Pith. sign in

REVIEW 5 major objections 5 minor 17 references

Better STEP, a format and dataset for boundary representation

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that any STEP B-rep can be preserved exactly as an open HDF5 dictionary that any program can read.

desk verdict Useful infrastructure paper—a well-specified HDF5 B-rep format and public converted datasets—but the equivalence claim is under-validated because nothing checks OpenCascade's parsing fidelity. read the letter →

arxiv 2506.05417 v1 pith:K6OEIQUS submitted 2025-06-04 cs.CV

Nafiseh Izadyar , Sai Chandra Madduri , Teseo Schneider This is my paper
classification cs.CV
keywords boundaryrepresentationSTEPfilesHDF5formatCADdatasetsparametricsurfaceshalf-edgetopologypointcloudsamplingmachinelearningon
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that boundary-representation CAD files in STEP format can be repackaged, without loss, into an openly specified HDF5 dictionary that any program can read. Such a format would remove the need for proprietary CAD kernels in large machine-learning pipelines, where per-node licenses are costly. The authors converted the Fusion 360 and ABC datasets, plus a million OnShape models, and built a Python library that samples points, normals, and curvature directly from the parametric surfaces. They verify the format's integrity by running four standard learning tasks (normal estimation, denoising, surface reconstruction, and segmentation) with existing models and getting accuracies close to published values.

What carries the argument

The central object is the HDF5 dictionary whose topology section is a standard half-edge data structure: each solid stores shells, each shell stores faces with orientation flags, each face stores loops and a surface index, each loop stores half-edges, each half-edge stores mates, an edge, and a 2D curve index, and each edge stores a 3D curve index and start and end vertices. This hierarchy is what lets the format preserve adjacency and trimming information exactly while remaining readable by any HDF5 client. The conversion tool uses OpenCascade to parse STEP files into this dictionary, and the processing library navigates it like a half-edge structure and samples points, normals, and derivatives from the continuous parametric surfaces.

What would settle it

Convert a set of STEP files with a second CAD kernel that implements its own STEP reader, then compare per-face surface types, parameter domains, and dense sampled point positions to those stored in the HDF5 files; any systematic discrepancy beyond numerical tolerance in surface type or trim curves would falsify the claimed equivalence.

Watch

Extended reading notes

Core claim

The paper claims that a STEP file's B-rep is fully determined by two layers: a geometry layer holding parametric curves and surfaces (with control points, knots, weights, transforms, and trim domains) and a topology layer holding a half-edge hierarchy of solids, shells, faces, loops, half-edges, edges, and vertices, with orientation flags and indices that tie every topological entity to its geometric counterpart. Storing these layers as an HDF5 dictionary, with an optional per-face mesh, gives an equivalent format to the original STEP file. The authors state this format is fully specified, cross-platform, and cross-language, and that their conversion pipeline and processing library make the parametric information directly accessible, something earlier mesh-based pipelines had to reconstruct heuristically.

Load-bearing premise

The load-bearing premise is that OpenCascade parses every STEP file correctly and completely, so the HDF5 dictionary inherits the true geometry and topology; if OpenCascade silently alters a surface or drops a trim curve, the format stores that error faithfully and the paper's validation would not detect it.

