acceptodds
Under review as a conference paper at ICLR 2027

Patchwork: A Polyhedral Implicit Representation for Highly Detailed 3D Shapes

Abstract

We introduce patchwork, a highly expressive shape representation that explicitly encodes 3D shapes as unions of polyhedra while admitting a simple closed-form zero-level-set equation. It is grounded in a rigorous mathematical framework that yields provable complexity bounds and guarantees approximation of arbitrary shapes to arbitrary precision. To fit patchwork to a given shape, we pair conventional gradient-based optimization with a novel combinatorial matching scheme. Experiments show that our approach achieves superior quality with 3–6 fewer parameters compared to existing alternatives on average, and scales to highly detailed artist-created assets. Patchwork further supports explicit polygonalization and conforming volume mesh extraction, making it a versatile, scalable, yet compact representation for downstream tasks, with potential for geometric learning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D representation, implicit shape fitting, mesh extraction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_3H5uOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper represents 3D surfaces as the zero level set of the difference between two maxima of affine functions and extracts interfaces between oppositely labeled cells in a power diagram. The fitting algorithm includes geometric initialization, normal-guided discrete matching, assignment stabilization, and pruning. Experiments demonstrate competitive reconstruction quality compared with the baseline.

Critical strengths

Defining surfaces through interfaces between positively and negatively labeled power cells is an interesting implicit–explicit formulation that accommodates complex topology. The optimization challenges are generally well explained, and the proposed solutoin are reasonably motivated.

The main fitting experiments demonstrate strong quantitative and qualitative results, with better reconstruction accuracy and substantially fewer parameters than the evaluated baselines.

Critical weaknesses

The method would be easier to follow with preliminaries connecting Voronoi diagrams, power distances, and affine-function maxima. Figures 3–4 would benefit from labeled sites/cells and numbered stages illustrating surface interfaces, candidate selection, and the effect of matching.

The “without DM” baseline in table 4 also changes smoothing, off-surface regularization, and initialization. Its poorer performance therefore cannot be attributed specifically to removing matching. A controlled ablation should be required.

The current fitting pipeline is limited to closed manifold surfaces and struggles with 3D shape with non manifold structure, open surface and self-intersection, which is widely presented in large 3D shape dataset like Objaverse. Preprocessing algorithm would cause loss of geometry.

The mean Chamfer improvement highlighted in Section 4.2 averages only the seven winning cases, rather than all ten evaluated shapes. ICLR_paper3 Averaging all ten entries in Table 9 (Appendix G.3, p. 24) yields approximately 8.898 × 10⁻³ for Patchwork (5M initialization), versus 3.805 × 10⁻³ for GeomDist. Presenting the favorable subset average without corresponding full-set statistics can be misleading.

The complex-asset scalability study evaluates only ten shapes, which provides limited evidence of robustness across diverse, high-detail geometry. Moreover, the experiments optimize each shape independently rather than train a shared encoder–decoder across a dataset. So it is not sure whether the proposed method could be used to train generative model.

Questions for authors

Can you provide a controlled ablation/intuitive explanation on how the result would be that replaces normal-guided matching with ordinary top-1 assignment while keeping the hard-max formulation, initialization, losses, pruning, and training budget unchanged, to isolate the contribution of matching itself?

3 comments
AuthorsOpenReview ↗

Author Response

We thank the reviewer for the positive, constructive feedback. In the revised manuscript, we have added preliminary background on Voronoi and power diagrams, power distance, and the maxima of affine functions (new Fig. 4). We have also added annotations to Figures 3 and 5 (previous Fig. 4) to improve clarity (W1).

W2 / Q1. Without discrete matching (DM) ablation. To isolate DM, we kept all configurations and the training budget unchanged, modifying only the top-1 assignment. Without DM, the optimization converges to a low-polygonal approximation, resulting in a substantially lower F-Score and greatly diminished detail (See revised Fig. 2).

