reconstructing each part through cross-section sketches and extrusions, and assembling the results in their original configuration.
Proposed Method
- CADonatello decomposes artifact-affected meshes into parts, reconstructs each independently with sketch parameterization and inferred extrusions, then assembles the part sequences in their original spatial configuration.The final CAD model is formed from the reconstructed part-level sequences.Source: Proposed Method · p. 3
Part Decomposition
- P3-SAM (Ma et al. 2025) automatically produces category-agnostic part masks from 3D geometry, using a model trained on more than 3 million examples to enable independent parametric reconstruction.Uniformly sampled surface prompt points produce per-face segmentation masks and geometrically coherent part submeshes.Source: Part Decomposition · p. 3
Part Reconstruction
- Part reconstruction identifies key slicing planes, parameterizes constrained cross-sections, and infers corresponding extrusion operations, adapting the cross-section paradigm of (Karadeniz et al. 2025b) to decomposed parts.Source: Part Reconstruction · p. 3
Orientation-aware Key Slicing Plane Estimations.
- Each part is canonicalized using its oriented bounding box to handle arbitrary orientations; single-plane or learned multi-plane strategies are selected by comparing extrusion reconstructions with the input geometry.The single-plane strategy is intended for parts reconstructible with one sketch–extrude operation; otherwise, a learning-based detector identifies multiple key planes.Source: Orientation-aware Key Slicing Plane… · p. 4
Scan Cross-section Parameterization.
- Rasterized cross-section loops are processed by a transformer encoder to predict sketch primitives and constraints; the network is pretrained on 1.5 million SketchGraphs sketches and fine-tuned on realistic scan-derived cross-sections to address the domain gap.At inference, parameterization is performed for each loop of each key plane for each part.Source: Scan Cross-section Parameterization · p. 5
Adaptive Extrusion.
- Adaptive extrusion tries canonical single-plane candidates first and accepts a reconstruction at the fixed error threshold δ=0.2; otherwise, it uses learned multi-plane candidates and optimizes extrusion heights.Extrusion type is determined from loop nesting relationships, and extrusion direction follows the key-plane normal.Source: Adaptive Extrusion · p. 5
Part Assembly
- The complete CAD model is formed by merging the reconstructed part-level sequences while preserving the parts’ original spatial configuration.Source: Part Assembly · p. 5
CC3D Cross-Section Dataset
CC3D Cross-Section pairs scan-derived cross-section images with corresponding parametric sketches from aligned CAD models, supporting sketch-model adaptation to scanning artifacts.
CC3D Cross-Section Dataset
- The dataset contains 44k training and 2k testing image–sketch pairs, built from CC3D (Cherenkova, Aouada, and Gusev 2020; Mallis et al. 2023) using aligned CAD models and scanned meshes sliced along three axes.CAD slices become constrained sketches, while corresponding mesh slices become rasterized binary images; both are centered and normalized within a unit bounding box.Source: CC3D Cross-Section Dataset · p. 5
- The paired dataset is used to fine-tune the sketch parameterization model and improve its robustness to scanning artifacts.Source: CC3D Cross-Section Dataset · p. 6
Experiments
CADonatello is evaluated on real-world and multi-part CAD benchmarks, where it achieves strong reconstruction results and generalizes to non-canonical extrusion directions. Ablations assess cross-section training data, segmentation methods, and extrusion strategies.
Experiments
- Evaluation covers 2,978 CC3D samples, 200 Fusion 360 assemblies with 3,780 parts, and 2,000 CC3D Cross-Sections samples.Meshes are normalized and segmented with P3-SAM; cross-sections are parameterized by a SketchGraphs-pretrained network fine-tuned on CC3D Cross-Sections.Source: Experiments · p. 6
Comparison with state-of-the-art
- CADonatello achieves the best ECD on full CC3D, the best CD and ECD on CC3D-Complex, and the best results across all metrics on both Fusion 360 settings.On CC3D, it significantly outperforms MiCADangelo and uses 171M parameters versus Cadrille’s 2B while maintaining competitive CD and IR.Source: Table 1 · p. 6, Comparison with state-of-the-art · p. 6, Table 2 · p. 6
- On a 200-solid benchmark of randomly oriented extrusions, CADonatello generalizes well while competing methods degrade substantially.The benchmark tests extrusions along random 3D axes, outside the canonical-axis-aligned directions favored by competing methods’ training data.Source: Comparison with state-of-the-art · pp. 6–7, Table 3 · p. 7
Ablation Study
- Fine-tuning on CC3D Cross-Sections gives the best sketch-parameterization performance among the evaluated training datasets.The comparison includes SketchGraphs-only and SketchGraphs with synthetic noisy augmentations; parameterization improvements also carry over to CC3D 3D reconstruction.Source: Table 4 · p. 7, Ablation Study · p. 7
- Native 3D foundation segmenters yield better reconstruction than 2D-derived or dedicated CAD-trained alternatives, and multi-extrusion improves over single extrusion.PartSAM and P3-SAM achieve better Seg. IoU results; the full extrusion method adds learned plane detection and multi-extrusion optimization to the canonical-plane baseline.Source: Table 5 · p. 7, Ablation Study · p. 7
Conclusion
CADonatello reconstructs parametric CAD models from real-world 3D scans by decomposing objects into parts and reconstructing each through cross-section-driven sketch parameterization and extrusion. Its experiments show state-of-the-art performance on CC3D-Complex and Fusion360-Assembly, while the framework remains limited to extrusion and direct fusion.
