acceptodds
Under review as a conference paper at ICLR 2027

VoroFacet: Junction-aware Mesh Extraction from Neural Unsigned Distance Fields

Abstract

Unsigned distance fields (UDFs) can represent surfaces of arbitrary topology, but recovering mesh connectivity from neural UDFs remains challenging, especially at non-manifold junctions where multiple sheets meet. To this end, we propose VoroFacet, a novel pipeline that extracts single-layer meshes from neural UDFs and reconstructs junction connectivity accurately through Voronoi facets. Specifically, we first use a near-surface double-layered proxy derived from the UDF as a local spatial partition for facet selection, where the facets belong to the Voronoi diagram of the sites generated by an adaptive ball cover of the proxy. We then estimate junctions from the mesh of this first extraction and place balls along them, so that in a re-extraction process, the resulting cospherical sites make incident sheets share Voronoi vertices and edges at these junctions. Finally, we propose a three-stage optimization to refine the geometry of the extracted mesh. Experiments demonstrate superior reconstruction quality over state-of-the-art methods, particularly at non-manifold junctions. The source code will be publicly available.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
surface reconstruction, implicit representation

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

This paper introduces a mesh extraction approach from (neural) UDFs. While SDFs can represent only closed & manifold surfaces, UDFs can represent open & non-manifold surfaces. Due to this property, the mesh extraction algorithms for UDFs should handle the non-manifoldness. The most prominent problem is the non-manifold edges, where multiple smooth surfaces meet with each other, and create junctions. This work delves into this problem, and provides a solution that extracts topologically cleaner meshes from UDFs than the other baseline methods.

The work seems to be inspired by the two baseline methods: 1) DCUDF2, and 2) VoroCrust. This work starts its extraction process from the double layer proxy of the surface from DCUDF2. This baseline work produces meshes of good geometric quality, but it removes the non-manifold edges by creating false structures near the junctions. To amend this issue, this work leverages the similar apporach as VoroCrust, to sample ball covers on the double layer proxy. With these ball covers and hand-crafted hyperparameters, this approach provides a theoretical guarantee about the surface approximation quality. Also, to produce geometrically more accurate mesh, this approach designs multi-stage approach, to get coarse approximation of the underlying surface in the first step and fine-tune the mesh in subsequent steps to get more accurate mesh.

To validate their approach, authors provide the comparisons with the other baseline methods in diverse scenarios: 1) objects with non-manifoldness, 2) objects with complex inner structure (car), and 3) manifold objects (other scenarios are in Appendix as well). They also provide ablations about the hyperparameters they selected in their algorithm.

Critical strengths
  • Robust mesh extraction for non-manifold objects: As UDFs can represent non-manifold shapes, the mesh extraction algorithms for UDFs should faithfully reconstruct the non-manifoldness in the output. In terms of this, this algorithm shows more robust outputs compared to the other approaches.

  • Theoretically sound approach: This algorithm is theoretically grounded on the Voronoi diagram and ball coverings, so geometrical & topological properties are guaranteed.

Critical weaknesses

While authors claim that their method can handle all kinds of geometry under single pipeline (line 428-431), I don't think this is correct statement. I think this algorithm is particularly useful for handling the non-manifold cases, while the baseline works (DCUDF, DCUDF2) focus on the closed manifold cases. As shown in the Table 1 and 2, these baseline works generate manifold surfaces, which are usually more desirable in many downstream applications. Therefore, I'm not convinced that this work is better than these baseline approaches for every geometrical structures. Also, as this method heavily relies on DCUDF2 to first extract the double layer structure (which also accounts for almost half of the computation time), I feel this method is kind of incremental work.

One follow-up question in terms of this -- can you also report the number of non-manifold cases in Table 3 (when the GT geometry is manifold)? If VoroFacet guarantees to generate manifold mesh for these cases I think the claim would become stronger. At least, as this method starts from DCUDF2's double layer structure from the beginning, I think the authors should show that it should not harm DCUDF2's manifoldness property when the GT shape is manifold as well.

Questions for authors

No particular questions that affect my assessment.

5 comments
AuthorsOpenReview ↗

Responses to Reviewer NGep (1/2)

Thank you for your time and thoughtful review. We appreciate your recognition of VoroFacet’s robust non-manifold reconstruction and its foundation in Voronoi geometry. We address your concerns below.

Comment 1: While authors claim that their method can handle all kinds of geometry under single pipeline (line 428-431), I don't think this is correct statement. I think this algorithm is particularly useful for handling the non-manifold cases, while the baseline works (DCUDF, DCUDF2) focus on the closed manifold cases. As shown in the Table 1 and 2, these baseline works generate manifold surfaces, which are usually more desirable in many downstream applications. Therefore, I'm not convinced that this work is better than these baseline approaches for every geometrical structures. ... One follow-up question in terms of this -- can you also report the number of non-manifold cases in Table 3 (when the GT geometry is manifold)? If VoroFacet guarantees to generate manifold mesh for these cases I think the claim would become stronger. At least, as this method starts from DCUDF2's double layer structure from the beginning, I think the authors should show that it should not harm DCUDF2's manifoldness property when the GT shape is manifold as well.