& CH \downarrow & HS \downarrow & FS \uparrow & NC \uparrow & IoU \uparrow \\\\ \hline \text{w.o. DM} & 3.74 / 4.17 & 5.77 / 9.32 & 37.78 / 31.85 & 94.90 / 94.49 & 92.24 / 94.34 \\\\ \text{Full} & \textbf{1.90} / \textbf{1.80} & \textbf{3.81} / \textbf{6.78} & \textbf{73.22} / \textbf{77.27} & \textbf{98.05} / \textbf{98.17} & \textbf{98.39} / \textbf{99.14} \end{array}$$ We have replaced the DM in Tab. 4 with the new results and renamed the previous smooth variant ablation to “Hard Maximum (HM)”. **W3. Closed manifold surface assumption**. We fully agree with the reviewer’s assessment regarding our closed-manifold assumption, and acknowledge that preprocessing comes at the cost of some geometric fidelity. This is precisely why we kept the shape datasets unprocessed wherever possible (ABC, Thingi10K, Roof Modeling, and GeomDist), preprocessing only when necessary (ShapeNet). Although large-scale datasets such as Objaverse exhibit far more diverse topology, converting shapes to manifolds remains a standard and well-established practice in 3D shape learning and generation [1, 2, 3]. We have emphasized our closed-surface assumption throughout the introduction. **W4. GeomDist metrics**. We did not intend to present a favorable subset. Of the 10 shapes evaluated for scalability and visualized in the Teaser (Fig. 1), the 7 we reported are those that satisfy our manifold assumption, and on all of them our method outperforms GeomDist. For the remaining 3 non-manifold shapes, while our reconstructions remain visually faithful from the exterior, interior artifacts arising from non-manifoldness disproportionately lower the averaged statistics over all 10 shapes. We openly acknowledge this limitation in the main manuscript, and provide full per-shape metrics and visualizations of the failure cases in the supplementary material. We think the current comparison is fair and prefer to keep experiments that are outside of the scope of the paper in the supplementary. Alternatively, we could compare to the processed shapes as GT, but since the competing methods are very slow (40 h on 4 A100s per shape), we cannot finish such a comparison in the time frame of the rebuttal because we cannot access 4 x A100s. **W5. Generative potential**. Since submission, we have made promising progress in converting patchwork into a regular, grid-like data structure. Specifically, we can analytically convert complex assets (e.g., Loong in Fig. 1) represented by patchworks into a pair of sparse 4K two-channel (RG) textures via octahedral mapping, near-losslessly and with storage equivalent to two sparse $256^3$ grids. We are therefore optimistic about its generative potential and leave this to future work. **References** - [1] Li, Yangguang, et al. "Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models." IEEE Transactions on Pattern Analysis and Machine Intelligence (2025). - [2] Chen, Rui, et al. "Dora: Sampling and benchmarking for 3d shape variational auto-encoders." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025. - [3] Hunyuan3D, Team, et al. "Hunyuan3d 2.1: From images to high-fidelity 3d assets with production-ready pbr material." arXiv preprint arXiv:2506.15442 (2025).
AuthorsOpenReview ↗

Follow-up on rebuttal

Thank you again for your thoughtful and positive review. As the discussion period draws to a close, we wanted to briefly follow up and highlight the main points of our rebuttal:

  • Discrete Matching (DM) ablation (W2 / Q1): The newly isolated ablation shows that DM substantially improves fitting quality; without it, the results degrade to low-polygon approximations.
  • Additional revisions (W1, W3, W4): We have added background on Voronoi/power diagrams, power distance, and max-affine functions; improved the figure annotations; clarified the manifold assumption; and explained the reported GeomDist metrics.

In our response, we have also discussed the method’s generative potential (W5). We would appreciate knowing whether our responses have addressed the concerns raised in your initial review, and we are happy to clarify anything that remains unresolved.

Area_Chair_tXohOpenReview ↗

Looking for Reviewer's reply !!!

Dear Reviewer 3H5u,

Thank you for providing your valuable review comments on this manuscript.
Do the authors’ responses address all of your questions?
We look forward to hearing from you.

