acceptodds
Under review as a conference paper at ICLR 2027

LADA: Learned Decomposition and Assembly for Non-Manifold Surface Reconstruction

Abstract

Explicit reconstruction of non-manifold surfaces remains challenging because, although most geometry is locally manifold-like, the global topology is governed by a small number of intricate non-manifold curves. Existing methods typically resolve this topology locally using marching-cubes-style algorithms, making them prone to missing, broken, or fragmented non-manifold edges. We view a non-manifold surface as a union of manifold patches meeting along non-manifold curves and propose a decomposition-and-assembly framework for explicit reconstruction from point clouds. First, a learned decomposition partitions the point cloud into overlapping blocks and predicts whether neighboring point pairs belong to the same manifold patch. Thresholding these affinities and applying union-find splits the points into a set of manifold patches as well as points on non-manifold curves. Second, the assembly step reconstructs each manifold patch with an off-the-shelf manifold surface reconstructor and stitches the patches with explicitly non-manifold edges. Because non-manifold structures are local, the network is trained only on shapes created by intersecting manifold meshes and generalizes zero-shot to unseen data. Across three benchmarks, including CAD models and point clouds sampled from neural implicit fields, our method matches strong baselines in Chamfer distance while achieving the lowest non-manifold Chamfer distance and substantially fewer topological errors on both clean and noisy inputs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Non-manifold, surface reconstruction

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_zD6dOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LADA, which is a non-manifold surface reconstruction method from point cloud by learning-based point cloud grouping. It reconstructs each patch using an existing manifold reconstruction method and assembles the patches along explicitly recovered non-manifold curves.

Critical strengths

1. Clear idea and formulation. This paper proposes a clear and well motivated decomposition-and-assembly framework. Separating manifold patch reconstruction from junction connectivity provides a coherent way to reuse established reconstruction methods while explicitly constructing the non-manifold connections between patches.

2. Improvements in topological accuracy and solid analysis.

  • Although chamfer distance drops slightly, the reported results show marked reductions in Betti number and Euler characteristic errors. On the clean benchmarks, LADA achieves the lowest Euler characteristic error across all three datasets and the lowest errors in all three Betti numbers on Constructed and Non-manifold ABC.

  • This work includes comparisons of affinity-head inputs, pairwise prediction against per-point classification and distance regression, and a sweep of the affinity threshold. The decomposition ablations support the proposed pair descriptor and pairwise formulation, while the threshold sweep shows relatively stable edge classification performance over a broad range of thresholds.

Critical weaknesses

1. Practical significance and application scope. This work addresses a rather specialized task of recovering explicit non-manifold connectivity from point clouds through a pipeline that combines several existing technical components. While the reported improvements in topological accuracy are promising, the broader practical significance of this particular problem formulation remains insufficiently demonstrated. The method seems primarily applicable to scenarios where a point cloud represents a complete object composed of multiple manifold parts, but the underlying non-manifold connectivity between these parts is unavailable. However, it is unclear how commonly such scenarios arise in practical applications, raising concerns about the relatively narrow scope and potential downstream utility of the proposed method. I believe the paper would be more convincing if the authors provided concrete application scenarios that motivate this specific reconstruction problem, along with experiments or demonstrations showing how the explicitly recovered non-manifold structures provide meaningful benefits in those applications. In particular, demonstrating a practical use case that directly leverages the recovered non-manifold connectivity would help strengthen the significance of the work.

2. Limited evidence for broader generalization and analysis of what the learned decomposition distinguishes.

  • Training uses about 1k procedural or mesh-derived shapes, and the three test sets contain less than 100 examples in total. Although the study includes CAD-derived geometry and learned-field examples, the coverage of distinct local junction configurations and more complex object geometries remains insufficiently addressed. It is therefore difficult to assess how broadly the method generalizes beyond the configurations represented in the current datasets. Additional qualitative results on a wider range of challenging objects, particularly those with junction configurations or local feature scales that differ from the training data, would strengthen the work. Showing both successful reconstructions and representative failures would also help establish its range of applicability.

  • The current experiments do not clearly show how reliably the learned affinities distinguish non-manifold junctions from sharp or highly curved manifold features. Although sharp manifold hinges are included in training, the paper does not provide a targeted evaluation of this distinction. Controlled qualitative comparisons between sharp manifold surfaces and actual non-manifold junctions, under comparable geometric and sampling conditions, would be informative. Visualizing the predicted affinities and resulting patch assignments would help clarify the model’s behavior and provide stronger evidence that it performs the intended decomposition.

Questions for authors

Please refer weakness part.

1. Practical significance. Could the authors clarify the practical application or future contribution that would benefit from the proposed decomposition-and-assembly framework? A concrete example illustrating the value of explicitly recovered patches and junction curves, beyond the reported reconstruction metrics, would help establish its broader significance.