Response:

Thank you for raising this concern. For non-manifold targets, DCUDF and DCUDF2 retain a manifold double cover rather than a single-layer surface with shared junction connectivity. Although these overlapping layers can render accurately, they contain extensive self-intersections: on Non-manifold Objects, 68.352% and 43.947% of triangles are affected on average, respectively (Appendix B, Table 8). Such outputs cannot directly serve as the single physical sheet required by applications such as thin-shell simulation. Manifold connectivity alone therefore does not establish suitability for these downstream uses.

DCUDF’s single-layer recovery requires user-specified surface types to select the appropriate post-processing, including min-cut separation for orientable surfaces with boundary. In our manifold-object experiments, we enable min-cut for both DCUDF and DCUDF2, yet two outputs from each method still retain double layers, substantially increasing their mean self-intersection percentages. VoroFacet instead reconstructs single-layer surfaces through Voronoi facet selection, without user-specified surface types or min-cut layer separation.

We have added the requested statistics for all methods on the 50 Manifold Objects in Appendix B, Table 9, with the relevant comparison below. Triangle percentages are reported as mean ±\pm standard deviation across all 50 objects; NM-incident triangles are those incident to non-manifold edges.

MethodModels with NM edgesNM-incident triangles (%)Models with self-intersectionsSelf-intersecting triangles (%)
DCUDF0/500.000±0.0000.000 \pm 0.00037/502.233±10.7102.233 \pm 10.710
DCUDF20/500.000±0.0000.000 \pm 0.00040/501.880±8.6981.880 \pm 8.698
VoroFacet46/501.321±2.1421.321 \pm 2.14243/500.070±0.1030.070 \pm 0.103

VoroFacet does not guarantee strict manifoldness: 46 outputs contain non-manifold edges. However, their incident triangles account for 1.321% of all triangles on average, while self-intersections involve only 0.070%. These reconstructions also achieve the lowest mean CD on Manifold Objects, 4.515 versus 4.529 for DCUDF and 4.528 for DCUDF2 (Table 2). Our unified-pipeline claim refers to this ability to reconstruct single-layer surfaces across manifold, open-boundary, and non-manifold inputs without switching extraction procedures according to surface type.

AuthorsOpenReview ↗

Responses to Reviewer NGep (2/2)

Comment 2: Also, as this method heavily relies on DCUDF2 to first extract the double layer structure (which also accounts for almost half of the computation time), I feel this method is kind of incremental work.

Response: Thank you for raising this concern. VoroFacet introduces a connectivity construction distinct from DCUDF2’s deformation-based extraction. DCUDF2 supplies a near-surface geometric proxy; VoroFacet neither modifies its deformation algorithm nor inherits its mesh connectivity. Instead, ball-induced Voronoi facet selection and junction-aware re-extraction construct a single-layer surface with shared vertices and edges at junctions. The proxy guides local facet selection, for which a double-cover topology is not itself a requirement.

We have also added a controlled experiment replacing DCUDF2 with DCUDF while keeping the learned UDFs and all downstream parameters unchanged (Appendix B, Table 6). Surface accuracy remains comparable, and NM-F1 reaches 62.15 with the DCUDF proxy, exceeding the strongest baseline, DCx (53.19). These results demonstrate that VoroFacet does not require DCUDF2 specifically. More broadly, separating geometric approximation from connectivity construction allows proxy generation to focus on geometric fidelity while VoroFacet reconstructs the shared surface connectivity.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer NGep,

Thank you again for your thoughtful review. In response to your concerns, we have:

  1. Clarified our unified single-layer extraction pipeline, which requires neither user-specified surface types nor min-cut layer separation. We also report the requested numbers of outputs with non-manifold edges and compare self-intersections across methods on manifold inputs (Table 9).
  2. Clarified the novelty beyond DCUDF2, explaining how Voronoi facet selection and junction-aware re-extraction construct new connectivity rather than inherit the proxy’s. The DCUDF-proxy experiment further shows that VoroFacet does not require DCUDF2 specifically (Table 6).

As the discussion period is coming to a close, we would be grateful to know whether these address your concerns, or if there is anything else we can clarify.

Best regards, The Authors

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

This paper presents VoroFacet, a pipeline for extracting single-layer meshes from neural unsigned distance fields (UDFs), with emphasis on correctly reconstructing non-manifold junctions where multiple sheets meet. The method uses the near-surface double-layered proxy from DCUDF2 as a geometric prior, constructs an adaptive ball cover of the proxy, generates Voronoi sites from exposed triple-sphere intersections, and selects Voronoi facets based on local geometric criteria to build the output connectivity. For junctions, the paper designs a two-pass extraction strategy: the first pass extracts an initial mesh and estimates junction curves, and the second pass places junction balls along them so that incident sheets share Voronoi vertices and edges. Finally, geometry-adaptive remeshing and a three-stage optimization refine the geometry. Experiments on four benchmarks (Non-manifold Objects, ShapeNet-Car, 3D Scenes, Open-boundary Objects) demonstrate state-of-the-art junction reconstruction quality.

Notably, the paper adopts a geometry-processing-oriented strategy rather than a learning-based approach, achieving strong results on junction reconstruction through careful algorithmic design (Voronoi facet selection, adaptive ball cover, junction-aware re-extraction) rather than neural network innovations.

