acceptodds
Under review as a conference paper at ICLR 2027

SIEVE: Topology-Flexible Mesh Compression via Surface-Intersection Coding

Abstract

Existing implicit representation-based 3D mesh compression methods typically adopt signed distance fields representation, which can struggle to preserve open boundaries and thin structures. In this paper, we present SIEVE, Surface Intersection Encoding via Voxel Edges, a 3D mesh compression framework that encodes local surface intersections to accommodate diverse surface geometries and topologies. Its compact Dual Vertex–Edge (DV-Edge) representation stores grid-edge intersection flags and precomputed dual vertex positions on a sparse regular grid, encoding local geometry and the information needed to construct mesh connectivity. A hierarchical autoencoder compresses DV-Edge into an entropy-coded bitstream for mesh reconstruction. Extensive experiments on challenging scanned meshes and geometrically irregular surfaces demonstrate improved rate–distortion performance over the evaluated baselines, with better preservation of fine surface structures. The code will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Mesh Compression

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2xtWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SIEVE is a learned lossy mesh codec that replaces SDFs with grid-edge surface intersections to better handle open boundaries and thin structures. Each mesh becomes DV-Edge, a sparse voxel grid storing edge-intersection flags and a precomputed dual vertex per voxel. A sparse-conv autoencoder with entropy coding compresses it, and the mesh is recovered by dual contouring. On five datasets, it reports large BD-rate savings over Draco, point-cloud codecs, and SIR-SNC (Chamfer distance and F-score), with fast encoding and decoding.

Critical strengths
  1. It brings modern 3D representations to mesh compression. The paper adapts SDF-free, intersection-based dual-grid representations from recent 3D generation work to learned compression. This addresses a real weakness of SDF-based codecs: open boundaries, thin structures, and fragile sign computation. It’s a sensible and timely direction for the field.

  2. Consistent empirical gains. SIEVE outperforms all baselines across five diverse test datasets, including real scans, with cross-dataset generalization. The gains also hold against a parameter-matched SIR-SNC, and encoding and decoding are fast. Within the evaluated setting, the results are strong.

Critical weaknesses
  1. Limited technical novelty and missing related work. a. DV-Edge, which stores per-voxel edge-intersection flags plus a precomputed QEF dual vertex on a sparse grid, closely matches existing representations: TRELLIS.2’s O-Voxel (essentially the same, minus splitting weights), Faithful Contouring, and Neural Dual Contouring. None of these is discussed as related work; TRELLIS.2 is cited only for its sparse-conv backend. b. The codec is also standard and largely borrowed: sparse-conv autoencoder, factorized entropy model (adopted from SIR-SNC), and G-PCC coordinate coding. c. The main ablation (Fig. 7) compares only against the authors’ own LSF variant, not these prior representations. As a result, the contribution reduces to combining existing components.

  2. Evaluation does not support the claims. a. Draco comparison is unequal. SIEVE remeshes and discards the original connectivity and attributes, while Draco preserves them. No simplify→Draco or V-DMC baseline is provided. b. SDF baseline is questionable. A minimal SDF should need fewer values per site, so SIR-SNC’s 2.7× more sites suggests configuration issues (band width, sign computation on open meshes, retraining) rather than an inherent limit. c. Metrics are topology-blind. The paper claims “topology-flexible” compression, yet CD and F-score barely register holes, spurious faces, flipped orientations, or non-manifold edges. These are the main failure modes of flag errors, which the paper itself acknowledges as holes. Together, these make the reported gains hard to interpret.