2. Evaluation coverage and generalization. Could the authors address more diverse topological configuration and geometric complexity, particularly whether the test sets contain configurations absent from training? Additional qualitative results on challenging held-out objects, together with representative failure cases, would help clarify the method’s generalization scope.

3. Distinguishing junctions from geometric features. Given that the training data include sharp manifold hinges, could the authors provide controlled comparisons between sharp or highly curved manifold surfaces and non-manifold junctions under comparable sampling conditions? Visualizations of the predicted affinities and resulting patch decompositions would help assess whether the method reliably distinguishes these cases.

I am willing to raise my score if my concerns are resolved.

4 comments

We sincerely thank you for the careful evaluation of our manuscript and the insightful and constructive comments. We address each concern in detail below.

Practical significance and concrete application scenarios of explicit non-manifold surface reconstruction (W1, Q1)

Non-manifold surfaces are essential in at least three practical scenarios. First, the medial axis extraction: medial axes are inherently non-manifold along sheet junctions, and while methods like Q-MDF learn accurate UDFs, our method overcomes the junction connectivity bottleneck to extract explicit meshes (shown on QMDF). Second, real-world scene reconstruction involving transparent objects, where transparent and opaque surfaces usually meet along non-manifold junctions; our method extracts explicit meshes from multi-view learned distance fields, as shown on three new scenes (see Fig. 5 in the Revision PDF). Third, remeshing AIGC-generated 3D models with explicit non-manifold connectivity to optimize topology for animation and gaming.

Evaluation coverage, broader generalization, and representative failure cases (W2, Q2)

Our training and validation sets are generated from no more than 4 planar meshes and 93 Thingi10K models, augmenting them via random poses, scales, and positions into 1,800 intersecting shapes (1,200 from planar meshes and 600 from Thingi10K). Hence, the training and validation data sets only involve less than 100 independent models. We evaluate zero-shot on 31 Constructed shapes, 50 Non-manifold ABC shapes, and 14 QMDF fields. Despite using intersection operations, the styles of ABC's CAD models are clear different from Thingi10K's models. To broaden the generalization evaluation, we further evaluate three multi-view reconstructions tasks (see Fig. 5 in the Rebuttal Revision PDF). Our method again recovers a continuous sequence of non-manifold edges along each non-manifold curve, without introducing edges elsewhere. By contrast, DCUDF detects no non-manifold edges. DMUDF and SALS distribute non-manifold edges across the surfaces. For the Jug, MIND produces fragmented chains with nearby holes and DCx yields broad, irregular bands of non-manifold edges.

The Limitations section shows failure cases. When a rectangular strip connecting two parallel squares narrows from 0.05 to 0.04, its two non-manifold edge chains merge and the strip vanishes. Because points across the two curves lie too close, the network cannot distinguish the thin structure from noise, and cross-curve KNN connections render both features and predicted affinities unreliable.

Distinguishing non-manifold junctions from sharp or highly curved manifold features (W2, Q3)

As shown in Fig. 3 (rows 1–2), our test sets contain abundant sharp manifold features (e.g., sharp edges of the tree model and the creases of the CAD model). Our model reliably groups points across sharp manifold creases into the same patch rather than incorrectly splitting them at non-manifold junctions. To further illustrate this behavior, we added targeted visualizations comparing patch decompositions at sharp manifold features and non-manifold junctions (see Fig. 12 in the revision PDF).

Even if manifold sharp features were split across manifold patches, this would not impair the reconstruction and might even help recover these sharp features.

To further evaluate the generalization of our mehtod, We conducted additional experiments on 50 non-manifold shapes from the ABC dataset, selected using a different random seed. We believe these additional experiments help strengthen the experimental validation and address your concern regarding the limited number of test examples. The results are shown in the table below.

DatasetMethod#SuccessGlobal CDNon-manifold CDNon-manifold [email protected][email protected][email protected]|Δb₀| ↓|Δb₁| ↓|Δb₂| ↓|Δχ| ↓
Non-manifold ABC (Additional 50)DCUDF500.00279NANANANA3.3826.160.9027.28
DMUDF500.003110.08900.1160.0620.99912.16727.744776.304060.72
MIND500.003180.01900.8280.7770.914129.5223.82127.82236.76
DCx500.002780.00730.9040.8460.9820.909.7840.1632.48
SALS500.002840.07580.2270.1310.99712.04893.62661.92543.38
Ours500.002840.00690.9370.9380.9370.040.460.480.94

I appreciate the response and the additional experiments. I agree that the method improves what it targets (especially non-manifold edge recovery), and the new multi-view results are consistent with this. However, I do not find this sufficient to fully resolve my concerns, so I will keep my score.