Best Regards.

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

This paper introduces "patchwork," a 3D shape representation that encodes surfaces as the zero level set of a difference-of-maxima of linear functions, corresponding geometrically to the interface between two power diagrams. Grounded in Viro's patchworking theory, the authors prove it can approximate arbitrary shapes to arbitrary precision and derive complexity bounds. A discrete matching optimization scheme is proposed to fit the representation while preserving hard maximum assignments. Experiments on ABC, Thingi10K, and ShapeNet show consistent improvements over neural implicit and foam-based baselines with 3–6× fewer parameters, along with support for explicit polygonalization and conforming volume mesh extraction.

Critical strengths
  1. Principled mathematical foundation with provable expressiveness. Unlike most neural implicit representations whose approximation capability is argued empirically, patchwork provides formal guarantees: it can approximate any shape to arbitrary precision, and the paper derives complexity lower bounds via classical algebraic geometry (Bézout's theorem, Harnack's inequality). This theoretical grounding is consequential because it establishes that the representation's capacity is not an artifact of over-parameterization but a provable property, lending confidence that scaling parameters will reliably improve fidelity.

  2. Strong empirical results with significantly fewer parameters. The quantitative improvements are substantial and consistent across three datasets of varying complexity. On ShapeNet, the IoU gap over the second-best method (96.69 vs. 70.99) is remarkable, achieved with only 17% of the matched parameter budget after pruning. This demonstrates that the discrete matching scheme successfully unlocks the theoretical expressiveness in practice, rather than the gains coming merely from a larger model.

Critical weaknesses
  1. Lack of algorithm pseudocode. The method section is information-dense, with five interleaved optimization steps (geometric initialization, discrete matching, assignment stabilization, cell fitting, and pruning) described across multiple paragraphs and equations. A consolidated Algorithm box summarizing the full pipeline would significantly improve readability and reproducibility.

  2. Lack of quantitative noise robustness analysis. Fig. 9 shows that noisy constraints degrade results to low-poly-like outputs, but the paper provides no quantitative characterization of this sensitivity — e.g., at what angular noise level in normals does quality begin to deteriorate? What fraction of missing samples causes failure? In practice, point-normal constraints are often estimated from neural rendering density fields or scanned data, both of which introduce noise. A more thorough discussion and empirical study of this limitation is needed to assess the method's real-world applicability.

Questions for authors
  1. The hyperparameter τ\tau (controlling initialization offset δ\delta = 0.1/τ\tau) varies across experiments (1.0, 2.0, 8.0). Could you provide guidance on how to select τ\tau for a new shape? How sensitive is the final result to this choice — e.g., what happens if τ\tau is off by a factor of 2?
  2. On Thingi10K, your Hausdorff Distance (6.78) is notably worse than several baselines (SPSR: 4.06, IGR: 4.08) despite having the best Chamfer Distance. Can you characterize the failure cases — are they primarily due to unpruned floaters, or overshooting polyhedra at thin features?
  3. The paper mentions potential for "geometric learning" but provides no preliminary evidence. Do you envision patchwork being used as a decoder output representation in a generative pipeline (e.g., a network predicting patchwork parameters)? If so, how would you handle the variable number of cells across shapes?
4 comments
AuthorsOpenReview ↗

Author Response

We thank the reviewer for their honest and constructive feedback. We have added the pseudocode to the supplementary material (Sec. H.5) to improve readability (W1).

W2. Angular Noise Robustness. Our method provides explicit and precise control over angular noise tolerance through the stabilization threshold described in Sec. 3, Assignment Stabilization. For clean shapes, we assume a 5∘5^\circ angular tolerance to account for the error inherent in approximating curved surfaces with flat patches. Normal deviations within this threshold are treated as noise and absorbed into the cell pair, while deviations exceeding it are treated as mis-coverage, and the gradient is reassigned to the most suitable pair in the neighborhood via discrete matching. We have rewritten this section for clarity and further provide visualizations in the supplementary material (Sec. H.3).

