acceptodds
Withdrawn from ICLR 2027

Few-Shot B-rep Local Structure Retrieval with Structure-Aware Pretraining and Hard-Example Fine-Tuning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing B-rep retrieval methods primarily model whole-part similarity, whereas engineering reuse often requires finding complete models that contain a desired local structure. This setting is challenging because target structures can vary substantially in geometry and topology, especially when a few query examples are available. We present LocBrep, a few-shot local-structure retrieval framework for B-rep models. It couples structure-aware pretraining and hard-example fine-tuning with a coarse-to-fine retriever to learn transferable local representations, accurately retrieve specific parts, and support retrieval for CAD assemble. LocBrep first employs SmoothDiff pretraining, which combines multi-scale neighborhood consistency with topological positional regularization to obtain robust face-level representations. It then performs query-guided hard-example fine-tuning to adapt the encoder to a target structure from limited examples. For scalable retrieval, hyperedge-level coarse search recalls candidate parts, and a query-aware cross-graph reranker refines their ranking according to local structural correspondence. We further introduce LocBrep-Bench, comprising 4500 B-rep models from 50 local-structure categories with face-level annotations. LocBrep outperforms existing methods on FabWave and LocBrep-Bench under both zero-shot and one-shot settings, with up to a 15.71 percentage-point improvement in MAP. Experiments on Fusion 360 further validate its applicability to retrieval in CAD assemblies.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
B-rep Retrieval, Graph, CAD, Few-Shot learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LoBTOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper achieves local structure retrieval via the proposed 2-stage training and a coarse-to-fine retriever. The method is evaluated under both 0-shot and 1-shot.

Critical strengths
  1. Combination of this 2-stage training with the 0-shot and 1-shot settings makes sense.
  2. The improvements brought by the 2 stages are reasonably demonstrated.
Critical weaknesses
  1. i'm concerned about the coarse in retriever. How sensitive is the final performance to this candidate-pool size? since once missed at this stage, it can never get back. It would be good to report results with different candidate-pool size,for example, top-20/50/100 after coarse retrieval.
  2. Though the proposed method achieves the best results, the final performance is still far from perfect. Could the authors provide a systematic failure analysis to clarify where it's from?
  3. I think the presentation of the method could be clearer. The relationship between pretraining, one-shot fine-tuning, 0-shot/1-shot evaluation does take some effort to understand. It may be clearer to directly show them in fig. 2.
Questions for authors
  1. There are global citation formatting issues, ie, missing parentheses or spacing.
  2. How well does the method generalize to local structures that are unseen during pretraining, rather than downstream?
Reviewer_SwyFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces LocBrep, a framework that retrieves complete 3D CAD models containing a specific local structure, such as a hole or slot, even when only a few query examples are available.

Its main contributions include a new pretraining method for learning local geometric and topological features, a hard-example mining strategy for few-shot learning, an efficient coarse-to-fine retrieval system, and LocBrep-Bench, a benchmark containing 4,500 CAD models across 50 local-structure categories.

The authors evaluate LocBrep on FabWave, LocBrep-Bench, and Fusion 360 CAD assemblies. Results show that it outperforms existing methods in both zero-shot and one-shot settings, with up to a 15.71 percentage-point improvement in MAP. Ablation studies further support the effectiveness of its main components.

Critical strengths
  1. LocBrep focuses on finding CAD models that contain a specific local structure, rather than only matching overall shapes. This better reflects real engineering needs and can help designers reuse existing parts with less manual work.

  2. The method is tested on two retrieval benchmarks and Fusion 360 CAD assemblies, showing better performance than existing methods in both zero-shot and one-shot settings. The ablation studies also support the value of its main components, making the results more convincing.

Critical weaknesses
  1. LocBrep-Bench is created by combining existing CAD components with base models. Although this makes evaluation easier to control, these constructed models may not fully represent the complexity of real-world engineering designs. This limits confidence in how well the method will perform in practical applications.