My final rating is mainly based on the following points.

  • Problem setting and significance (W1). The quantitative benchmarks sample points from ground-truth meshes, including interior sheets that typical surface scanning would not observe. The applications named in the rebuttal start from neural fields rather than point clouds, and no scanned point cloud is used directly as input. They show where non-manifold surfaces appear, but not that the recovered connectivity brings a downstream benefit.
  • Generalization (W2). Non-manifold ABC, including the additional 50 shapes, uses different base models but the same intersection procedure as the training data. The out-of-distribution cases (14 QMDF fields, 3 multi-view scenes) have no ground-truth curves and rely on manual topology labels.

That said, I consider this a borderline paper. Although I am slightly leaning toward rejection, the improvement on its target is real, and my rating reflects the remaining concerns above rather than a flaw in the method itself.

We sincerely thank you for the careful evaluation and response. We would like to clarify your concerns.

  1. Problem setting and significance. We would like to clarify a potential misunderstanding regarding our target problem. The primary applications of non-manifold reconstruction are not conventional surface scanning, since internal structures are generally invisible to surface scanners. Instead, as discussed in our paper, important applications include transparent multi-view reconstruction and medial axis extraction. Indeed, the output of many medial axis methods is point cloud instead of distance field. We use point clouds as input because they provide a more general representation than UDFs. In particular, point clouds can be readily derived from a UDF by projecting sampled points onto its zero level set, whereas learning a UDF from point clouds is non-trivial. Our framework also supports UDF-derived inputs. We will clarify this distinction in the revised manuscript.

  2. Generalization. Intersections provide a simple and effective way to generate abundant non-manifold models for quantitative evaluation, following the data-generation procedure used in SALS. Although the same intersection procedure is adopted, we do not believe that the resulting models necessarily share the same distribution, since their geometric characteristics depend heavily on the underlying shapes. In particular, the ABC CAD models have substantially different geometric styles and structures from our training data, which consists of no more than 4 planar shapes and 93 Thingi10K models. Thus, these experiments evaluate generalization across different underlying geometric models, rather than merely reproducing the training examples. Moreover, the Constructed data (See Appendix) are not generated from intersections. Furthermore, medial axis extraction and transparent multi-view reconstruction demonstrate the applicability of our method to substantially different practical scenarios. Although these applications do not provide ground-truth non-manifold curves for quantitative evaluation, they offer complementary evidence of the practical value and versatility of our method beyond intersection-based benchmarks.

Reviewer_dQRYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents LADA, a non-manifold surface reconstruction method that decomposes surfaces into manifold parts and then assembles them. It does this by learning pairwise affinities (same-patch scores) between points in overlapping blocks that used to detect connected manifold components and disconnected non-manifold curves. It is trained on a synthetic dataset of non-manifold synthetically created shapes and evaluated on unseen datasets, and it mostly improves topology metrics.

Critical strengths
  • Treating the non-manifold surface reconstruction with machine learning over local pairwise features is interesting and effective.

  • This design of learning only the local pairwise affinities, while the grouping and the assembly are non-learned, seems effective and provides strong generalization. The results overall are convincing quantitatively for the non-manifold CD, Betti numbers, and Euler characteristic, and qualitatively (Figure 3).

These are consequential to my overall assessment because they show a novel and effective approach that clearly improves topology over existing methods.

Critical weaknesses
  • It is not clear why the Chamfer distance is higher compared to several competitors, and where the gap is. Furthermore, saying “It also attains a competitive CD, second only to DCx" (L309) is simply wrong when looking at the numbers.

  • The experiments with noise are nice, but they do not convince of the ability of the method to work on raw scans of non-manifold objects.

These are consequential to my overall assessment because they leave the geometric accuracy and the real-world applicability of the method unclear.

Questions for authors

Please address the weaknesses.

6 comments

We sincerely thank you for the careful evaluation of our manuscript and the insightful and constructive comments. We address each concern in detail below.

Sources of the global Chamfer Distance (CD) gap compared to competitors (W1)

The primary challenge in non-manifold reconstruction lies in recovering the correct topology rather than achieving geometric accuracy, necessitating a trade-off between topological correctness and geometric fidelity.

Our global CD gap stems from three factors: (1) We take point clouds directly as input, whereas UDF baselines rely on continuous fields trained by DEUDF on 300K points, which is much more than our about 40K input points on average; (2) Our performance drops on thin structures (see Limitations); (3) We adopt a modified iPSR for manifold patch reconstruction, which has cannot preserving sharp features, while the Constructed and non-manifold ABC date sets contain many sharp edge features. In contrast, DCx and SALS use high-resolution meshes or dual-based surface reconstruction, enabling better preservation of sharp edges. Although they achieve slightly better CD, dual-contouring-based surface extraction is known to introduce spurious non-manifold edges that compromise topological correctness. Moreover, the iPSR-based manifold reconstruction can be readily replaced with other non-watertight surface reconstruction methods.

