acceptodds
Under review as a conference paper at ICLR 2027

Anastasis: Recovering CAD Modeling History from Meshes via Search over Feature Proposals

12 views

Abstract

CAD reverse engineering is essential for recovering editable models from existing designs. In industrial workflows, meshes are widely used as representations for downstream applications such as simulation. Reconstructing the underlying parametric CAD structure from meshes is therefore important when the original editable model is unavailable. Existing approaches tackle this inverse problem from different perspectives. Geometry-driven methods directly fit the observed shape but rely on weak design priors, often yielding feature decompositions inconsistent with the original modeling logic, whereas learned models offer stronger priors for feature alignment but struggle with precise parameter recovery. We present Anastasis, a mesh-to-CAD framework that recursively explains the input geometry by recovering executable parametric features from proposals and searching over their composition into a modeling history. Specifically, a learned proposal model identifies feature-level surface evidence and Boolean roles from the input mesh, which are then converted into executable parametric features by a geometry-based recovery stage that recovers their continuous parameters. Finally, a learned feature prior ranks and filters the recovered candidates, after which combinatorial search selects and orders them into a modeling history, while unresolved geometric residuals are recursively processed by the same procedure. Experiments on public CAD reconstruction benchmarks show that Anastasis achieves state-of-the-art performance with 0.981 IoU in-distribution, generalizes well across datasets, and produces modeling histories that better reflect the underlying design process.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Mesh-to-CAD, CAD Modeling History, Feature-Based Modeling, Parametric CAD

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_8xCrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents anastasis as a mesh-to-cad reconstruction framework to recover parametric modeling sequences from meshes. Specifically, a mesh-based network first groups surface evidence into candidate features and predicts their boolean roles. A geometric recovery stage then estimates the continuous parameters of extrusion and revolution operations. The recovered candidates are ranked by a learned feature prior, after which subset dynamic programming jointly determines feature selection and execution order. Remaining geometric discrepancies are recursively reconstructed through additional boolean operations. The method is evaluated on several datasets against some recent reconstruction approaches. The reported results show good geometric reconstruction performance, together with improved operation-level recall and competitive runtime. Ablation experiments further examine the contributions of different components.

Critical strengths

The most compelling technical contribution is the subset dynamic programming formulation. By exploiting the observation that the final material state of a spatial point is determined by the last boolean operation on it, the authors express the contribution of an operation in terms of the set of subsequent operations. This enables joint feature selection and ordering with a relatively controlled computational complexity, rather than explicitly enumerating all possible permutations. The proposed formulation provides a principled way to account for interactions between additive and subtractive operations within a bounded candidate set. Although learning-guided CAD search has been investigated previously, this subset-based formulation offers a useful alternative to stepwise and greedy construction strategies.

Critical weaknesses
  1. A central motivation of the paper is to recover parametric CAD modeling histories rather than merely reconstruct final geometry. However, the objective optimized by the proposed subset DP primarily measures geometric agreement between the reconstructed and target solids. For a fixed target, maximizing such objective is equivalent to minimizing the volumetric symmetric difference between the two shapes. This objective does not explicitly evaluate the plausibility, compactness, or editability of the resulting construction sequence. Although the paper introduces operation-level recall, this metric alone does not comprehensively assess modeling-history quality. In particular, it does not explicitly measure operation-order consistency, redundant operations, program complexity, or parameter-level accuracy and editability. These aspects are important when distinguishing a meaningful parametric reconstruction from a geometrically accurate but unnecessarily complicated sequence. This concern is reinforced by the results in Table 3. While geometric reconstruction remains relatively strong on the complex subset, operation-level recall decreases to 0.440 for models containing more than five operations, compared with 0.884 for simple models. I recognize that the original modeling history is not uniquely identifiable from the final geometry, and the paper appropriately acknowledges this ambiguity in Section 3.1. Therefore, exact recovery of the original construction sequence should not be required. However, this makes it especially important to clarify what constitutes a successful recovered history and to evaluate whether the resulting sequences are structurally meaningful and practically editable. As the recovery of modeling histories is central to the paper's stated contribution, the current evaluation leaves an important gap between the research objective and the evidence presented.

  2. The subset-DP formulation is technically interesting, but its effectiveness depends strongly on the quality and coverage of the candidate feature set. As described in Appendix A.3, each search node retains at most 16 candidates, and the search objective is estimated using 16,384 uniformly sampled points. The DP is therefore exact only with respect to the sampled objective over the retained candidate set. However, if a necessary feature is missing from the initial proposals or removed by candidate pruning, the incomplete proposals then can produce distorted residuals, causing errors to accumulate during recursive reconstruction. However, the paper does not systematically examine how performance depends on candidate-set coverage, the top-n budget, the number of interacting operations, or residual recursion depth. Furthermore, the main search ablation compares subset DP with a specific greedy selection strategy. While this experiment demonstrates the benefit of the proposed method over that baseline, it does not establish how much of the advantage is specific to the subset-DP formulation relative to other search strategies, such as beam search or alternative budget-matched combinatorial methods. This comparison is relevant because related approaches, including zone graphs, cadfit, and cadena, have already explored learning-guided or geometry-guided CAD construction and search. The combination of learned geometric evidence and explicit parameter recovery also has methods like Point2Cyl. The distinctive methodological contribution of Anastasis is therefore primarily the proposed search formulation and its integration with feature proposals, rather than the general principle of combining learning, geometric fitting, and program search. A more systematic characterization of the search algorithm's contribution and scaling behavior would considerably strengthen the novelty and practical significance of the paper.