  2. Most baselines are originally designed for whole-model retrieval or classification, rather than local-structure retrieval. The paper also does not compare against external methods in the Fusion 360 assembly experiment. Therefore, it is still unclear how much LocBrep improves over existing approaches specifically designed for local structure matching.

Questions for authors
  1. Since LocBrep-Bench is built by combining existing CAD components, how well does it represent real-world engineering designs? Have the authors tested whether the method performs similarly on real, unmodified CAD models?

  2. Could the authors compare LocBrep with more methods specifically designed for local-structure retrieval and include baseline comparisons on Fusion 360 assemblies? This would help clarify the method's advantages in practical applications.

Reviewer_rBjzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LocBrep, claimed as the first learning-based framework for few-shot B-rep partial retrieval: given a masked local face sub-graph as query, retrieve complete CAD models containing a matching local structure. The pipeline has three stages: (1) BrepDenoise, a conditional diffusion pretraining over face features on the hybrid-graph backbone adopted from PP-Brep (Hao et al., 2026), explicitly adapting the low-frequency conditional diffusion formulation of SDMG (Zhu et al., 2025), plus a multi-scale neighborhood-consistency loss (Eqs. 2–3) and topological positional regularization (Eq. 4); (2) label-free one-shot fine-tuning that mines hard examples transductively from the full downstream candidate corpus, adapting AUGCL's affinity-partition uncertainty estimation, with DINOv2 multi-view rendering filtering; (3) a coarse-to-fine retriever with hyperedge-level indexing and a cross-graph reranker trained with supervised BCE on MFInstSeg instance labels. Evaluation uses the new LocBrep-Bench (4,500 Boolean-union composites on DeepCAD bases, 50 categories), an author-re-annotated FabWave, and a baseline-free Fusion 360 Gallery Assembly experiment (8,251 assemblies). Against four author-adapted whole-model baselines, the paper reports zero-shot MAP 64.72 vs BRT's 49.01 on LocBrep-Bench — the abstract's "up to 15.71 percentage-point" claim — and one-shot MAP 80.32 vs 72.63.

Critical strengths

A real, apparently unoccupied problem, with numerically careful and unusually candid execution. No learning-based B-rep partial-retrieval prior work was found in an active web sweep this session; the nearest verified candidates are handcrafted (a descriptor-based dual-level CAD retrieval approach, CAD&A 2023) or whole-model (GC-CAD, a self-supervised whole-part retrieval preprint), so the "first" claim survives as scoped. A full arithmetic audit found every claimed delta in Sections 5.3–5.4 and the abstract exact against Tables 1–8 — no numerical errors anywhere. The paper also discloses its own vulnerabilities: Appendix E.3 states plainly that no baseline gets DINOv2 filtering or transductive corpus mining, E.4 explains the exclusion of classical partial-retrieval systems, A.2 declines causal over-reading of its own spectral statistics, and all results average five fixed seeds.

Genuine internal controls beyond the norm. Table 5's pathway diagnostics rule out diffusion-conditioning leakage (condition-only collapses to 9.15 MAP against 64.72 for the full model). Table 7's decoupled control (pretraining on DeepCAD, reranker on MFInstSeg) still reaches 61.93/77.73 MAP, supporting reranker transferability. Table 3 holds the top-100 candidate pool fixed so reranking effects are isolated. Appendix D.1 documents deliberate anti-shortcut measures in benchmark construction: seam removal, Boolean union, and geometry validation.