Although neural density fields are widely used for real-world object capture, they are not the only representation. As shown in Figure 8, our method integrates naturally with surface-based neural rendering methods such as Neural Deferred Shading [1] and NeuS variants [2, 3], enabling fracture simulation directly on real-world captures.

Q1. Hyperparameter τ\tau. We add an ablation of τ\tau in the supplementary material (Sec. H.4). τ\tau sets the offset of the Voronoi sites at initialization. If τ\tau is too large and overshoots the primary features, the field becomes noisy, the candidate set fails to converge for pruning, and the resulting geometry is noisy. If τ\tau overshoots only small features while preserving the large structure, fine geometric details are lost and cannot be recovered. Decreasing τ\tau recovers more detail but increases the number of remaining parameters, and beyond a certain point the quality gains diminish while the parameter count keeps growing.

In practice, we choose τ\tau by visualizing our initialization, which is training-free and can be meshed directly via Marching Cubes or explicit polygonization. As demonstrated in our experiments, a single τ\tau is reasonably robust across all shapes within a dataset, despite the varying feature sizes within each shape.

Q2. Thingi10K Hausdorff explanation. We characterize two failure cases on Thingi10K:

  1. Non-manifold (ID: 53920, 76277, 77319, 77513, 81528, 92880, 93497, 96637), where the surface has self-intersections and negated surface normals, resulting in overshooting polyhedra.
  2. Non-closed (ID: 64444, 85540, 85546), where inside the bounded shape there are floating open polygon soups resulting from scanning noise;

We perform an analysis to exclude those subsets and re-evaluate the metrics, along with SPSR and IGR, and report the metrics as below.

MethodCD↓HD↓F1↑NC↑IoU↑IGR1.944.0858.6098.9598.06BaselineSPSR1.884.0665.7197.8297.59Ours1.806.7877.2798.1799.14IGR1.93(−0.01)3.87(−0.21)58.60(0.00)98.95(0.00)98.07(+0.01)w.o.Non-manifoldSPSR1.91(+0.03)4.08(+0.02)65.12(−0.59)97.76(−0.06)97.48(−0.11)Ours1.74(−0.06)5.17(−1.61)76.93(−0.34)98.22(+0.05)99.17(+0.03)IGR1.92(−0.02)3.97(−0.11)59.26(+0.66)98.95(0.00)98.04(−0.02)w.o.Non-closedSPSR1.87(−0.01)3.93(−0.13)66.01(+0.30)97.82(0.00)97.55(−0.04)Ours1.51(−0.29)5.91(−0.87)77.77(+0.50)98.42(+0.25)99.32(+0.18)IGR1.91(−0.03)3.75(−0.33)59.31(+0.71)98.96(+0.01)98.05(−0.01)w.o.BothSPSR1.90(+0.02)3.94(−0.12)65.42(−0.29)97.77(−0.05)97.43(−0.16)Ours1.42(−0.38)4.16(−2.62)77.46(+0.19)98.50(+0.33)99.37(+0.23)\begin{array}{c|cccccc}\hline&\text{Method}&\text{CD}\downarrow&\text{HD}\downarrow&\text{F1}\uparrow&\text{NC}\uparrow&\text{IoU}\uparrow\\\\\hline&\text{IGR}&1.94&4.08&58.60&\textbf{98.95}&98.06\\\\\text{Baseline}&\text{SPSR}&1.88&\textbf{4.06}&65.71&97.82&97.59\\\\&\textbf{Ours}&\textbf{1.80}&6.78&\textbf{77.27}&98.17&\textbf{99.14}\\\\\hline&\text{IGR}&1.93(-0.01)&\textbf{3.87}(-0.21)&58.60(0.00)&\textbf{98.95}(0.00)&98.07(+0.01)\\\\\text{w.o.Non-manifold}&\text{SPSR}&1.91(+0.03)&4.08(+0.02)&65.12(-0.59)&97.76(-0.06)&97.48(-0.11)\\\\&\textbf{Ours}&\textbf{1.74}(-0.06)&5.17(-1.61)&\textbf{76.93}(-0.34)&98.22(+0.05)&\textbf{99.17}(+0.03)\\\\\hline&\text{IGR}&1.92(-0.02)&3.97(-0.11)&59.26(+0.66)&\textbf{98.95}(0.00)&98.04(-0.02)\\\\\text{w.o.Non-closed}&\text{SPSR}&1.87(-0.01)&\textbf{3.93}(-0.13)&66.01(+0.30)&97.82(0.00)&97.55(-0.04)\\\\&\textbf{Ours}&\textbf{1.51}(-0.29)&5.91(-0.87)&\textbf{77.77}(+0.50)&98.42(+0.25)&\textbf{99.32}(+0.18)\\\\\hline&\text{IGR}&1.91(-0.03)&\textbf{3.75}(-0.33)&59.31(+0.71)&\textbf{98.96}(+0.01)&98.05(-0.01)\\\\\text{w.o.Both}&\text{SPSR}&1.90(+0.02)&3.94(-0.12)&65.42(-0.29)&97.77(-0.05)&97.43(-0.16)\\\\&\textbf{Ours}&\textbf{1.42}(-0.38)&4.16(-2.62)&\textbf{77.46}(+0.19)&98.50(+0.33)&\textbf{99.37}(+0.23)\\\\\hline\end{array}