Editorial extensions

If this is right

  • Machine-learning pipelines can sample a CAD shape's smooth parametric surfaces directly, obtaining exact point positions, normals, and curvature without meshing the whole model.
  • Datasets converted once (Fusion 360, ABC, OnShape) can be distributed as HDF5 files, so a cluster node needs no CAD kernel license to read or process them.
  • Tasks that need parametric information, such as primitive or degree segmentation, can get it directly from the topology and geometry instead of recovering it from meshes with heuristics.
  • Because the format is fully specified and HDF5-based, other languages and platforms can interoperate with the same files, reducing the version-incompatibility problem of STEP across kernels.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the equivalence is exact, the format could serve as a neutral interchange layer between CAD kernels, so a model stored once could be read by any kernel that implements the dictionary.
  • Editorial inference: a natural stress test would be to convert a random sample of STEP files with a second independent kernel and compare face-level surface types, trim curves, and sampled geometry; the paper's validation does not yet include such a comparison.
  • Editorial inference: because the paper found OpenCascade's mesher fails on roughly 1 to 9 percent of models depending on dataset, downstream users should check which faces have empty meshes; learning on point clouds is unaffected, but mesh-based methods may need fallbacks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes a new HDF5-based dictionary format for boundary representation (B-rep) CAD geometry, together with a converter (STEPTOHDF5) based on OpenCascade, a Python processing library (ABS), and a large converted dataset derived from ABC and Fusion 360. The authors argue that this format is an equivalent, fully specified, open alternative to STEP, removing the need for a proprietary CAD kernel in learning pipelines. They demonstrate the library on four use cases: normal estimation, denoising, surface reconstruction, and segmentation, reporting accuracy numbers that are broadly in line with published results.

Significance. If the equivalence claim is established, the format and dataset would be a valuable community resource for machine learning on CAD geometry, enabling large-scale processing without per-node kernel licenses. The paper's strengths include a detailed format specification in Appendix A, released source code, a substantial conversion effort, and a library that provides parametric sampling, normals, and curvature. However, the validation is currently insufficient to support the central claim: the conversion is never checked against an independent CAD kernel, and the use-case experiments are circular with respect to conversion fidelity. The dataset itself, if properly verified, would be a meaningful contribution; the current manuscript does not yet provide that verification.

major comments (5)
  1. [§3.1, §5, §6] The central claim that the HDF5 format is 'equivalent' to the original STEP B-rep is not validated. Section 3.1 states that STEPTOHDF5 uses OpenCascade to parse and extract all geometric and topological information, with no independent verification against another kernel. The use cases in Section 5 only compare off-the-shelf model accuracies to published values; they never compare the HDF5 geometry or topology against the original STEP files. A silent OpenCascade parsing error (e.g., a misparameterized trim curve, a dropped seam edge, or a changed face adjacency) would be propagated faithfully into the HDF5 file and would remain undetected. Section 6 even acknowledges that 'different STEP files require different kernels,' which is an admission of incomplete coverage, but no attempt is made to measure how often OpenCascade misparses a file or how large the resulting deviations are. The authors should add a direct comparison between the HDF5 representation and the original STEP geometry/topology, ideally using an independent kernel or a set of analytic ground-truth shapes, with quantitative error metrics.
  2. [§5, Listings 3–5] The ground-truth normals, points, and labels used in all four use cases are computed by ABS from the HDF5 files produced by the same OpenCascade-based pipeline, so those experiments are circular with respect to conversion fidelity. They cannot catch a conversion error: if OpenCascade silently distorts a surface, the sampled points and normals will be consistent with the distorted surface, and the downstream model may still produce plausible outputs. The comparison with BPNet's published accuracy in Table 2 provides some external anchoring, but it is not a test of geometric equivalence. At least one experiment should use an independent source of ground truth, such as evaluating the analytic parametric surfaces directly or reading the original STEP file with a different kernel, to confirm that the HDF5 data matches the STEP geometry.
  3. [Table 1] The surface-reconstruction results in Table 1 are reported without error bars or any measure of variance, and several entries are non-monotonic and counterintuitive. For example, at 4,000 samples the Chamfer distance is 3.39 at σ=0 but 1.91 at σ=0.005, and at 8,000 samples it is 1.63 at σ=0 but 2.56 at σ=0.025. Such values suggest either high variance across the 1,000 selected models or sensitivity to the specific random sampling. Without means and standard deviations over multiple runs or bootstrap resamples, the table does not provide convincing evidence that the format produces data consistent with the method's expectations. This is load-bearing because the table is presented as evidence of the format's integrity.
  4. [§4, Table 2] The reported meshing failure rates (1.56% for ABC, 8.82% for Assembly, 1.04% for Joint, 0.02% for Reconstruction, 0.07% for Segmentation) show that OpenCascade's meshing is not fully robust, yet the paper does not explain how these failures affect dataset usability. It is unclear whether a model with a failed mesh is skipped entirely, whether only some faces are missing, and whether the failure rate correlates with model complexity or face type. This matters because the format stores a mesh per face (Appendix A.3), and a failed mesh means the HDF5 file contains an empty dataset, which downstream users must handle. The authors should clarify the handling of failed meshes and quantify the impact on the dataset's completeness.
  5. [§1, §4] The introduction claims 'we convert the Fusion 360 and ABC datasets and add another million models from OnShape,' but the dataset section (§4) lists only one million ABC models and the four Fusion 360 subsets; the OnShape portion is never described in terms of size, selection criteria, or conversion statistics. This discrepancy makes the dataset's actual composition unclear. Please clarify whether the released dataset includes OnShape models and, if so, provide the corresponding details and statistics.
minor comments (5)
  1. [Listing 1] The comment above Listing 1 says 'Example of computing normal at every point,' but the code computes a binary label and returns 1 or 0, not a normal. The comment should be corrected, or the listing should be replaced with the normal-computing code from Listing 3.
  2. [Listing 5] Listing 5 contains Python syntax errors: 'elif :' on line 11 has a misplaced colon, and line 13 'degree [(2,3),(3,2)]' is missing an equals sign. Since the paper demonstrates the library's ease of use, the code should be syntactically valid.
  3. [§3.1 and Figure 2] The caption of Figure 2 says 'Point-cloud for a model where OpenCascade fails to generate a mesh,' while the text says 'approximately 5% of the models fail to produce a mesh.' It is unclear whether the failure is per model or per face, and whether the point cloud in Figure 2 is generated by an alternative fallback or by sampling the parametric surfaces. Please clarify.
  4. [Appendix A.1] In the description of b-spline curves and surfaces, the text says 'We also track if the curve is periodic or if it closed' and later 'if it is u_closed/v_closed.' The first should read 'if it is closed' for grammatical consistency.
  5. [§4, Conclusion] The dataset section says 'Our dataset includes one million models from ABC' and the Fusion subsets, but the conclusion says 'several million models' have been converted. The total count should be made consistent, and the OnShape contribution, if any, should be stated explicitly.