Critical weaknesses
  1. The headline margins are not attributable to the claimed contributions; both comparisons are information-asymmetric. One-shot: the method mines hard examples from the full test-time corpus and uses DINOv2 filtering, while baselines update from the single exemplar only (E.3). The paper's own Table 4 shows removing only the 2D filtering drops MAP from 80.32 to 71.53 — below BRT's 72.63 (Table 1) and below the label-guided Graph Prompt ablation at 72.59. Derived from these rows, the DINOv2 component alone accounts for 8.79 points, i.e., the entire one-shot MAP advantage on LocBrep-Bench rests on a component denied to every baseline (the P@100/R@100 margins do survive: 64.89/70.63 vs 61.38/65.51). Section 5.4 presents the 9.72/15.62/7.69 gaps without connecting them to this ablation. Zero-shot: the pipeline includes a reranker supervised on MFInstSeg, while baselines rank by cosine similarity between a local query embedding and a global part embedding (E.1). Coarse-only MAP is 45.18 — below BRT (49.01) and PP-Brep (48.77) — and the supervised reranker supplies +19.54 (Table 3). Since the reranker consumes generic Eq. 7 aggregates, attaching it to baseline encoders was feasible and absent. The mitigating evidence is Coarse+Cos at 56.33 > 48.77 — a real supervision-free encoder advantage — but far from the advertised 15.71. Compounding this, no experiment establishes BrepDenoise over existing B-rep self-supervision: Table 2's only comparator is GraphMAE, which the paper itself concedes "does not provide a common reference" for component gains; the paper's own cited B-rep pretraining methods BRep-BERT and BrepMAE are never run, and there is no no-pretraining row. Each stage is, by the paper's own wording, an adaptation (SDMG, AUGCL, the PP-Brep backbone, DINOv2), yet "a novel diffusion framework" sits unreconciled with the provenance statement a page later.

  2. Both evaluation grounds are author-constructed and under-reported. LocBrep-Bench positives are author-composited; the component source corpus is never named, so within-category geometric/topological variation — the paper's stated motivation — is unquantified, and possible overlap with the MFInstSeg machining-feature taxonomy supervising the reranker is unexamined. FabWave ground truth is an author re-annotation whose inter-annotator agreement is "recorded" but never reported, and the evaluated FabWave subset size and category count are not stated. Per-category positive counts — the P@100/R@100 denominators and ceilings — appear nowhere; no standard deviation or interval is reported despite the five-seed protocol, although decisive ablation calls sit within about one point (71.53 vs 72.63/72.59). The abstract's Fusion 360 sentence ("further validate") over-reaches a disclosed internal-only, baseline-free experiment whose numbers (36.19/46.92 MAP, under silently switched P@5/R@5 metrics) appear only in appendix Table 8. All artifacts are promised only upon acceptance, so nothing is verifiable at review time.

Questions for authors
  1. Symmetric one-shot protocol: report your method restricted to exemplar-only adaptation (no corpus mining, no DINOv2), or grant BRT/PP-Brep the same transductive mining and DINOv2 filtering. Given 71.53 (Table 4) vs BRT's 72.63 (Table 1), does any one-shot MAP advantage survive under matched information access?
  2. Reranker attribution: train the same MFInstSeg-supervised cross-graph reranker on PP-Brep or UV-Net face features (it consumes Eq. 7 aggregates), and add to Table 2 BRep-BERT, BrepMAE (both in your bibliography), and a no-pretraining control on the same backbone. What fraction of the 15.71-point zero-shot margin remains attributable to BrepDenoise?
  3. Report per-category positive counts for LocBrep-Bench and the evaluated FabWave subset, per-seed standard deviations for Tables 1–4, and the recorded FabWave inter-annotator agreement value.
  4. What is the source corpus of the composed components, how many distinct component instances exist per category, and what is the category overlap between MFInstSeg annotations and the 50 LocBrep-Bench categories? Which Fusion 360 Assembly data field supplies the "parametric design timeline" used for ground truth?
  5. Section 4.2 cites no prior graph-matching literature; please position the similarity-normalized bidirectional cross-graph reranker (Eq. 9) against existing cross-graph matching architectures and state the delta, and consider complementing LocBrep-Bench with naturally occurring, non-composited local structures.

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

Related papers

Discussion (0)

Sign in to comment.