Results
- The approach achieves state-of-the-art performance on the CC3D-Complex and Fusion360-Assembly benchmarks while scaling to complex multi-part objects.The conclusion reports these results without providing quantitative values.Source: Conclusion · p. 7
Method
- CADonatello combines foundation-scale 3D part segmentation with cross-section-driven reconstruction to parameterize sketches and extrude structurally meaningful components.It decomposes complex scanned objects and reconstructs their parts independently.Source: Conclusion · p. 7
Dataset
- CC3D Cross-Sections bridges clean CAD sketches and scan-derived cross-sections, improving generalization to realistic scanning artifacts.Source: Conclusion · p. 7
Limitations and future work
- The framework supports only extrusion operations, like prior extrusion-based methods (Khan et al. 2024; Karadeniz et al. 2025b; Rukhovich et al. 2025; Kolodiazhnyi et al. 2025), and assembles parts by direct fusion rather than explicit joints (Willis et al. 2021a).Future work will add operations and assembly-joint prediction.Source: Conclusion · p. 7
Supplementary Material
The supplementary material details CADonatello’s datasets, evaluation metrics, implementation, and additional experiments, and expands the qualitative evaluation. It also documents runtime and filtering, and analyzes failure cases.
Datasets
- Evaluation covers 2,978 CC3D test samples, including 1,528 complex cases, 200 Fusion 360 assemblies with 3,780 parts, and 200 synthetic non-axis-aligned solids.The complex Fusion subset contains 126 assemblies with more than four parts; synthetic solids contain 2–7 extrusion operations with directions tilted up to 35° from the previous direction.Source: Datasets · p. 10
- CC3D Cross-Sections pairs rasterized scan-derived cross-sections with parametric sketches from aligned CAD BRep–scan pairs, yielding 44k training and 2k test pairs.The dataset samples three slicing planes along each canonical axis, filters sketches to at most 16 entities and SCD ≤ 1, and averages 8.3 primitives per sketch.Source: Datasets · p. 10, Figure 5 · p. 10
Evaluation Metrics
- Geometric and validity evaluation uses median mesh Chamfer Distance (CD), edge-focused Chamfer Distance (ECD), and Invalidity Ratio (IR).CD samples 8,192 surface points, ECD samples 4,096 edge points, and IR measures the proportion of predicted CAD models that cannot be converted into valid meshes.Source: Evaluation Metrics · pp. 10–11
- Sketch evaluation includes token-level Accuracy after bipartite matching, Primitive F1 with parameters within 5 quantization units, and connected-component counts using 0.01 endpoint tolerance.PF1 uses 6-bit parameter quantization, while a single connected sketch component is ideal.Source: Evaluation Metrics · p. 11
- Sketch Chamfer Distance (SCD) measures alignment between rasterized 128 × 128 sketch images after centering and unit-bounding-box normalization.The distance is computed between point sets extracted from foreground pixels.Source: Evaluation Metrics · p. 11
Additional Implementation Details
- The implementation segments 30k sampled surface points with P3-SAM, detects planes from 40 slices per canonicalized part, fine-tunes sketch parameterization for 20 epochs, then optimizes extrusion parameters.It discards parts below volume 0.1, caps sketches at 16 primitives, and fits extrusion over 8 candidate directions for 200 iterations before combining parts with FreeCAD Boolean operations.Source: Additional Implementation Details · p. 11
Additional Experiments
- Final reconstruction CD remains relatively stable under moderate segmentation and sketch errors, while part-level CD has a stronger monotonic relationship with final error.The Fusion 360 Assembly analysis indicates gradual pipeline degradation rather than catastrophic compounding of moderate intermediate inaccuracies.Source: Additional Experiments · p. 11, Figure 6 · p. 11
- Fine-tuning on CC3D Cross-Sections modestly improves sketch-level SCD and PF1 and yields small downstream CD and ECD gains on clean Fusion 360 Assembly inputs.Fusion 360 Assembly is used exclusively at test time; its meshes are triangulated from clean BRep models and lack the scanning artifacts targeted by the fine-tuning data.Source: Additional Experiments · pp. 11–12, Table 6 · p. 12