Circularity Check

1 steps flagged · score 2.0 of 10

Use-case validation is self-referential: 'compliance with STEP' is checked only against data produced from the converted HDF5 files, so conversion errors would be invisible.

  1. other [Abstract; Section 5 (Use Cases), Listings 3 and 5; Section 3.2]
    "We developed four standard use cases (normal estimation, denoising, surface reconstruction, and segmentation) to assess the integrity of the data and its compliance with the original STEP files. ... Note that we do not fine-tune or retrain the models; we evaluate them directly using our dataset."

    The claimed 'compliance with the original STEP files' is assessed only against data produced from the converted HDF5 dictionary itself. Ground-truth normals are computed by topo.normal(points) (Listing 3), segmentation labels are derived from topo.surface.shape_name (Listing 5), and the reconstruction metric compares against 'a dense sampling (15 000 points) of the parametric surfaces' stored in the same HDF5 file (Section 5). No experiment compares the converted B-rep back to the original STEP file or to an independent CAD kernel.

full rationale

The central format construction is not circular: the HDF5 dictionary is a direct serialization of OpenCascade's parsed B-rep (Section 3.1, Appendix A), and no fitted parameter or prediction is defined in terms of the target result. There are no load-bearing self-citations, imported uniqueness theorems, or ansatz-via-citation steps. The only circular-adjacent element is the validation claim: the four use cases are said to assess compliance with the original STEP files, but every ground-truth normal, label, and reference surface is generated by the same ABS library from the converted HDF5 files, so the experiments cannot detect conversion errors. This is a genuine validation loop, but it is not load-bearing for the format's internal construction—the format's equivalence to STEP is assumed from OpenCascade's parsing rather than derived from the use cases. The paper's own conclusion acknowledges that different STEP files require different kernels, which is a correctness limitation rather than a circularity. Therefore the circularity score is low, reflecting a minor self-referential validation rather than a forced derivation.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim depends on OpenCascade's parsing of STEP (a domain assumption), on the validity of the source datasets, and on standard properties of half-edge topology and parametric surface equations. No free parameters are fitted; the experimental thresholds and point counts are evaluation settings, not parts of the format. No new entities are postulated.