Critical strengths

S1. Precise and important problem formulation. The core difficulty of mesh extraction from UDFs lies in determining connectivity, especially at non-manifold junctions. The paper clearly decomposes the problem into "geometry extraction" and "connectivity construction," and identifies the systematic failure modes of existing methods at junctions (holes, spurious connections, gaps). The motivation is well-grounded.

S2. Novel Voronoi facet selection formulation. Casting mesh extraction as facet selection in a Voronoi diagram is a conceptual contribution. Unlike pseudo-sign or cell adjacency methods, Voronoi partitioning is inherently gap-free and overlap-free; selected facets meet along shared Voronoi edges, providing topological guarantees for connectivity.

S3. Clever junction-aware re-extraction design. The two-pass strategy is logically clean: the first pass observes geometry and estimates junction positions; the second pass enforces shared connectivity through junction balls. The 2D illustration in Figure 4 clearly demonstrates this process. The three-stage optimization (manifold fitting, non-manifold vertex fitting, manifold refinement) decouples the refinement to avoid pulling non-manifold vertices toward local centroids in joint optimization.

Critical weaknesses

W1. Strong dependence on DCUDF2 proxy without independent analysis. VoroFacet's entire pipeline builds on the double-layered proxy generated by DCUDF2. Although the paper claims the proxy serves only as a "geometric prior," the proxy quality directly determines the input to ball cover and all subsequent steps. The paper does not analyze: (1) how VoroFacet performs when the DCUDF2 proxy is poor (large gaps, artifacts); (2) the impact of using proxies of different quality (e.g., DCUDF vs DCUDF2) on final results; (3) the relationship between the proxy's double-layer gap width and ball radius parameters. If the proxy gap is non-uniform, can the adaptive ball cover always handle it correctly?

W2. Low computational efficiency. VoroFacet takes 428 seconds per model on average (Table 1), approximately 4.7x slower than DCUDF2 (91s) and 5.7x slower than DCx (75.6s). Proxy construction accounts for 41.2% and the two extraction passes for 33.9% of runtime. For batch processing or real-time interactive applications, this overhead is a significant bottleneck. The paper does not discuss acceleration strategies or parallelization possibilities.

W3. Many hyperparameters; generalization across datasets is uncertain. The method involves 7+ key hyperparameters (R_init, R0_gap, R1_gap, R_G, γ, δ_G, plus multiple weights for the three-stage optimization). Although Appendix D shows parameter sensitivity, all experiments use the same default parameters. The paper does not demonstrate generalization on different data types (e.g., clothing, architecture, mechanical parts). For small changes in gap threshold (Tables 8-9), NM-F1 varies by about 3-4 points, which could be more sensitive under more extreme data distributions.

W4. Limited evaluation dataset scale. Non-manifold Objects has only 20 models, ShapeNet-Car only 20 models, and 3D Scenes only 5 scenes. Although these datasets cover various topological complexities, the small sample sizes raise questions about statistical significance. No standard deviations or confidence intervals are reported. Additionally, no out-of-distribution testing is performed.

W5. Reliability of ground-truth junction extraction. The CAD mesh junction extraction algorithm in Appendix A itself depends on multiple heuristic thresholds (distance tolerance ε, 3° angle threshold). These threshold choices affect ground-truth quality, which in turn affects the reliability of NM-CD and NM-F1 metrics. The paper does not validate these ground-truth junctions against manual annotations, nor report how metrics change under different threshold settings.

Questions for authors

Q1. How robust is VoroFacet to proxy quality? If using a proxy generated by DCUDF (rather than DCUDF2), or if the proxy has large gaps or artifacts, how much does junction reconstruction quality degrade? Is there a quantitative degradation analysis?

Q2. In purely manifold scenarios (no junctions), why does VoroFacet's surface accuracy decrease compared to DCUDF2 (Table 1: CD 5.609 vs 5.320)? Can this cost be avoided by skipping re-extraction in manifold regions?

Q3. The main computational bottlenecks are proxy construction (41.2%) and ball cover (33.9%). Are there plans to introduce GPU parallel acceleration (e.g., parallel Voronoi computation) or approximate algorithms to reduce overhead?

Q4. For very thin sheets (e.g., paper, cloth), when the double-layer gap approaches or falls below R0_gap, can the adaptive ball cover still correctly distinguish the two sides of the same sheet? Is the opposing-triangle test in Figure 3 reliable in this case?

Q5. NM-LR deviates from 1 for all methods (Table 1: VoroFacet 1.486, DCx 3.575). Does this suggest a systematic bias in the current junction evaluation protocol? How much does the ground-truth junction extraction threshold choice affect NM-LR?

8 comments
AuthorsOpenReview ↗

Responses to Reviewer CjdD (1/3)

Thank you for your detailed and constructive review. We appreciate your recognition of our problem formulation, the novelty of our Voronoi facet-selection approach, and the design of our junction-aware re-extraction strategy. We address your concerns below.

Comment 1: Strong dependence on DCUDF2 proxy without independent analysis. ... The paper does not analyze: (1) how VoroFacet performs when the DCUDF2 proxy is poor (large gaps, artifacts); (2) the impact of using proxies of different quality (e.g., DCUDF vs DCUDF2) on final results; (3) the relationship between the proxy's double-layer gap width and ball radius parameters. If the proxy gap is non-uniform, can the adaptive ball cover always handle it correctly? ... How robust is VoroFacet to proxy quality? If using a proxy generated by DCUDF (rather than DCUDF2), or if the proxy has large gaps or artifacts, how much does junction reconstruction quality degrade? Is there a quantitative degradation analysis?