With those two sets removed, our Hausdorff distance (HS) drops to a similar level as SPSR and IGR, while our Chamfer distance (CD), F-Score (FS), and Intersection-over-Union (IoU) remain the top throughout the analysis. This suggests a global SDF is more robust to topological defects, while our method excels at accuracy. We have added a dedicated section (Sec. H.2) in the supplementary material with visualizations (Fig. 19) to explain the two failure cases.

Q3. Geometric Learning potential. For patchwork, the most direct integration for geometric learning is via a transformer, which handles varying length by padding or an end-of-sequence token. A related predecessor is MASH [4], which, like our method, encodes unordered surface patches compactly as coefficients. They successfully use the VecSet [5] to learn the coefficient distribution for generative tasks, which should apply directly to ours. We are also actively exploring a regular grid structure, such as Trellis [6]. Since submission, we have made promising progress in converting patchwork coefficients analytically into a pair of sparse two-channel (RG) textures under octahedral mapping. Even for highly complex shapes such as the Loong in Fig. 1, two 4K textures are sufficient for near-lossless conversion, that is, of the same size as two sparse 256³ grids. We are optimistic about its geometric learning potential.

References

  • [1] Worchel, Markus, et al. "Multi-view mesh reconstruction with neural deferred shading." 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022.
  • [2] Wang, Peng, et al. "Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction." arXiv preprint arXiv:2106.10689 (2021).
  • [3] Liu, Yuan, et al. "Nero: Neural geometry and BRDF reconstruction of reflective objects from multiview images." ACM Transactions on Graphics (ToG) 42.4 (2023): 1-22.
  • [4] Li, Changhao, et al. "Mash: Masked anchored spherical distances for 3d shape representation and generation." Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Papers. 2025.
  • [5] Zhang, Biao, et al. "3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models." ACM Transactions On Graphics (TOG) 42.4 (2023): 1-16.
  • [6] Xiang, Jianfeng, et al. "Native and compact structured latents for 3d generation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.
AuthorsOpenReview ↗

Follow-up on rebuttal

Thank you once more for your careful review and constructive feedback. With the discussion period nearing its end, we would like to briefly summarize the key points of our rebuttal:

  • Angular noise robustness (W2): We have explained and demonstrated that our method provides explicit control over angular noise tolerance.
  • Hyperparameter τ\tau (Q1): We have ablated τ\tau, illustrating its trade-off between reconstruction quality and parameter size, and provided practical guidance for choosing it.
  • Thingi10K Hausdorff distance (Q2): Through a series of analyses, we have shown that the failure cases stem from the inputs themselves: non-manifold or non-closed meshes, both of which violate our assumptions.
  • Additional revisions (W1): We have added algorithm pseudocode to facilitate reproducibility.

