Source-linked AI summary
ABC: A Big CAD Model Dataset For Geometric Deep Learning
Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, Daniele Panozzo
TL;DR
Geometric deep learning lacks large, high-quality datasets with reliable geometric ground truth. The paper introduces a million-scale parametrically represented CAD dataset and processing pipeline, then benchmarks surface-normal estimation; its comparison finds that point-only deep methods outperform analytical methods, whereas connectivity-aware methods do not.
Problem
High-quality geometric datasets and reliable ground truth for differential properties and geometric features remain limited because 3D data are irregular and difficult to annotate.
Method
The paper builds a CAD dataset of parametrically represented models, extracts analytic and topological ground truth, and provides a pipeline for deep-learning representations and surface-normal benchmarking.
Results
Deep learning methods using only 3D points outperform analytical methods, but connectivity-aware methods are considerably worse than the uniform analytical method.
Takeaways & Limitations
The dataset supports large-scale benchmarks and serves as a stress test for robust geometry-processing algorithms.
Takeaways & Limitations
The study trains methods to report a single normal per point and leaves sharp-edge and corner cases with multiple normals for future work.
Abstract
from arXiv · showhide
We introduce ABC-Dataset, a collection of one million Computer-Aided Design (CAD) models for research of geometric deep learning methods and applications. Each model is a collection of explicitly parametrized curves and surfaces, providing ground truth for differential quantities, patch segmentation, geometric feature detection, and shape reconstruction. Sampling the parametric descriptions of surfaces and curves allows generating data in different formats and resolutions, enabling fair comparisons for a wide range of geometric learning algorithms. As a use case for our dataset, we perform a large-scale benchmark for estimation of surface normals, comparing existing data driven methods and evaluating their performance against both the ground truth and traditional normal estimation methods.
1. Introduction
The paper addresses the scarcity of high-quality geometric datasets and reliable ground truth by introducing a large CAD dataset with explicit parametric representations and a surface-normal benchmark.
- 3D geometric datasets remain limited because high-quality models are difficult to create and surface data are inherently irregular.Existing datasets also often lack reliable ground truth annotations.
- Ground truth is difficult to generate for differential properties and manually marked geometric features.Differential properties are usually approximated from sampled surfaces, while feature annotation is laborious.
- Over 1 million high-quality geometric models are represented by parametric surfaces with patch, sharp-feature, and analytic differential annotations.The representation supports resampling at arbitrary resolutions, with or without connectivity, as point clouds or meshes.
- The benchmark evaluates surface-normal estimation locally and globally using seven existing deep learning algorithms as dataset size increases.The benchmark compares data-driven methods with traditional normal estimation approaches.
- An open-source processing pipeline converts CAD models into representations suitable for deep learning and supports comparisons on realistic manmade objects.The dataset targets geometric deep learning applications involving human-created mechanical shapes.
2. Related Work
Related work spans geometric datasets, point-cloud and graph-based neural networks, and analytical surface-normal estimation methods.
- 3D Deep Learning Datasets: Existing geometric datasets support tasks such as segmentation, classification, large-scene analysis, mesh registration, and 2D/3D alignment.The proposed dataset is distinguished by its analytic surface and curve representations for quantitative evaluation.
- Point Cloud Networks: Point-cloud networks use permutation-invariant processing, local patches, volumetric mappings, or learned neighborhood operations.PointNet and PointNet++ avoid dependence on point ordering, while PCPNet targets local shape properties.
- Networks on Graphs and Manifolds: Graph and manifold networks apply spectral, spatial, or differential-geometry operators to graph-structured and surface data.Spatial filters are described as more suitable for localized quantities.
- Analytic Normal Estimation Approaches: Analytical normal estimators fit tangent planes or average adjacent triangle normals using uniform, area, or angle weighting.These methods operate on point-set neighborhoods or mesh connectivity.
- Analytic Normal Estimation Approaches: Higher-order estimators fit polynomial surfaces over tangent planes, optionally using robust statistics to weight neighborhood points.Osculating jets recompute normals from the tangents of the fitted polynomial surface.