Response: Thank you for requesting this analysis. We have added a controlled experiment replacing DCUDF2 with DCUDF on the 50 Non-manifold Objects, keeping the learned UDFs and all downstream parameters unchanged. VoroFacet maintains comparable surface accuracy and strong junction reconstruction with either proxy, demonstrating that it does not require DCUDF2 specifically. Quantitative results and qualitative comparisons are provided in Appendix B, Table 6 and Figure 11 of the revised manuscript.

Results from Table 6 are shown below as mean ±\pm standard deviation. CD is scaled by 10310^3; F1 scores are percentages. Bold and underline indicate the best and second-best means, respectively.

ProxyCD ↓F1 (%) ↑NM-F1 (%) ↑
DCUDF5.563‾±0.826\underline{5.563} \pm 0.82680.32‾±8.94\underline{80.32} \pm 8.9462.15‾±23.16\underline{62.15} \pm 23.16
DCUDF25.549±0.831\mathbf{5.549} \pm 0.83180.46±8.98\mathbf{80.46} \pm 8.9865.80±21.37\mathbf{65.80} \pm 21.37

With the DCUDF proxy, VoroFacet achieves 62.15 NM-F1, exceeding the strongest baseline, DCx (53.19). Figure 11 also illustrates successful junction recovery using either proxy. These results show that the method remains effective with an alternative proxy constructor, although junction accuracy is affected by proxy quality.

Regarding whether the adaptive cover can always handle non-uniform gaps, the key is whether the proxy clearly distinguishes duplicate-layer gaps from genuine gaps between distinct target sheets. The former should be covered and the latter preserved. When their widths become comparable, the ball-radius constraints may merge distinct sheets, as shown in the lower example of Figure 10. All compared baselines also fail to faithfully reconstruct the two closely spaced sheets in this example, highlighting a challenge shared across the evaluated methods.

Comment 2: Low computational efficiency. VoroFacet takes 428 seconds per model on average (Table 1), approximately 4.7x slower than DCUDF2 (91s) and 5.7x slower than DCx (75.6s). Proxy construction accounts for 41.2% and the two extraction passes for 33.9% of runtime. For batch processing or real-time interactive applications, this overhead is a significant bottleneck. The paper does not discuss acceleration strategies or parallelization possibilities. ... The main computational bottlenecks are proxy construction (41.2%) and ball cover (33.9%). Are there plans to introduce GPU parallel acceleration (e.g., parallel Voronoi computation) or approximate algorithms to reduce overhead?

Response: Thank you for highlighting the missing discussion of acceleration and parallelization. For clarity, Table 1 reports 418.0 s for VoroFacet and 153.0 s for DCUDF2, approximately 2.7×. VoroFacet targets high-quality single-layer meshes with shared junction connectivity; the additional computation reconstructs this geometry and connectivity from the proxy.

Our implementation already uses GPU acceleration for proxy optimization and three-stage refinement, and CPU multithreading for ball-radius constraints, independent site construction, facet tests, and remeshing across sheets. Greedy ball selection and dependent connectivity updates within each sheet’s remeshing remain sequential in the current implementation. We have expanded Appendix F to explain this parallelization and discuss two directions for reducing the dominant costs: faster, geometrically accurate proxy construction tailored to VoroFacet, and incremental second-pass extraction that reuses unaffected parts of the first-pass ball cover and Voronoi structure.

AuthorsOpenReview ↗

Responses to Reviewer CjdD (2/3)

Comment 3: Many hyperparameters; generalization across datasets is uncertain. The method involves 7+ key hyperparameters (R_init, R0_gap, R1_gap, R_G, γ, δ_G, plus multiple weights for the three-stage optimization). Although Appendix D shows parameter sensitivity, all experiments use the same default parameters. The paper does not demonstrate generalization on different data types (e.g., clothing, architecture, mechanical parts). For small changes in gap threshold (Tables 8-9), NM-F1 varies by about 3-4 points, which could be more sensitive under more extreme data distributions. ... Additionally, no out-of-distribution testing is performed.

Response:

Thank you for raising this concern. VoroFacet itself is not trained across object categories; it performs geometric extraction and optimization on each input neural UDF. Our existing evaluation covers non-manifold CAD shapes, ShapeNet cars, real scanned scenes, and MGN garments, using the same geometric parameters and refinement weights across these datasets (Tables 1, 3, 10, and 11).

In particular, the ShapeNet-Car evaluation retains all interior faces, including thin, closely spaced sheets and numerous non-manifold junctions. On these challenging geometries, VoroFacet achieves the best junction accuracy, with NM-F1 of 36.67 versus 30.80 for the strongest baseline. These experiments assess VoroFacet’s applicability across different data types. We have made this existing coverage explicit in Section 5.2 of the revised manuscript.

Comment 4: Limited evaluation dataset scale. Non-manifold Objects has only 20 models, ShapeNet-Car only 20 models, and 3D Scenes only 5 scenes. Although these datasets cover various topological complexities, the small sample sizes raise questions about statistical significance. No standard deviations or confidence intervals are reported.