In addition, we use point cloud as input, while DCx/DCUDF/DMUDF/MIND use UDFs as input, which are learned from much more points than the ours. To make fair comparisons with the UDF baselines, we evaluated a variant that refines non-manifold curve geometry via a UDF (Ours (refine)) for a fairer comparison. This consistently reduces the global CD while achieving the best non-manifold CD and F1 scores.

The Beti number b₀ represents the number of connected components, b₁ represents the number of independent loops (genus), and b₂ represents the number of enclosed cavities. The Euler characteristic (χ) provides an overall topological descriptor by combining the Betti numbers across different dimensions. Our method achieves substantially better topological correctness while maintaining geometric CD errors comparable to those of the baselines.

DatasetMethod#SuccessGlobal CDNon-manifold CDNon-manifold [email protected]
ConstructedDCUDF310.00292NANA
DMUDF310.003340.09650.145
MIND310.004170.05260.661
DCx310.002890.01880.907
SALS310.002950.09330.213
Ours310.003050.00710.837
Ours(refine)310.002940.00580.912
Non-manifold ABCDCUDF500.00278NANA
DMUDF500.003110.09350.109
MIND490.003060.01310.922
DCx500.002760.00700.960
SALS500.002810.07960.293
Ours500.002860.00680.921
Ours(refine)500.002830.00480.961
QMDFDCUDF140.00331--
DMUDF80.04114--
MIND130.00240--
DCx140.00257--
SALS140.00249--
Ours140.00258--
Ours(refine)140.00255--

Applicability to real-world raw captures of non-manifold objects (W2)

Thank you for the suggestion. We added evaluations on neural implicit fields learned directly from multi-view photographs of scenes with transparent objects (one scene from α-NeuS and two from NU-NeRF), where transparent surfaces intersect other geometry along non-manifold curves. We provide qualitative comparisons (Fig. 5 in the revision PDF).

Our method again recovers a continuous sequence of non-manifold edges along each non-manifold curve, without introducing edges elsewhere. By contrast, DCUDF detects no non-manifold edges. DMUDF and SALS distribute non-manifold edges across the surfaces. For the Jug, MIND produces fragmented chains with nearby holes and DCx yields broad, irregular bands of non-manifold edges.

Inaccurate statement regarding global CD ranking in Line 309 (W1)

Thank you for catching this mistake. We will revise the narrative in the main text to ensure full consistency with the quantitative tables. We delete "second only to DCx" in Sec 4.2 in the revision PDF.

Thank you for the response. My CD concern is resolved. I still have a concern about the real data because there are only 3 examples, and one is a synthetic scene. Could you report Betti and Euler errors using manual topology labels, as you did for QMDF?

We sincerely appreciate your prompt and constructive feedback.

We report Betti and Euler errors using manual topology labels on neural implicit fields learned from multi-view images of a synthetic scene (Jug) by α\alpha-NeuS and two real scenes by NU-NeRF (Ball Statue and Bottle), as we did for QMDF. Our method achieves the best metrics on neural implicit fields learned from multi-view images (with zero error against manual topology labels), significantly outperforming other baselines.

MethodJug ∣Δb0∣↓\vert\Delta b_0\vert\downarrowJug ∣Δb1∣↓\vert\Delta b_1\vert\downarrowJug ∣Δb2∣↓\vert\Delta b_2\vert\downarrowJug ∣Δχ∣↓\vert\Delta\chi\vert\downarrowBall statue ∣Δb0∣↓\vert\Delta b_0\vert\downarrowBall statue ∣Δb1∣↓\vert\Delta b_1\vert\downarrowBall statue ∣Δb2∣↓\vert\Delta b_2\vert\downarrowBall statue ∣Δχ∣↓\vert\Delta\chi\vert\downarrowBottle ∣Δb0∣↓\vert\Delta b_0\vert\downarrowBottle ∣Δb1∣↓\vert\Delta b_1\vert\downarrowBottle ∣Δb2∣↓\vert\Delta b_2\vert\downarrowBottle ∣Δχ∣↓\vert\Delta\chi\vert\downarrow
DCUDF4534425324023025
DMUDF120725602354022883888160737763769
MIND19836419428000051095141015
DCx012164152241311013141
SALS69132996080733454279143550384
Ours000000000000

The table below shows the manual topology labels of these inputs.

b0b_0b1b_1b2b_2χ\chi
Jug1243
Ball statue1012
Bottle1023

The topological metrics are highly dependent on the underlying principles and characteristics of the multi-view reconstruction method. The Betti numbers and Euler characteristic derived from our manual annotations are provided for reference only.