assumptions (4)
  • domain assumption OpenCascade correctly and completely parses STEP files into the geometric and topological entities stored in the format.
    Section 3.1: STEPTOHDF5 uses OpenCascade to parse and extract all information; no independent verification against another kernel is provided.
  • domain assumption The source STEP files in ABC and Fusion 360 are valid and represent the intended B-reps.
    The converter reads these files as-is; the paper does not repair or validate the source B-reps beyond OpenCascade's processing.
  • standard math The half-edge topology with only top-down links is sufficient to reconstruct reverse adjacency information.
    Appendix A.2 stores shell-to-face and loop-to-half-edge links, and the library recovers mates and other reverse links, which is standard for half-edge structures.
  • standard math The parametric equations in Appendix A correctly represent the geometry of each surface and curve type.
    These equations are used for sampling and normal computation; their correctness is standard mathematical knowledge for these primitive types.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Better STEP, a format and dataset for boundary representation." pith.science (2026). https://pith.science/paper/K6OEIQUS

@misc{pith2026250605417,
  author       = {Pith},
  title        = {Pith review of: Better STEP, a format and dataset for boundary representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K6OEIQUS}},
  note         = {Machine review of arXiv:2506.05417}
}
read the original abstract

Boundary representation (B-rep) generated from computer-aided design (CAD) is widely used in industry, with several large datasets available. However, the data in these datasets is represented in STEP format, requiring a CAD kernel to read and process it. This dramatically limits their scope and usage in large learning pipelines, as it constrains the possibility of deploying them on computing clusters due to the high cost of per-node licenses. This paper introduces an alternative format based on the open, cross-platform format HDF5 and a corresponding dataset for STEP files, paired with an open-source library to query and process them. Our Python package also provides standard functionalities such as sampling, normals, and curvature to ease integration in existing pipelines. To demonstrate the effectiveness of our format, we converted the Fusion 360 dataset and the ABC dataset. We developed four standard use cases (normal estimation, denoising, surface reconstruction, and segmentation) to assess the integrity of the data and its compliance with the original STEP files.

Figures

Figures reproduced from arXiv: 2506.05417 by the authors.

Figure 1
Figure 1. Examples of a few models from our different datasets. We also randomly sample the models [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Point-cloud for a model where OpenCascade fails to generate a mesh. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Hierarchical structure of our format. From the root structure (solid) to the leaf (vertices). [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Example of a complex fan model (left) sampled with an increasing number of points. As [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Example of point clouds sampled using Listing 1; we highlight the feature edges in yellow. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Example of models with thousands of patches. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Distribution of of faces per model for the different datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Average distribution of faces (top) and curves (bottom) types across the different datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 9
Figure 9. Figure 9: Example of denoising point cloud using PathNet [12]. [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: Example of reconstructed surfaces from randomly sampling our dataset using NKSR [ [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 9 canonical work pages

  1. [1]

    Deepfit: 3d surface fitting via neural network weighted least squares

    Yizhak Ben-Shabat and Stephen Gould. Deepfit: 3d surface fitting via neural network weighted least squares. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 20–34. Springer, 2020

    work page 2020

  2. [2]

    Open cascade.https://dev.opencascade.org/

    Open CASCADE. Open cascade.https://dev.opencascade.org/

    work page

  3. [3]

    Shapenet: An information-rich 3d model repository.arXiv preprint arXiv:1512.03012, 2015

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository.arXiv preprint arXiv:1512.03012, 2015

    arXiv 2015

  4. [4]

    Constraint-Aware Feature Learning for Parametric Point Cloud

    Xi Cheng, Ruiqi Lei, Di Huang, Zhichao Liao, Fengyuan Piao, Yan Chen, Pingfa Feng, and Long Zeng. Constraint learning for parametric point cloud.arXiv preprint arXiv:2411.07747, 2024

    work page Pith review arXiv 2024

  5. [5]