Response:

Thank you for raising this concern. We would like to note that the main evaluations on Non-manifold Objects and Manifold Objects each contain 50 models (Tables 1 and 2); the 20-model subset is used for component ablations (Table 4). The 20 ShapeNet-Car models and five 3D Scenes complement these main evaluations with complex interior structures and real scanned scenes. Their sample sizes match those used by our baselines HR-UDF (20 cars) and CAP-UDF (five scenes), respectively. Following your suggestion, we now report mean ±\pm standard deviation across objects or scenes in Tables 1–6 and 8–16 of the revised manuscript, making the observed variability explicit.

AuthorsOpenReview ↗

Responses to Reviewer CjdD (3/3)

Comment 5: Reliability of ground-truth junction extraction. The CAD mesh junction extraction algorithm in Appendix A itself depends on multiple heuristic thresholds (distance tolerance ε, 3° angle threshold). These threshold choices affect ground-truth quality, which in turn affects the reliability of NM-CD and NM-F1 metrics. The paper does not validate these ground-truth junctions against manual annotations, nor report how metrics change under different threshold settings. ... NM-LR deviates from 1 for all methods (Table 1: VoroFacet 1.486, DCx 3.575). Does this suggest a systematic bias in the current junction evaluation protocol? How much does the ground-truth junction extraction threshold choice affect NM-LR?

Response:

Thank you for raising this concern. The source models underlying our datasets come from diverse contributors on public online platforms, without junction annotations. Their sheets may intersect or nearly touch without sharing mesh edges, so distance and angular tolerances are needed to identify junctions from geometry. These thresholds affect the reference junction length and hence NM-LR, even when the predictions remain unchanged; deviation from 1 alone does not establish systematic evaluation bias.

We have tested seven GT extraction settings while keeping all reconstructed meshes fixed, and added the results to Appendix A, Table 5 of the revised manuscript. Below we summarize mean NM-LR (ideal value 1) for VoroFacet and DCx, the strongest baseline in NM-F1 on both datasets. Bold/underline indicate the best/second-best values within each two-method comparison.

GT extraction settingNon-manifold Objects: DCxNon-manifold Objects: VoroFacetShapeNet-Car: DCxShapeNet-Car: VoroFacet
Default6.629‾\underline{6.629}2.167\mathbf{2.167}1.696‾\underline{1.696}0.982\mathbf{0.982}
Distance tolerance ε0/2\varepsilon_0/28.618‾\underline{8.618}2.615\mathbf{2.615}1.764‾\underline{1.764}1.022\mathbf{1.022}
Distance tolerance 2ε02\varepsilon_05.448‾\underline{5.448}1.885\mathbf{1.885}1.527‾\underline{1.527}0.886\mathbf{0.886}
Direction merging 1∘1^\circ5.266‾\underline{5.266}1.681\mathbf{1.681}1.660‾\underline{1.660}0.959\mathbf{0.959}
Direction merging 5∘5^\circ6.689‾\underline{6.689}2.196\mathbf{2.196}1.715‾\underline{1.715}0.993\mathbf{0.993}
Axis–plane alignment 3∘3^\circ6.631‾\underline{6.631}2.168\mathbf{2.168}1.699‾\underline{1.699}0.983\mathbf{0.983}
Axis–plane alignment 7∘7^\circ6.627‾\underline{6.627}2.167\mathbf{2.167}1.694‾\underline{1.694}0.981\mathbf{0.981}

Although NM-LR varies with the thresholds, VoroFacet remains closer to 1 than DCx in every setting. More importantly, it achieves the lowest mean NM-CD and highest mean NM-F1 and VA-NM-F1 under all seven settings on both datasets, as detailed in Table 5 of the revised manuscript. This shows that its junction reconstruction advantage is not specific to the default reference thresholds.

Comment 6: In purely manifold scenarios (no junctions), why does VoroFacet's surface accuracy decrease compared to DCUDF2 (Table 1: CD 5.609 vs 5.320)? Can this cost be avoided by skipping re-extraction in manifold regions?

Response: Thank you for raising this question. The purely manifold results are reported in Table 3 of the original submission, now Table 2 in the revised manuscript, rather than Table 1. In this evaluation, VoroFacet achieves the lowest mean CD among all compared methods, with 4.515 versus 4.528 for DCUDF2, and a higher F1 of 90.33 versus 90.15. These full-pipeline results do not show the suggested CD degradation relative to DCUDF2.

Comment 7: For very thin sheets (e.g., paper, cloth), when the double-layer gap approaches or falls below R0_gap, can the adaptive ball cover still correctly distinguish the two sides of the same sheet? Is the opposing-triangle test in Figure 3 reliable in this case?

Response: Thank you for this question. For duplicate layers representing the same target sheet, a small gap is deliberately covered rather than preserved. In Eq. (3), triangle pairs with centroid distance dij≤2Rgap0d_{ij}\leq 2R_{\mathrm{gap}}^0 are excluded from the opposing-pair constraints.