In our response, we have also discussed the method’s potential for geometric learning (Q3). We would be grateful to hear whether these responses have resolved your concerns. If any questions remain, we would be glad to address them before the discussion closes.

I thank the authors for their detailed responses and additional experiments, which address my concerns. In view of the paper’s contributions, I maintain my original recommendation.

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

The paper presents a power-diagram-based implicit representation of 3D shapes that supports both implicit queries and explicit polygonal geometry. Its main algorithmic contribution is a new matching-based optimization scheme for fitting the representation to oriented surface samples. Experiments on ABC, Thingi10k, ShapeNet, a roof dataset, and ten highly detailed assets report improvements in reconstruction accuracy and parameter efficiency, alongside demonstrations of polygonal extraction and fracture. The method primarily targets clean, oriented samples of closed surfaces, with limited robustness to noise and non-manifold internal structures.

Critical strengths

The method selects a positive/negative cell pair according to normal alignment and uses a coverage loss to make the selected pair active. This allows samples with incompatible normals to be assigned to different surface patches, reducing conflicting updates while retaining hard maxima and sharp interfaces. Combined with assignment stabilization and pruning, it provides a practical approach to fitting detailed geometry.

The central optimization insight is interesting: incompatible normal constraints should be assigned to different local faces rather than smoothed into the same active face.

Critical weaknesses

A closely related recent baseline, DCCVT, is missing. DCCVT also combines a polyhedral Voronoi tessellation with an implicit shape field for surface reconstruction. Its use of an auxiliary SDF is a methodological difference, not sufficient grounds to exclude it from comparison. A quantitative comparison is needed to establish whether Patchwork's SDF-free design improves reconstruction quality or efficiency over this recent alternative.

The main baselines are matched to the initial parameter budget, while the proposed method is subsequently pruned. This is reasonable evidence of compact final representations, but not a complete rate-distortion comparison at matched achieved error or final storage. It also does not account for auxiliary fitting state, acceleration structures, extraction cost, or encoded mesh size.

Questions for authors
  1. Please report the actual generator counts entering the optimizer and reconcile them with Table 8. Can you provide results using the original initialization under matched actual parameter budgets?

  2. What specific result supports the contribution that Patchwork “bridges neural networks and known mathematical theories”? You should show some applications that 3D Voronoi can't do but your representation based on power diagram can, or tone down this contribution.

4 comments
AuthorsOpenReview ↗

Author Response

We thank the reviewer for their thoughtful evaluation and constructive feedback.

W1. DCCVT baseline. We are grateful to the reviewer for pointing out the missing baseline. We set the DCCVT primitives count to match the same budget as ours (Tab. 8) and report quantitative results as follows, where our method consistently outperforms DCCVT in terms of CD, FS, and IoU. The qualitative results are available in our revised manuscript.