Reviewer_Tmf8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes LADA, a two-stage framework for explicit non-manifold surface reconstruction from point clouds. LADA formulates the problem as a decomposition-and-assembly process. First, a learned decomposition network (based on Point Transformer V3) predicts pairwise affinities within local blocks to partition the point cloud into independent manifold patches and non-manifold curve points. Second, a deterministic assembly step reconstructs each manifold patch using an off-the-shelf reconstructor (iPSR) and stitches them together along explicitly extracted non-manifold edges. LADA train purely on simple synthetic composite shapes. On three zero-shot benchmarks, the method matches strong baselines in Chamfer distance while achieving the lowest non-manifold Chamfer distance and substantially fewer topological errors on both clean and noisy inputs.

Critical strengths
  1. Problem framing is clean and well-motivated. The observation that non-manifold structure is fundamentally local, and that global topology is decided by a small one-dimensional subset, is accurate and directly motivates the block-wise learned decomposition.

  2. Decoupling non-manifold reconstruction into manifold decomposition and deterministic assembly is a highly intuitive and effective approach. It cleverly bypasses the topological ambiguities that direct meshing or UDF-extraction methods face.

  3. Training exclusively on simple, procedurally generated composite shapes and generalizing to complex CAD models and neural implicit fields is impressive. The method achieves the lowest non-manifold CD and dramatically smaller topological errors across all settings.

  4. Experimental evaluation is thorough. Three benchmarks, including a constructed set with ground-truth non-manifold curves, a CAD-derived set, and a neural-field-derived set. Noise robustness and hyperparameter sensitivity are evaluated systematically.

Critical weaknesses
  1. While the learned decomposition is fast (~1.64s), the deterministic assembly step is computationally expensive. This heavy reliance on complex geometric processing limits the method's scalability to large scenes or real-time applications.

  2. Thin-structure limitation is acknowledged but under-analyzed. It is unclear whether this is a fundamental limit of the local affinity formulation, a KNN parameter issue, or a block granularity issue. Since assembly stage is not learned, I concerned it is the main bottleneck.

Questions for authors
  1. The current method processes object-centric point clouds. How would LADA scale to large, scene-level point clouds?

  2. The assembly step is entirely deterministic and heavily dependent on the success of the learned decomposition. Is it possible to construct feedback loop to correct decomposition errors.

3 comments

We sincerely thank you for the careful evaluation of our manuscript and the insightful and constructive comments. We address each concern in detail below.

Computational efficiency of the assembly stage and scalability to large, scene-level point clouds (W1 + Q1)

We target offline, high-quality reconstruction rather than real-time applications. Decomposition takes ~1.64 s, while assembly takes ~349.5 s, which almost entirely spent on the iPSR based patch reconstruction. Since iPSR is a plug-and-play module, faster non-watertight surface reconstructors can directly replace it. Even so, end-to-end (point cloud to mesh), our pipeline is faster and uses less GPU memory than UDF baselines that require slow per-shape UDF field training.

Our method also scales well to scenes. Manifold patch decomposition is inferred independently per block, indicating that the GPU memory footprint per block is fixed and total cost grows linearly with the number of blocks. Manifold patches can be reconstructed in parallel, and non-manifold curve extraction scales with curve length rather than scene volume. Thus, our method should reconstruct scene-level point clouds as well.

Cause of the thin-structure limitation and the bottleneck (W2, Q1)

This stems from sampling resolution rather than block size or KNN settings. When thickness approaches a few times the average point spacing sˉ\bar{s}, points across two non-manifold curves become mutual KNN neighbors, creating local geometric ambiguity that degrades features and affinities (Sec. 4.6 in the revision PDF). Moreover, slight dilation in predicted curve points causes adjacent curve bands to merge, hindering decomposition. Hence, the bottleneck is in the decomposition step.

The feasibility of a feedback loop to correct decomposition errors (W2, Q2)

We thank the reviewer for this constructive suggestion. While using assembly feedback to refine decomposition is promising, existing manifold construction methods assume valid manifold inputs and cannot reliably identify violations. We therefore prioritize robust decomposition through diverse procedural training data while maintaining a modular pipeline. Incorporating assembly feedback is left for future work.

Thanks for your response. I keep my rating.

We sincerely appreciate your positive and encouraging feedback.