If you are referring instead to two distinct target sheets whose separation is too small to preserve in the first pass but becomes distinguishable at the smaller second-pass threshold, then yes: pairs satisfying 2Rgap1<dij≤2Rgap02R_{\mathrm{gap}}^1<d_{ij}\leq 2R_{\mathrm{gap}}^0 can be classified as opposing in the second pass when the normal-alignment condition also holds. We have added Figure 5 in Section 4.2 of the revised manuscript to demonstrate this case: the first extraction merges the two sheets, while the second restores their separation using the smaller Rgap1R_{\mathrm{gap}}^1.

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

This paper proposes VoroFacet, a method for extracting single-layer meshes from neural unsigned distance fields, with a focus on nonmanifold junctions. It uses a double-layered surface proxy to guide Voronoi facet selection, followed by junction-aware re-extraction, remeshing, and geometric refinement. Experiments demonstrate improved junction reconstruction while maintaining competitive overall surface accuracy across different surface types and complex geometries.

Critical strengths
  1. The method constructs mesh connectivity using selected Voronoi facets and junction-aware re-extraction to connect incident sheets, addressing a key limitation of existing neural UDF extraction methods.
  2. VoroFacet demonstrates notable improvements in nonmanifold junction reconstruction over the evaluated baselines while maintaining competitive overall surface accuracy.
Critical weaknesses
  1. The junction metrics primarily measure geometric proximity and junction length rather than directly assessing correct connectivity. Although the authors explain why removing junction-aware re-extraction improves NM-F1, the evaluation still lacks direct evidence that the reconstructed junctions preserve the correct connectivity between incident surfaces.
  2. Although the paper evaluates complex geometries, the robustness of the method under challenging conditions remains insufficiently characterized. In particular, the evaluation does not systematically assess closely spaced surfaces, complex junction configurations, or self-intersections introduced during refinement. Together with the substantial computational overhead, this makes it difficult to assess the method's practical reliability.
Questions for authors

Q1. Could the authors provide additional qualitative results, possibly in a revised supplementary material, particularly representative failure cases involving closely spaced surfaces and complex junctions, to better assess the method's robustness and limitations?

Q2. Beyond NM-CD, NM-F1, and NM-LR, what evidence demonstrates that the reconstructed junctions preserve the correct connectivity between incident surfaces?

Q3. How robust is VoroFacet on geometries with densely connected nonmanifold junctions, particularly in configurations more challenging than those currently demonstrated?

7 comments
AuthorsOpenReview ↗

Responses to Reviewer xooF (1/3)

Thank you for your time and thoughtful review. We appreciate your recognition of our Voronoi-based connectivity construction and the improvements in junction reconstruction while maintaining competitive surface accuracy. To address your concerns, we have added valence-aware junction evaluation and targeted analyses of robustness and failure cases, as detailed below.

Comment 1: Although the paper evaluates complex geometries, the robustness of the method under challenging conditions remains insufficiently characterized. In particular, the evaluation does not systematically assess closely spaced surfaces, complex junction configurations, or self-intersections introduced during refinement. Together with the substantial computational overhead, this makes it difficult to assess the method's practical reliability. ... Could the authors provide additional qualitative results, possibly in a revised supplementary material, particularly representative failure cases involving closely spaced surfaces and complex junctions, to better assess the method's robustness and limitations?

Response: Thank you for this suggestion. In Appendix B of the revised manuscript, we have added qualitative comparisons and representative failure cases covering closely spaced surfaces, high-valence junctions, and sharp features, together with explanations of the observed failures.

We also quantify self-intersections after extraction, remeshing, and each refinement stage, and compare against all baselines on the 50 Non-manifold Objects (Appendix B, Table 8). Stage III reduces the mean percentage of self-intersecting triangles from 0.411% to 0.064%. This final proportion is substantially lower than those of DCUDF (68.352%) and DCUDF2 (43.947%). Results from Table 8 are shown below as mean ±\pm standard deviation. Stage I is evaluated with the temporarily removed faces restored.

Method / stageSelf-intersecting triangles (%)
MeshUDF0.005±0.0050.005 \pm 0.005
CAP-UDF0.000±0.0000.000 \pm 0.000
GeoUDF0.789±1.0520.789 \pm 1.052
MIND3.451±3.5283.451 \pm 3.528
HR-UDF0.000±0.0000.000 \pm 0.000
DualMesh-UDF17.941±4.53017.941 \pm 4.530
DCx0.276±0.7380.276 \pm 0.738
DCUDF68.352±10.89568.352 \pm 10.895
DCUDF243.947±7.75443.947 \pm 7.754
VoroFacet (extraction)0.000±0.0000.000 \pm 0.000
VoroFacet (remeshing)0.001±0.0010.001 \pm 0.001
VoroFacet (Stage I)0.228±0.5230.228 \pm 0.523
VoroFacet (Stage II)0.411±0.5540.411 \pm 0.554
VoroFacet (Stage III, final)0.064±0.0530.064 \pm 0.053

We have also expanded the runtime discussion in Appendix F to describe the implementation's parallelization and the main opportunities for acceleration.

AuthorsOpenReview ↗

Responses to Reviewer xooF (2/3)

Comment 2: The junction metrics primarily measure geometric proximity and junction length rather than directly assessing correct connectivity. Although the authors explain why removing junction-aware re-extraction improves NM-F1, the evaluation still lacks direct evidence that the reconstructed junctions preserve the correct connectivity between incident surfaces. ... Beyond NM-CD, NM-F1, and NM-LR, what evidence demonstrates that the reconstructed junctions preserve the correct connectivity between incident surfaces?