Roof Modeling

\text{Dataset} & \text{Method} & CH \downarrow & HS \downarrow & FS \uparrow & NC \uparrow & IoU \uparrow & \\#\text{Params} \\\\ \hline \text{Roof} & \text{DCCVT} & 3.10 & 5.08 & 45.10 & 96.46 & 96.84 & 8232 \\\\ & \text{Ours} & \textbf{1.21} & \textbf{0.66} & \textbf{82.16} & \textbf{99.74} & \textbf{99.98} & 2195 \\\\ \hline \text{ABC / Thingi10K} & \text{DCCVT} & 2.84 / 2.62 & 4.45 / \textbf{6.02} & 55.86 / 59.81 & 91.56 / 91.61 & 77.76 / 84.71 & 128625 \\\\ & \text{Ours} & \textbf{1.90} / \textbf{1.80} & \textbf{3.81} / 6.78 & \textbf{73.22} / \textbf{77.27} & \textbf{98.05} / \textbf{98.17} & \textbf{98.39} / \textbf{99.14} & 45192 \\\\ \hline \text{ShapeNet} & \text{DCCVT} & 3.52 & 4.73 & 24.93 & 68.04 & 24.17 & 526848 \\\\ & \text{Ours} & \textbf{1.65} & \textbf{3.28} & \textbf{64.26} & \textbf{97.77} & \textbf{96.69} & 90808 \end{array}$$ Although the global auxiliary SDF helps resolve topological defects, as indicated by the lower HD on Thingi10K, it also introduces two failure modes: 1) When the SDF gives poor initialization for later optimization stages (DCCVT paper, Sec. 11), DCCVT produces failure cases similar to those reported in their paper (e.g., shape 398259). 2) The difficulty of recovering sharp creases on CAD models, possibly due to its curvature regularization and the inherent smoothness of the neural SDF initialization. In contrast, our SDF-free design can circumvent both of them. **W2. Complete rate-distortion comparison**. We agree that a complete rate-distortion comparison would be a valuable addition. However, such a comparison is not well-defined across all baselines and would require us to write quite a bit of additional code for the different algorithms we compare to. Writing and tuning the code is too time-consuming for the rebuttal, but we will make a best effort to include such a comparison in an eventual final version. Similarly, although the parameter count does not capture auxiliary costs, such as fitting state, acceleration structures, extraction cost, or encoded mesh size, these costs are not intrinsic to the representations being compared: they vary across implementations and often reflect a trade-off between memory and speed. This is precisely why we compare all methods at a matched initial parameter budget, which is intrinsic to each representation, well-defined for every method, and if anything, shows competitors to be more favorable than they are compared to our method. At this setting, our method achieves lower error while simultaneously using far fewer parameters than the initial budget. In addition, we report our per-iteration time and VRAM usage (Tab. 5) to further reflect our auxiliary cost in practice. Since the results of our method are so strong, we expect to dominate all competing work in the rate-distortion comparison. **Q1. Report the actual generator counts entering the optimizer, and results using matched actual parameter budgets**. The number of generators reported in Tab. 8 is precisely the number entering the optimizer. The average number of parameters after the optimization is reported in Tables 2, 3, and 10. We are not sure if we understood the question correctly. If we add the average final number of generators (with parameters matched to Tables 2, 3, and 10) to Table 8, would that help answer the question? As described in Sec. 3 (“Geometric Initialization”), we sample N oriented surface points per shape and offset each along and opposite its normal, producing two generators per sample (one for positive and one for negative cells), i.e., 2N generators, ALL enter the optimizer. Each generator corresponds to a linear function parameterized by $a \in \mathbb{R}^3$ and $c \in \mathbb{R}$ (4 floats), giving 8N parameters in total. Tab. 8 reports these values directly: #Prims = 2N, #Params/Prim = 4, and #Params = 8N, matching the parameter budgets used in all experiments: we use N = 1024 for Roof Modeling (Suppl. Sec. G.4), N = 16384 for ABC and Thingi10K (Sec. 3, “Model capacity”) and N = 65536 for ShapeNet (Sec. 3, “Model capacity”). This yields, as listed in Tab. 8, 2048 / 32768 / 131072 generators (#Prims) and 8192 / 131072 / 524288 parameters (#Params) for ours.

Q2. Bridges neural networks and known mathematical theories. Patchwork is a particular case of a neural network, the 2-layer maxout network [1]. We state and prove multiple theorems on this particular type of networks (see Theorems 1–4 in Appendices C–D ).