- Scope: The study focuses on selected normal-estimation techniques and leaves multiple-normal prediction at sharp edges and corners for future work.The evaluated machine-learning methods report only one normal per point.
3. Dataset
The dataset contains over one million CAD models in parametric B-Rep form, with ground-truth geometry and topology that can be resampled and converted into learning-ready formats. An open-source pipeline loads, discretizes, labels, filters, and exports these models, while acknowledging an uncurated source collection and CAD-specific scope.
- 3. Dataset: The proposed dataset satisfies large size, ground-truth, parametric representation, and expandability requirements, while focusing on CAD models with substantial within-class shape variation.The CAD restriction trades broader model coverage for similar samples and high-quality ground truth.
- 3. Dataset: Over 1 million models were collected from Onshape’s publicly available CAD collection over four months.The source models are freely usable for research.
- 3. Dataset: The vectorial representation supports arbitrary-resolution resampling and analytical computation of differential signals used as ground truth.This representation can be converted into discrete learning formats while retaining correspondences to the continuous model.
- 3. Dataset: CAD models are encoded as B-Reps containing topology and geometric entities such as surfaces, curves, points, faces, edges, and vertices.Surfaces include analytic and NURBS-based types, while curves include standard conics and NURBS curves.
- 3.2. Processing Pipeline: The uncurated Onshape collection includes defective, duplicate, and primitive-only models, so geometric and topological criteria are used to filter low-quality entries.Semantic, crowd-sourced filtering and annotation remain future work.
- 3.2. Processing Pipeline: The processing pipeline translates STEP B-Reps, samples and triangulates patches with uniform or curvature-adaptive options, transfers topology labels, and supports quality-based post-processing.Labels identify surface-patch membership for triangles and sharp feature edges for mesh edges.
4. Supported Applications
ABC supports geometric learning and geometry-processing applications through explicit surface, curve, feature, and differential ground truth. Its applications include vectorization, differential-quantity estimation, feature detection, reconstruction, image generation, and robustness evaluation.
- 4. Supported Applications: The dataset is intended as training data or a benchmark for applications benefiting from geometric ground truth.The paper highlights tasks spanning surface representations, differential quantities, features, reconstruction, images, and robustness.
- Surface Vectorization: B-Rep surface vectorization provides an arbitrary-resolution counterpart to image vector representations, supporting research on mesh-to-B-Rep conversion.Data-driven methods for this conversion are described as being in their infancy.
- Estimation of Differential Quantities: Analytic normals and curvature values enable objective evaluation of differential-quantity prediction on point clouds or triangle meshes.The ground truth avoids relying solely on approximations from sampled surfaces.
- Sharp Feature Detection: Explicit topological encoding of sharp features provides ground truth for sharp-feature prediction on point clouds and meshes.Feature lines also separate the naturally defined surface patches in each object.
- Shape Reconstruction: Known B-Rep geometry supports simulated scanning and quantitative evaluation of reconstruction and point-cloud upsampling errors.The same exact geometry supplies the reference for evaluating reconstructed outputs.
- Image Based Learning Tasks: The dataset’s rendering module generates ground-truth image data across poses, lighting, materials, cameras, and depth, color, or contour modes.The images have no geometric approximation error.
- Robustness of Geometry Processing Algorithms: Normal estimation remains a robustness challenge: most evaluated methods fail on at least one model by producing zero-length or NaN normals.The dataset is presented as a setting for studying and solving these failures.
5. Normal Estimation Benchmarks
The benchmark evaluates local-patch and full-model normal estimation using CAD models with ground-truth normals, varying dataset size, sampling resolution, and input connectivity. It finds modest gains from more training samples, resolution-dependent learning behavior, and a strong advantage for analytic methods when connectivity is available.
- Benchmark design: The benchmark evaluates normal estimation on local patches and full models, exploiting ground-truth normals and varying model sampling densities.It is presented as a large-scale study intended to inform both data-driven and analytic methods.