Reviewer_aa8pOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LADA reconstructs explicit non-manifold triangle meshes from unoriented point clouds by splitting the problem into learned decomposition and deterministic assembly. A Point Transformer V3 backbone with an MLP predicts symmetric affinities for neighbouring point pairs inside overlapping blocks, and thresholding plus union-find gives patch point sets and predicted curve points. Assembly uses iterative Poisson reconstruction per patch, falls back to local Delaunay and ball pivoting for small patches, then clips patches against an envelope, contracts it to curves with MCF skeletonization, and snaps and stitches the interfaces. Training uses 1,080 construction-labelled shapes and validation 720, with zero-shot tests on Constructed (31), Non-manifold ABC (50) and QMDF (14 learned fields). Against five baselines it reports the lowest non-manifold Chamfer distance on the two sets with curve ground truth, and Euler errors about 6.1-10.4x lower than the best baseline on all three sets. Surface-Chamfer means rank fourth of six everywhere, and non-manifold F1 trails DCx on both curve-labelled sets.

Critical strengths
  • The formulation fits the problem. Junction decisions become learnable from construction-derived labels, and assembly then stitches patches along estimated curves. On validation shapes the pairwise form reaches ARI 0.836 against 0.264 and 0.260 for per-point alternatives. Pair classification has precedent in PrimitiveNet, so the contribution is what gets decomposed and how it is supervised.
  • The topological gain is large and impressive.
Critical weaknesses
  • Generalization evidence is narrow. Both curve-labelled test sets are author-built, ABC uses mesh intersections that also appear in part of the training construction, and the Constructed recipe is only illustrated. QMDF comes from collected clouds but has no curve ground truth. The zero-shot protocol is explicitly disjoint, so this is a scope limit rather than a leak, but broader curve recovery is not shown. The values are also means from a validation-selected model, so the small ABC margin over DCx (0.0068 vs 0.0070) has no uncertainty attached.
  • The narrative needs to acknowledge the trade-off and fix a ranking. Non-manifold [email protected] is 0.907 against 0.837 on Constructed and 0.960 against 0.921 on ABC, all in DCx's favour, yet the text never discusses F1 and the conclusion still says edges lie exactly along the curves. It also calls surface Chamfer second only to DCx while the means rank fourth of six. Different orderings are not contradictory by themselves, but the wording should match the tables.
Questions for authors
  • Using the existing predictions, what explains lower non-manifold CD but lower [email protected] than DCx? Directional precision and recall plus a few overlays would separate displacement from missing segments. If actual chain errors show up, which stage is responsible?
  • Could you give paired per-shape non-manifold CD differences against DCx on ABC, or an uncertainty summary, to show whether the 0.0002 mean advantage is broad or driven by a few shapes? If the second seed is readily available, its evaluation would help.
2 comments

We sincerely thank you for the careful evaluation of our manuscript and the insightful and constructive comments. We address each concern in detail below.

Narrow generalization evidence and the use of author-built curve-labelled test sets (W1, Q1)

Since there is few public non-manifold data set, self-constructed test sets are commonly used (e.g., MIND, DCx, SALS). Our training and validation sets are generated from no more than 4 planar meshes and 93 Thingi10K models, augmenting them via random poses, scales, and positions into 1,800 intersecting shapes (1,200 from planar meshes and 600 from Thingi10K). ABC is not in the training data set. We perform zero-shot testing on Constructed (31), Non-manifold ABC (50), and QMDF (14). Despite using intersection operations, the styles of ABC's CAD models are clear different from Thingi10K's models.

We also test on multi-view implicit fields of transparent objects from α-NeuS and NU-NeRF using extracted zero-level-set point clouds (Fig. 5 in the revision PDF). Our method again recovers a continuous sequence of non-manifold edges along each non-manifold curve, without introducing edges elsewhere. By contrast, DCUDF detects no non-manifold edges. DMUDF and SALS distribute non-manifold edges across the surfaces. For the Jug, MIND produces fragmented chains with nearby holes and DCx yields broad, irregular bands of non-manifold edges.

Uncertainty of the non-manifold CD margin over DCx on Non-manifold ABC (W1, Q2)

Our non-manifold CD advantage over DCx on Non-manifold ABC is not incidental: per-shape comparisons across the 50 shapes show that our method outperforms DCx on 27 models. Moreover, since UDF-based methods inherently favor lower CD than point-cloud approaches, we also evaluate a variant with UDF-based curve refinement (Ours (refine)) for a fairer comparison, which further widens our margin.