Q2. 3D Voronoi. We expect that 3D Voronoi and our representation can handle the same applications. However, our representation has two key advantages:

  1. Our representation is always more expressive. Our formulation originates from tropical geometry, where the linear functions ⟨a,x⟩+c\langle a, x \rangle + c impose no restriction on cc. The resulting cell decomposition is exactly a power diagram, with the Voronoi diagram as the special case c=−14∣a∣2c = -\frac{1}{4}|a|^2.
  2. Our representation is always more compact. Compared to the Voronoi diagram, the generator of the power diagram can land outside its cell, offering extra flexibility. To quantify this, we re-run all experiments with 3D Voronoi diagram parameterization (i.e., c=−14∣a∣2c = -\frac{1}{4}|a|^2 as explained in Sec. 3 “Geometric Initialization”), and report results below. Our power diagram parameterization outperforms the Voronoi-parameterized variant, while requiring 32–53% fewer parameters.
\text{Dataset} & \text{Method} & CH \downarrow & HS \downarrow & FS \uparrow & NC \uparrow & IoU \uparrow & \\#\text{Params} \\\\ \hline \text{Roof} & \text{Ours (Voro)} & 1.67 & 2.02 & 81.89 & 99.63 & 99.92 & 4137 \\\\ & \text{Ours} & \textbf{1.21} & \textbf{0.66} & \textbf{82.16} & \textbf{99.74} & \textbf{99.98} & 2195 \\\\ \hline \text{ABC / Thingi10K} & \text{Ours (Voro)} & 1.95 / 1.89 & 4.77 / 7.48 & 72.19 / 75.13 & 97.99 / 97.96 & \textbf{98.41} / 98.93 & 66570 \\\\ & \text{Ours} & \textbf{1.90} / \textbf{1.80} & \textbf{3.81} / \textbf{6.78} & \textbf{73.22} / \textbf{77.27} & \textbf{98.05} / \textbf{98.17} & 98.39 / \textbf{99.14} & 45192 \\\\ \hline \text{ShapeNet} & \text{Ours (Voro)} & 1.82 & 3.74 & 59.50 & 95.13 & 90.91 & 191928 \\\\ & \text{Ours} & \textbf{1.65} & \textbf{3.28} & \textbf{64.26} & \textbf{97.77} & \textbf{96.69} & 90808 \end{array}$$ **References** - [1] Goodfellow, Ian, et al. "Maxout networks." International conference on machine learning. PMLR, 2013.
AuthorsOpenReview ↗

Follow-up on rebuttal

We sincerely appreciate the time and effort you put into reviewing our paper. As the discussion phase comes to an end, we would like to kindly follow up on our response.

  • DCCVT baseline (W1): We have included DCCVT as an additional baseline and provided its qualitative and quantitative results for all experiments in the revised manuscript.
  • Complete Rate-distortion comparison (W2): We have explained that our method achieves both lower error and fewer final parameters than the baselines, which already places it favorably in terms of rate-distortion. We will make our best effort to include the full rate-distortion comparison in the final version, as it requires more implementation and tuning than the rebuttal period allowed.
  • Generator counts and parameter budgets (Q1): We have clarified that Table 8 reports exactly the generator counts entering the optimizer, using our proposed initialization, and that the average final parameter counts are reported in Tables 2, 3, and 10.. We would greatly value your clarification on what “matched actual parameter budgets” refers to beyond these numbers, and we are ready to provide the additional results if helpful.
  • Bridges neural networks and known mathematical theories (Q2): We have emphasized that patchwork is a two-layer maxout network and pointed to the theorems we proved for this network class in the supplementary material.
  • 3D Voronoi (Q2): We have clarified that, while addressing the same applications as 3D Voronoi, our representation is always more expressive and more compact, as demonstrated by additional ablations against the 3D Voronoi parameterization.

Please let us know whether our responses have addressed the points raised in your review. We remain happy to provide further clarification on any remaining concerns.

Thanks for the careful rebuttal. Many of the issues are well solved. I raise my recommendation from weak rejection to weak accept.

Comments on the paper

1 comment
Area_Chair_tXohOpenReview ↗

Looking for reviewers' reply !!!

Dear Reviewers,

The authors have posted their rebuttal. Please read it and reply in the discussion, covering:

Whether each of your questions and concerns has been addressed. Whether your rating and recommendation change, and why. Any remaining concern that must be settled before the meta-review. Best.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.