- Benchmark design: Datasets contain randomly sampled patches ranging from 512 vertices to entire models, with four training-set sizes from 10k to 250k patches.All benchmark datasets use an 80% training and 20% test split.
- Evaluation protocol: The study compares 12 representative algorithms—five traditional and seven machine-learning methods—using a common normal-estimation loss.The loss does not penalize normals flipped by 180°, treating orientation as a separate postprocessing issue.
- Results: Increasing training-set size reduces error for all methods on patches and full models, but the improvement is modest.Figures compare angle-deviation error at sample sizes including 10k, 50k, and 100k.
- Results: DGCNN, PCNN, and PwCNN benefit from higher full-model sampling resolution, whereas PN++ shows no clear improvement and Laplace Surface Network fails to converge at the highest resolution.The authors link PN++’s behavior to spatial subsampling that preserves sublinear training time but prevents use of extra resolution.
- Results: Analytic methods remain consistent across dataset sizes and improve with resolution, while connectivity-based methods heavily outperform K-nearest-neighbor methods and learning methods when connectivity is available.The same data-driven-versus-analytic pattern is also observed on a synthetic NURBS-patch dataset.
6. Conclusion
The paper contributes a large CAD model dataset and conversion tools, then demonstrates that extracted ground truth supports large-scale differential-quantity benchmarks. The normal-estimation comparison offers guidance for geometric deep learning, while exposing robustness problems in geometry processing.
- The dataset and conversion tools support large-scale learning benchmarks using ground-truth signals extracted from CAD models.The authors demonstrate this capability with differential surface quantities.
- Deep-learning methods using only 3D points outperform analytical methods, unlike methods that also use connectivity information.The comparison indicates that connectivity-aware graph architectures do not exploit connectivity effectively in this benchmark.
- Connectivity-aware methods are considerably worse than the uniform analytical method, which averages normals of neighbouring faces.
- Roughly 100 models produce invalid normals with the uniform algorithm because of floating point errors.The dataset’s size and complexity make it a stress test for robust geometry processing algorithms.
7. Distribution
The dataset and associated information are publicly distributed under specified licensing and packaging arrangements. The processing pipeline is separately released as containerized software.
- The dataset and all associated information are available online under the MIT license.Data are split into chunks of 10k models for each data type.
- Model copyrights remain with their respective creators, while the geometry processing pipeline is released under the GPL license.The pipeline is provided as Docker and Singularity containers runnable on suitable machines.
Supplementary Material
The supplementary material describes filtering, representations, and processing choices for the CAD models. Models are provided in explicit CAD formats and discrete meshes, with the pipeline enabling resampling and differential-quantity extraction.
- Figure 8 illustrates models with differently colored patches and highlighted sharp feature curves.The left view highlights sharp feature curves, while the right shows all feature curves.
- Automatic filters remove empty, primitive-only, shape-healing, and excessively large models from the Onshape collection.Users can modify the filters for their applications.
- Each ABC-Dataset model is stored in three representations and multiple filetypes.
- STEP and Parasolid preserve explicit CAD topology and geometry, while processing supports arbitrary-resolution sampling, mesh generation, and differential-quantity extraction.
- The Onshape-generated STL meshes are faithful approximations but may have poor aspect ratios and highly non-uniform sampling.
- The pipeline-generated OBJ meshes provide a fairly regular alternative with uniform vertex distribution and most triangle angles near 60°.
B.3. Curve and Patch Features
The dataset stores CAD curves and patches through explicit geometric types, parameters, mesh references, and feature annotations. These descriptions cover standard primitives, spline representations, and surfaces formed by revolution or extrusion.
- YAML files store surface and curve definitions together with references to corresponding mesh vertices and faces.The files contain patch lists and curves describing patch boundaries.
- Curves are categorized as Line, Circle, Ellipse, BSpline, or Other, with type-dependent geometric information.Curve records may include sharp-feature status and mesh correspondence.
- Curves: Circle and ellipse records encode locations, local axes, radii, focal points, and plane normals as applicable.Their parametric forms use cosine and sine coordinates in local systems.