Method
Ours0.0298680.0035860.007510.0042090.0041050.0037840.0047520.0046350.0031720.003974
DCx0.0031980.0048720.0055650.0057750.0106240.0048890.0101930.0043590.0017290.007464
Ours0.0047420.0040840.0038440.0052600.0657380.0030060.003260.0050320.0029070.004216
DCx0.0067830.0027920.0059180.0098010.012480.0066680.0035680.0040020.0024570.002767
Ours0.0033860.0032490.0050170.0105830.0041680.0034290.0069330.0037690.0041010.003879
DCx0.0022040.0094750.0028080.0068720.0059630.0084720.0048020.0032830.0069650.012767
Ours0.0069690.0049120.0072380.0034050.0029850.0044190.0042230.0053870.0032470.003645
DCx0.0011880.0053030.0042840.0020860.0117530.0037370.0040260.0060380.0128830.001263
Ours0.0061970.0040650.003040.0066760.0034380.0067190.0049680.0299220.0045230.003961
DCx0.0034720.0015260.0084910.0101770.0094110.0129560.0045980.0583390.0055520.004431
DatasetMethod#SuccessNon-manifold CDNon-manifold [email protected][email protected][email protected]
ConstructedDCUDF31NANANANA
DMUDF310.09650.1450.0810.979
MIND310.05260.6610.7310.720
DCx310.01880.9070.8710.973
SALS310.09330.2130.1270.920
Ours310.00710.8370.8600.820
Ours(refine)310.00580.9120.9290.900
Non-manifold ABCDCUDF50NANANANA
DMUDF500.09350.1090.0580.998
MIND490.01310.9220.8940.968
DCx500.00700.9600.9300.995
SALS500.07960.2930.1760.996
Ours500.00680.9210.9310.917
Ours(refine)500.00480.9610.9620.966

Trade-off and discrepancy between lower non-manifold CD and lower [email protected] compared to DCx (W2 + Q1)

We report the precision and recall at 0.01 (see above table) to clarify this trade-off.

Extracting non-manifold curves by contracting an isosurface envelope via MCF skeletonization inherently introduces slight geometric displacement rather than missing segments, causing the original [email protected] to trail DCx. We use point cloud as input, while DCx uses UDFs as input, which are learned from much more points than the our input points. To make fair comparisons with the UDF baselines, we further refine our mesh around the non-manifold regions by the UDF. The result is shown in the last row (ours(refine)). It achieves the best performance in both non-manifold CD and [email protected]. Conversely, DCx generates wide, spurious non-manifold bands (inflating recall over precision), broken chains, and bubble artifacts, which are reflected from the topological indicators.

Narrative and ranking revisions to match the quantitative tables (W2)

We will revise the narrative in the main text to ensure full consistency with the quantitative tables.

  • We delete "second only to DCx" in Sec 4.2
  • We discuss F1 in Sec 4.2
  • We change "exactly" to "nearly" in Conclusion

Thank you for suggesting that we evaluate the robustness of our non-manifold CD results to different random seeds used for model selection. Following your suggestion, we repeated the evaluation on 50 different non-manifold ABC shapes selected using a different random seed. As shown below, our method consistently achieves the lowest mean non-manifold CD among all compared methods (0.0069, compared with 0.0073 for DCx), further supporting the robustness of our results. This time, our [email protected] also achieves the best without further UDF optimization. The results are shown in the table below.

DatasetMethod#SuccessGlobal CDNon-manifold CDNon-manifold [email protected][email protected][email protected]|Δb₀| ↓|Δb₁| ↓|Δb₂| ↓|Δχ| ↓
Non-manifold ABC (Additional 50)DCUDF500.00279NANANANA3.3826.160.9027.28
DMUDF500.003110.08900.1160.0620.99912.16727.744776.304060.72
MIND500.003180.01900.8280.7770.914129.5223.82127.82236.76
DCx500.002780.00730.9040.8460.9820.909.7840.1632.48
SALS500.002840.07580.2270.1310.99712.04893.62661.92543.38
Ours500.002840.00690.9370.9380.9370.040.460.480.94

Comments on the paper

2 comments
AuthorsOpenReview ↗

Common Concerns Across Reviewers

We sincerely thank all reviewers for their insightful and constructive comments, and we summarize our responses to common questions below.

Zero-shot generalization, dataset composition, and real-world multi-view evaluations (Reviewer aa8p-W1/Q1, Reviewer dQRY-W2, Reviewer zD6d-W1/W2a/Q1/Q2)

Since there are few public non-manifold datasets, self-constructed test sets are commonly used (e.g., MIND, DCx, SALS). Our training and validation sets are generated from no more than 4 planar meshes and 93 Thingi10K models (<100 independent models in total), augmenting them via random poses, scales, and positions into 1,800 intersecting shapes (1,200 planar and 600 Thingi10K). We perform zero-shot testing on Constructed (31), Non-manifold ABC (50), and QMDF (14), none of which appear in training. Despite using intersection operations, the styles of ABC's CAD models are clearly different from Thingi10K's models.

To broaden the generalization evaluation and demonstrate applicability to real-world captures, we also test on neural implicit fields learned directly from multi-view photographs of scenes with transparent objects (one scene from α-NeuS and two from NU-NeRF) using extracted zero-level-set point clouds, where transparent surfaces intersect other geometry along non-manifold curves (Fig. 5 in the revision PDF). Our method again recovers a continuous sequence of non-manifold edges along each non-manifold curve. By contrast, DCUDF detects no non-manifold edges, DMUDF and SALS distribute non-manifold edges across the surfaces, MIND produces fragmented chains with nearby holes on the Jug, and DCx yields broad, irregular bands of non-manifold edges.