Response: Thank you for this valuable suggestion. We have added valence-aware NM-F1 (VA-NM-F1), where valence denotes the number of incident surface sheets. It requires spatially matched junction samples to have the same valence, penalizing geometrically close junctions that connect the wrong number of sheets. Its definition is provided in Appendix A, and results are reported alongside the existing junction metrics in Tables 1, 3–5, 7, and 12–16 of the revised manuscript.

On Non-manifold Objects and ShapeNet-Car (Tables 1 and 3), VoroFacet achieves the highest VA-NM-F1 among the evaluated methods, indicating better recovery of junctions with the correct incident-sheet counts. The results below are percentages reported as mean ±\pm standard deviation; --- denotes no explicit non-manifold edges. Bold and underline indicate the best and second-best means, respectively.

MethodVA-NM-F1: Non-manifold Objects ↑VA-NM-F1: ShapeNet-Car ↑
MeshUDF7.66‾±5.39\underline{7.66} \pm 5.392.84±1.122.84 \pm 1.12
CAP-UDF------
GeoUDF4.55±3.084.55 \pm 3.0811.20±3.3011.20 \pm 3.30
MIND2.33±4.032.33 \pm 4.034.78±4.294.78 \pm 4.29
HR-UDF------
DualMesh-UDF3.82±2.043.82 \pm 2.0412.71±3.5612.71 \pm 3.56
DCx6.22±8.646.22 \pm 8.6415.09‾±7.22\underline{15.09} \pm 7.22
DCUDF------
DCUDF2------
VoroFacet37.07±17.23\mathbf{37.07} \pm 17.2318.91±8.50\mathbf{18.91} \pm 8.50

In the 20-object ablation study (Table 4), removing junction-aware re-extraction slightly increases NM-F1 but substantially reduces VA-NM-F1 and increases junction redundancy. Bold and underline compare the two variants shown below.

VariantNM-F1 (%) ↑VA-NM-F1 (%) ↑NM-LR → 1
Without junction-aware re-extraction75.67±11.39\mathbf{75.67} \pm 11.396.09‾±15.04\underline{6.09} \pm 15.041.916‾±0.510\underline{1.916} \pm 0.510
Full method74.79‾±14.04\underline{74.79} \pm 14.0438.50±15.88\mathbf{38.50} \pm 15.881.615±0.643\mathbf{1.615} \pm 0.643

This demonstrates that re-extraction substantially improves recovery of the correct junction valence, which NM-CD and NM-F1 do not directly assess.

AuthorsOpenReview ↗

Responses to Reviewer xooF (3/3)

Comment 3: How robust is VoroFacet on geometries with densely connected nonmanifold junctions, particularly in configurations more challenging than those currently demonstrated?

Response: Thank you for this suggestion. We have added qualitative and quantitative tests on synthetic junctions with 3, 6, and 9 sheets meeting along a common edge at equal angles, examining increasing junction valence and decreasing inter-sheet angles. All extractors query the same learned UDF for each configuration. Appendix B of the revised manuscript provides quantitative results in Table 7 and qualitative comparisons, including a representative failure case, in Figure 12.

Results from Table 7 are shown below, grouped by junction configuration. NM-CD is scaled by 10310^3; both F1 scores are percentages. The 2×2\times setting doubles (Rinit,Rgap0,Rgap1,RG,δG)(R_{\mathrm{init}}, R_{\mathrm{gap}}^0, R_{\mathrm{gap}}^1, R_G, \delta_G), keeping the UDF and all other parameters unchanged. --- denotes undefined NM-CD for an empty predicted junction set. Bold and underline indicate the best and second-best results within each configuration, including ties.

ConfigurationMethodNM-CD ↓NM-F1 ↑VA-NM-F1 ↑NM-LR → 1
3 sheetsMeshUDF---0.000.000.000
GeoUDF18.23496.5596.551.068
MIND51.80786.9386.931.270
DualMesh-UDF198.2065.923.79147.333
DCx1.07199.33‾\underline{99.33}99.33‾\underline{99.33}1.046
VoroFacet0.230\mathbf{0.230}99.39\mathbf{99.39}99.39\mathbf{99.39}0.982\mathbf{0.982}
VoroFacet (2×2\times scales)0.301‾\underline{0.301}99.2599.250.977‾\underline{0.977}
6 sheetsMeshUDF72.5490.000.000.035
GeoUDF116.98138.710.000.410
MIND42.42621.790.006.608
DualMesh-UDF196.3966.890.00302.469
DCx2.525‾\underline{2.525}99.100.004.153
VoroFacet3.52099.12‾\underline{99.12}97.01\mathbf{97.01}1.035‾\underline{1.035}
VoroFacet (2×2\times scales)0.389\mathbf{0.389}99.84\mathbf{99.84}96.74‾\underline{96.74}1.015\mathbf{1.015}
9 sheetsMeshUDF63.78418.980.00‾\underline{0.00}0.088\mathbf{0.088}
GeoUDF148.56529.340.00‾\underline{0.00}3.701
MIND58.0861.910.00‾\underline{0.00}10.304
DualMesh-UDF204.8383.360.00‾\underline{0.00}459.370
DCx5.202‾\underline{5.202}71.51‾\underline{71.51}0.00‾\underline{0.00}11.874
VoroFacet11.8521.330.00‾\underline{0.00}8.289
VoroFacet (2×2\times scales)2.135\mathbf{2.135}95.87\mathbf{95.87}23.23\mathbf{23.23}2.938‾\underline{2.938}