    Cherenkova, Anis Kacem, Sk Aziz Ali, Ilya Arzhannikov, Gleb Gusev, and D

    Elona Dupont, K. Cherenkova, Anis Kacem, Sk Aziz Ali, Ilya Arzhannikov, Gleb Gusev, and D. Aouada. Cadops-net: Jointly learning cad operation types and steps from boundary- representations.2022 International Conference on 3D Vision (3DV), pages 114–123, 2022. doi: 10.1109/3DV57658.2022.00024

    arXiv 2022

  6. [6]

    BPNet: B\'ezier Primitive Segmentation on 3D Point Clouds

    Rao Fu, Cheng Wen, Qian Li, Xiao Xiao, and Pierre Alliez. Bpnet: Bézier primitive segmenta- tion on 3d point clouds.arXiv preprint arXiv:2307.04013, 2023

    work page Pith review arXiv 2023

  7. [7]

    Neural kernel surface reconstruction

    Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4369–4379, 2023

    work page 2023

  8. [8]

    Jayaraman, Aditya Sanghi, J

    P. Jayaraman, Aditya Sanghi, J. Lambourne, T. Davies, and Hooman Shayani. Uv-net: Learning from boundary representations.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11698–11707, 2021. doi: 10.1109/CVPR46437.2021.01153

    arXiv 2021

Show all 17 references
  1. [9]

    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, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. InThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    2019

  2. [10]

    Lambourne, Karl D.D

    Joseph G. Lambourne, Karl D.D. Willis, Pradeep Kumar Jayaraman, Aditya Sanghi, Peter Meltzer, and Hooman Shayani. Brepnet: A topological message passing system for solid models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1...

    2021

  3. [11]

    Qi, Hao Su, Kaichun Mo, and L

    C. Qi, Hao Su, Kaichun Mo, and L. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation.2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 77–85, 2016. doi: 10.1109/CVPR.2017.16

    2017 doi

  4. [12]

    Pathnet: Path-selective point cloud denoising.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4426–4442, 2024

    Zeyong Wei, Honghua Chen, Liangliang Nan, Jun Wang, Jing Qin, and Mingqiang Wei. Pathnet: Path-selective point cloud denoising.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4426–4442, 2024

    2024

  5. [13]

    Karl D. D. Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G. Lambourne, Armando Solar-Lezama, and Wojciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences.ACM Transactions on Graphics (TOG), 40(4), 2021

    2021

  6. [14]

    Joinable: Learning bottom-up assembly of parametric cad joints.arXiv preprint arXiv:2111.12772, 2021

    Karl DD Willis, Pradeep Kumar Jayaraman, Hang Chu, Yunsheng Tian, Yifei Li, Daniele Grandi, Aditya Sanghi, Linh Tran, Joseph G Lambourne, Armando Solar-Lezama, and Wojciech Matusik. Joinable: Learning bottom-up assembly of parametric cad joints.arXiv preprint arXiv:2111.12772, 2021

    2021 arXiv

  7. [15]

    Deepcad: A deep generative network for computer- aided design models.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6752–6762, 2021

    Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer- aided design models.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6752–6762, 2021. doi: 10.1109/ICCV48922.2021.00670

    2021

  8. [16]

    Brep2seq: a dataset and hierarchical deep learning network for reconstruction and generation of computer-aided design models.J

    Shuming Zhang, Zhidong Guan, Hao Jiang, Tao Ning, Xiaodong Wang, and Pingan Tan. Brep2seq: a dataset and hierarchical deep learning network for reconstruction and generation of computer-aided design models.J. Comput. Des. Eng., 11:110–134, 2024. doi: 10.1093/jcde/ qwae005

    2024 doi

  9. [17]

    Thingi10k: A dataset of 10,000 3d-printing models.arXiv preprint arXiv:1605.04797, 2016

    Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models.arXiv preprint arXiv:1605.04797, 2016. A File format The root of the HDF5 file includes one unique group called parts and has one string attributeversion (currently version 2.0). The part group c...

    2016 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.