Fair comparisons with UDF baselines via UDF-based refinement (Reviewer aa8p-W1/W2/Q1/Q2, Reviewer dQRY-W1a)

We use a point cloud as input (~40K points on average), while DCx/DCUDF/DMUDF/MIND use continuous UDFs learned by DEUDF from 300K points, giving UDF-based methods an inherent geometric advantage in achieving lower CD. Moreover, extracting non-manifold curves by contracting an isosurface envelope via MCF skeletonization inherently introduces slight geometric displacement rather than missing segments, causing our original [email protected] to trail DCx. To make fair comparisons with UDF baselines, we evaluated a variant that further refines our mesh around non-manifold regions via a UDF (Ours (refine)). As shown below, this consistently reduces global CD while achieving the best performance in both non-manifold CD and [email protected] across benchmarks. Conversely, DCx generates wide, spurious non-manifold bands (inflating recall over precision), broken chains, and bubble artifacts, as reflected in the topological indicators.

DatasetMethod#SuccessGlobal CDNon-manifold CDNon-manifold [email protected][email protected][email protected]
ConstructedDCUDF310.00292NANANANA
DMUDF310.003340.09650.1450.0810.979
MIND310.004170.05260.6610.7310.720
DCx310.002890.01880.9070.8710.973
SALS310.002950.09330.2130.1270.920
Ours310.003050.00710.8370.8600.820
Ours(refine)310.002940.00580.9120.9290.900
Non-manifold ABCDCUDF500.00278NANANANA
DMUDF500.003110.09350.1090.0580.998
MIND490.003060.01310.9220.8940.968
DCx500.002760.00700.9600.9300.995
SALS500.002810.07960.2930.1760.996
Ours500.002860.00680.9210.9310.917
Ours(refine)500.002830.00480.9610.9620.966
QMDFDCUDF140.00331----
DMUDF80.04114----
MIND130.00240----
DCx140.00257----
SALS140.00249----
Ours140.00258----
Ours(refine)140.00255----

Cause of the thin-structure limitation, bottleneck, and representative failure cases (Reviewer Tmf8-W2/Q1, Reviewer dQRY-W1a, Reviewer zD6d-W2a/Q2)

This stems from sampling resolution rather than block size or KNN settings. As shown in Limitations, when a rectangular strip connecting two parallel squares narrows from 0.05 to 0.04 (approaching a few times the average point spacing sˉ\bar{s}), its two non-manifold edge chains merge and the strip vanishes. Because points across the two curves lie too close, the network cannot distinguish the thin structure from noise, and cross-curve mutual KNN connections create local geometric ambiguity that degrades both features and predicted affinities (Sec. 4.6 in the revision PDF). Moreover, slight dilation in predicted curve points causes adjacent curve bands to merge, hindering decomposition. Hence, the bottleneck is in the decomposition step.

Narrative and ranking revisions to match the quantitative tables (Reviewer aa8p-W2, Reviewer dQRY-W1b)

Thank you for catching these mistakes. To ensure full consistency with the tables in the revision PDF, we deleted "second only to DCx" and added a discussion of F1 in Sec. 4.2, and changed "exactly" to "nearly" in the Conclusion.

AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and All Reviewers,

We sincerely appreciate the time and effort all reviewers have devoted to evaluating our paper, as well as your insightful comments and constructive suggestions. We summarize the main updates below:

  • Clarified the composition of our training data and the zero-shot evaluation settings on Constructed, Non-manifold ABC, and QMDF.
  • Evaluated our method on 50 additional Non-manifold ABC shapes selected with a different random seed, achieving the lowest non-manifold CD and highest [email protected] among the compared methods.
  • Added qualitative comparisons on three multi-view reconstructions (Fig. 5 in the revised manuscript) and reported Betti numbers and Euler characteristic errors (ours is 0), against manual annotations.
  • Reported additional results for UDF-based refinement, including global CD, non-manifold CD, [email protected], precision, and recall, to provide a more comprehensive comparison.
  • Expanded the discussion of thin-structure failures and added visualizations comparing patch decompositions around sharp manifold features and non-manifold junctions (Sec. 4.6 and Fig. 12 in the revised manuscript).
  • Clarified the computational cost of the assembly stage, the role of iPSR, and the scalability of block-wise decomposition.
  • Revised the discussion of CD and F1 in Sec. 4.2 and replaced “exactly” with “nearly” in the Conclusion to ensure more precise wording.
  • Clarified the relationship between point cloud and UDF inputs and practical applications.

Thank you again for your valuable time and consideration.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.