Questions for authors
  1. Could the authors provide additional evidence that the recovered CAD programs are meaningful as parametric modeling histories, beyond achieving high geometric similarity? For example, evaluations of operation precision, sequence complexity, redundant operations, and operation-order consistency would be informative. A small but controlled editability experiment, in which recovered sketch or feature parameters are modified, could also help distinguish geometric fitting from recovery of useful parametric structure.

  2. Could the authors compare subset DP against additional search alternatives, such as beam search, while keeping the feature candidates, learned scores, computational budget, and geometric evaluation procedure fixed? This would clarify whether the observed gains primarily arise from the proposed DP formulation rather than differences in candidate selection or search budget.

  3. How sensitive are the results to the top-n candidate budget, particularly for models involving many interacting operations? An analysis of proposal recall before candidate pruning, together with performance stratified by the number of operations and boolean interactions, would help identify the main bottlenecks. It would also be valuable to quantify how often recursive residual reconstruction successfully recovers initially missing features versus propagating earlier errors.

  4. Section 4.1 states that the predicted and reference models are independently normalized before evaluation. Could the authors clarify how parameter-recovery accuracy is assessed in the original coordinate system, where absolute dimensions may matter? Additionally, the "w/o learned prior" ablation is described as retaining all valid recovered candidates, while Appendix A.3 specifies a maximum of 16 candidates per search node. How is the candidate budget handled in this ablation when more than 16 valid candidates are available?

  5. Given that the method assumes tessellated meshes and relies on surface normals and mesh connectivity for feature recovery, how robust is it to variations in tessellation resolution, geometric noise, and incomplete surface observations? Could the authors provide a controlled sensitivity analysis reporting both geometric accuracy and operation-level recall? This would help clarify whether the reported improvements remain consistent under less ideal input conditions.

Reviewer_75ceOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Anastasis, a mesh-to-CAD reconstruction framework that recovers executable parametric modeling histories from input meshes. Instead of directly predicting CAD sequences, Anastasis combines learned feature proposals with geometry-based recovery of continuous parameters. A learned feature prior and subset dynamic programming are then used to select and order candidate features, while unresolved geometric residuals are handled recursively. Experiments on DeepCAD, FutureCAD, and Fusion360 demonstrate state-of-the-art geometric reconstruction and modeling-history recovery, with strong cross-dataset generalization.

Critical strengths
  1. The paper presents an interesting combination of learned feature proposals, analytical geometric recovery, and combinatorial search. Separating discrete feature discovery from continuous parameter estimation is particularly appealing and provides a meaningful alternative to direct autoregressive CAD generation.
  2. This work achieves substantial improvements over competitive recent baselines, including CADENA, CADFit, and cadrille. Also, very strong qualitative examples.
  3. Beyond conventional geometric metrics, the paper evaluates operation-level recovery using [email protected]. The significant improvements on this metric provide additional evidence that the method recovers meaningful CAD feature structures rather than merely approximating the target geometry.
  4. The ablation studies clearly demonstrate the contributions of surface grouping, learned feature priors, and subset dynamic programming. In particular, the increasing advantage of combinatorial search over greedy selection on complex models supports the methodological motivation.
Critical weaknesses
  1. The current framework primarily supports extrusion and revolution, leaving more sophisticated CAD operations, such as fillets, chamfers, and lofts, unexplored. This limits its applicability to complex industrial CAD models. An industrial application will be limited without a more complex set of operations.
  2. Figure 1 is too informative and can be clearer. Maybe considering split it into more understandable subfigures?
Questions for authors

See above. I do not have a lot of questions. To me this manuscript is a clear acceptance.

Reviewer_wdqXOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper considers a mesh2cad problem, with goal to recover history of construction operations from mesh. Unlike most of current methods, which uses VLM or agent-based approaches, this paper proposes a custom model, which recursively explains the input geometry by recovering executable parametric features from proposals and searching over their composition into a modeling history The model can use extrusion and revolve operations, and demonstrate strong performance on several benchmarks, outperforming current LLM-based approaches.

Critical strengths
  1. The model demonstrate very strong performance, outperforming a number of current SOTA methods, mostly based on LLMs, which is almost a bit suprising, but remarkable.
  1. Interesting technical contributions - separation of feature discovery, parameter fitting and joing sequece search. The subset-DP formulation provides a concrete alternative to greedy assembly, and as shown by ablations studies bring substantial benefit in geometric accuracy for complex examples. Overall, the methods probably would be interesting to analyse in mesh2cad community and can influece new methods.
Critical weaknesses
  1. Method considers only 2 main construction operations - extrusion and revolve. While important, they only 2 of many important operations, and competitive methods (like CADENA) can reconstruct models based on them. So the comparison with other models can confuse the readers, because on benchmarks with limited number of operations model outperform others, but on wider benchmarks its just don't applicable. It also raised a question how the model can be extended to whole set of CAD operations?

  2. On of the key contribution of th ework is focus on joint ordering of operations. However, limited attention is paid on measuring the precision of recovered operation set, execution-order agreement, etc. How well the design process (desing intent) is recovered by the model?

Questions for authors
  1. How model can be extended to the wider set of construction operations? What it will require and how it can influence the performance?

  2. What evidence directly measures the contribution of ordering and the usefulness of the recovered history? Currently, only operation recall is measured.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.