- Curves: B-Spline curves are represented by control points, knots, weights, rationality, closure, continuity, and polynomial degree.
- Patches: Patches are categorized into analytic, revolution, extrusion, B-Spline, and other surface types, with mesh indices and surface parameters recorded.
- Patches: Plane, cylinder, cone, sphere, and torus patches use explicit parametric equations with locations, axes, radii, angles, and quadric coefficients where applicable.Revolution surfaces rotate a curve around an axis, while extrusion surfaces extend a curve along a direction.
- Patches: Cylinders, cones, spheres, and tori record coordinate axes and shape-specific parameters such as center axes, apexes, radii, and angles.
- Patches: Revolution and extrusion patches retain the generating curve plus either a rotation axis or linear extrusion direction.B-Spline patches additionally store rationality, closure, continuity, degrees, poles, and knot vectors in both parameter directions.
C. Implementation Details
The comparison standardizes training and model settings across learning-based and traditional normal-estimation methods. Neural methods generally adapt segmentation architectures to predict normalized three-dimensional normals, while analytic methods use specified neighborhood or fitting parameters.
- Common settings: All compared methods use the cosine loss and a maximum allowed time of 3 days.The overall parameter counts are listed in Table 3, and running times in Table 5.
- Learning-based methods: Point Convolutional Neural Networks produce 3 outputs instead of 10, with input-layer sizes matched to input size and batch sizes of 32 or 16.Training uses stochastic gradient descent with step size 10^-3.
- Learning-based methods: Surface Networks train for 300 epochs with Adam, starting at learning rate 10^-3 and halving it every 20 epochs after epoch 100.Degenerate-triangle Laplacian coefficients that overflow are replaced with 1.
- Learning-based methods: PointNet++, DGCNN, PwCNN, and PointCNN adapt segmentation architectures to output 3 values representing normalized normals.The methods use 100 training epochs and batch sizes of 16, except PointCNN, which uses batch size 8.
- Learning-based methods: PCPNet retains its original architecture and trains with Adam at step size 1 × 10^-3 for up to 2000 epochs without weight decay.Its optimizer settings use β = (0.9, 0.99) and ϵ = 1 × 10^-8.
- Traditional methods: Traditional methods use author defaults or local neighborhoods: 10 nearest neighbors for plane fitting, jet fitting, or normal refinement.Mesh versions use default parameters for RoSt and Osculating Jets implementations.
D. Running Times
Running times are reported separately for data-driven training and analytic normal estimation, with different measurement conditions. Training is capped at three days, whereas analytic runtimes are measured on one CPU core.
- Timing protocol: Training times for data-driven methods are reported in Table 4, while analytic-method runtimes are reported in Table 5.The two categories therefore use separate timing tables and evaluation procedures.
- Analytic methods: Analytic-method runtimes are measured on one Intel Core i7 CPU core for normal estimation.The training-time measurements use different machines and GPU models.
- Data-driven methods: Training time is capped at 4320 minutes, and this timeout occurs for all PCPNet and ExtOp trainings.Table 4 records GPU coloring and notes when more than two GPUs were used serially.
- NURBS timing-related benchmark: Figure 9 plots angle-deviation error for the 512-point NURBS benchmark at sample sizes of 10k and 50k.The figure compares error curves across the two sample-size settings.
E. NURBS Patches Experiment
The additional experiment evaluates normal estimation on synthetic random bi-variate BSpline patches, where restricted shape variation favors data-driven methods. It varies dataset size and sampling density to compare performance across conditions.
- Dataset construction: The experiment uses synthetic datasets containing 10k and 50k random bi-variate BSpline surface patches.The restricted shape space is intended to make normals easier to infer and gives data-driven methods an advantage over traditional methods.
- Sampling densities: Each synthetic dataset contains patches sampled at 512, 1024, and 2048 vertices.Figure 11 shows example NURBS patches at the three sampling densities.
- Evaluation: Figure 10 plots angle-deviation error for the medium-resolution 1024-point NURBS benchmark at sample sizes of 10k and 50k.The plot is organized from top to bottom by the two sample-size settings.