With default parameters, VoroFacet achieves 99.39% and 97.01% VA-NM-F1 on the 3- and 6-sheet junctions, respectively, whereas all listed baselines obtain zero VA-NM-F1 on the 6-sheet case. The 9-sheet configuration challenges all compared methods: both the baselines and default VoroFacet obtain zero VA-NM-F1. For VoroFacet, the default cover balls cannot span the wider gap left by the proxy at the acute junction. Doubling the extraction scales improves coverage of this gap and raises NM-F1 to 95.87% and VA-NM-F1 to 23.23%, the highest values in the 9-sheet comparison.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback and recognition of our novel Voronoi facet-selection approach. We have added experiments and analyses addressing their concerns about junction connectivity, proxy dependence, robustness, and evaluation reliability, and expanded the runtime discussion. Revisions are highlighted in magenta in the revised manuscript.

The key revisions are summarized below:

  1. Junction connectivity (Appendix A; Tables 1, 3–4): We introduce VA-NM-F1 to evaluate both junction location and incident-sheet count. Comparisons and ablations demonstrate improved recovery of junction valence and the contribution of junction-aware re-extraction.

  2. Proxy dependence (Appendix B; Table 6; Figure 11): We replace DCUDF2 with DCUDF while keeping all downstream parameters unchanged. Comparable surface accuracy and strong junction reconstruction show that VoroFacet does not require DCUDF2 specifically.

  3. Robustness and failure cases (Appendix B; Tables 7–9; Figures 10, 12–13): We add tests on high-valence junctions, closely spaced surfaces, and sharp features. We also quantify self-intersections throughout the pipeline and across baselines, and report non-manifold edges and self-intersections on manifold inputs.

  4. Evaluation reliability (Appendix A, Table 5; Tables 1–6, 8–16): We test seven GT junction extraction settings and add standard deviations to the aggregate results. VoroFacet retains the best mean NM-CD, NM-F1, and VA-NM-F1 under every tested setting on both junction datasets.

  5. Runtime and acceleration (Appendix F): We describe existing GPU and CPU parallelization, remaining sequential operations, and opportunities for faster proxy construction and incremental second-pass extraction.

We also provide an interactive project page where reviewers can rotate, zoom, and clip reconstructed 3D meshes in synchronized side-by-side views to examine surface geometry and junction connectivity.

Detailed responses are provided in the corresponding reviewer discussions.

AuthorsOpenReview ↗

Looking forward to your further feedback

Dear Reviewers,

Thank you for your time and effort in reviewing our submission. We have provided detailed responses to each of the concerns raised. Could you please let us know whether your concerns have been clearly addressed? We look forward to your further feedback.

Best Regards,

The authors

AuthorsOpenReview ↗

Rebuttal Summary for the Area Chair

Dear Area Chair,

We kindly summarize the reviewers’ comments and our responses from the rebuttal phase, hoping to assist you in efficiently evaluating this process. We sincerely thank the reviewers for their time and constructive feedback. We are pleased that the reviewers recognized VoroFacet’s strong performance in non-manifold junction reconstruction.

The main concerns raised during the review process were:

  • The novelty of our approach (Reviewer NGep).
  • Dependence on the DCUDF2 proxy (Reviewers CjdD and NGep).
  • The need for additional validation of junction connectivity, robustness, and evaluation reliability (Reviewers xooF and CjdD).
  • Further justification regarding applicability across surface types and computational cost (Reviewers NGep, CjdD, and xooF).

We have carefully addressed these concerns in our responses and the revised manuscript as follows:

  • Novelty clarification: We elaborated on VoroFacet’s distinct contribution in constructing single-layer connectivity through Voronoi facet selection and junction-aware re-extraction, rather than inheriting the proxy’s connectivity.
  • Proxy dependence analysis: We added controlled experiments with an alternative proxy, demonstrating that VoroFacet does not require DCUDF2 specifically and retains strong reconstruction performance.
  • Additional experimental validation: We introduced valence-aware junction evaluation, added robustness and ground-truth sensitivity studies, and reported standard deviations, strengthening the evidence for junction reconstruction and evaluation reliability.
  • Enhanced applicability and efficiency discussion: We clarified the existing dataset coverage, provided manifold-input and self-intersection statistics, and expanded the discussion of computational cost and acceleration opportunities, clarifying the scope and limitations of our unified pipeline.

We are encouraged by the positive feedback from the discussion: Reviewer xooF stated that the additional experiments “largely address my concerns” and expressed willingness to raise the rating from Weak Reject to Weak Accept. Reviewer CjdD confirmed that our responses and additional experiments “address my concerns” and indicated that they would consider raising their score. Reviewer NGep has not yet provided further feedback. We are confident that our detailed responses and additional evidence address their concerns and clarify the distinct contribution and applicability of our method.

We have carefully addressed the key concerns raised by the reviewers and strengthened our manuscript. We sincerely thank the reviewers for their insightful comments and you for your time and effort.

Best regards,
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.