Questions for authors
  1. (Clarification) How does DV-Edge differ from TRELLIS.2’s O-Voxel, Faithful Contouring, and Neural Dual Contouring? What is novel beyond applying such a representation to compression?
  2. (Clarification) Is a voxel “active” if any of its 12 edges is crossed, or only its 3 owned edges? Is xᵢ local to its voxel and clamped inside it? How are degenerate QEFs handled, given that no regularization is used?
  3. (Clarification) With unsigned flags and no normals, how are face orientation and quad triangulation determined?
  4. (Clarification) How are cases handled where an edge is predicted as crossed but one of its 4 incident voxels is predicted inactive? Is this the source of the holes in Fig. 13?
  5. (Clarification) How does SIR-SNC preprocess meshes and compute signs on open meshes? What truncation band is used, and was default SIR-SNC retrained on the same data?
  6. (Clarification) Were the thresholds τᵤ and τₑ selected on the test set?
  7. (Experiment) Please report topology metrics (hole/boundary-edge count, non-manifold edges, orientation consistency) and Hausdorff distance.
  8. (Experiment) Please provide a rate breakdown (coordinate vs. feature bits), and compare against a tight-band SDF codec on a watertight subset.
  9. (Future work) Consider stronger entropy models (hyperprior or context models), and loss weighting that prioritizes flag accuracy over vertex positions.
7 comments

We thank the reviewer for the careful review and for recognizing the timely direction, the consistent gains across five datasets, the parameter-matched comparison, and the fast encoding and decoding. We address each concern below. Revisions are marked in red in the updated PDF.

W1a/W1c/Q1. Relation to prior DC-based representations and our compression contribution. We disagree that sharing a dual-contouring parameterization reduces SIEVE to a combination of existing components. Our contribution is a mesh compression framework, which codes sparse support and surface attributes into a bitstream and learns to reconstruct the mesh under an explicit bitrate constraint. O-Voxel, Faithful Contouring, and NDC target generation or reconstruction; they do not provide this mesh compression framework. DV-Edge shares their DC foundation, while its sparse flags-and-vertices parameterization serves as the input to our codec. O-Voxel additionally stores splitting weights, while Faithful Contouring uses anchor normals and signed edge flux. NDC learns to predict flags and vertices from geometric inputs, whereas SIEVE learns to compress and reconstruct these attributes under a bitrate constraint.

In SIEVE, DV-Edge is derived from LSF by retaining its intersection flags and replacing intersection positions and normals with precomputed QEF vertices, reducing the attributes from 15 to 6. Fig. 7 demonstrates the rate--distortion benefit of this choice under the same codec. Table 1 further reports 50--61% BD-BR savings in CD over the SDF-based SIR-SNC baseline. These results support our compression design; the contribution does not depend on claiming a new per-cell DC primitive. We have clarified the relationship to these methods in Sec. 3.2 of the revision.

W1b. Codec components. We do not claim novelty in the individual codec components: the sparse-convolutional autoencoder, factorized entropy model, and G-PCC coordinate coding are established techniques. Their reuse does not negate the compression contribution described above. SIEVE achieves its rate--distortion gains with these standard components, without requiring a more elaborate network or entropy model. This further underscores the value of our representation choice for mesh compression.

W2a. Comparison with V-DMC beyond Draco. To address your concern about the Draco comparison, we add V-DMC, an MPEG mesh codec incorporating mesh simplification. As shown in Fig. 4 and Table 1 of the revision, SIEVE achieves 11--80% BD-BR savings for CD and 7--75% for F-score over V-DMC across the evaluated datasets. This comparison supports SIEVE's compression advantage beyond connectivity-preserving Draco. For convenience, we reproduce the relevant results from Table 1 below.

BD-BR (%) of SIEVE relative to V-DMC, reported as CD/F-score. Negative values indicate bitrate savings.

BaselineShapeNetScanNetAOM + RWTTObjaverse
V-DMC−80.4/−74.7−11.0/−7.2−44.3/−42.7−59.2/−54.8

W2b/Q5. SDF baseline and site count. We evaluate SIR-SNC using its official code and released checkpoints, which were trained on the same 3,000 AMA/DT4D/Thingi10K meshes as SIEVE. SIR-SNC-32 is also retrained on the same dataset. Meshes are only normalized to [−1,1]3[-1,1]^3 , and signs are computed by ray-casting parity as in the SIR-SNC paper. The default distance threshold is one voxel diagonal, following SIR-SNC.

The SIR-SNC paper ablates the truncation band (Sec. 4.3 and Fig. 8 of the SIR-SNC paper) and shows that distortion increases once the band is narrower than about 0.75 voxel diagonals. We therefore use its official default threshold in our comparison. Fewer attributes per site do not imply fewer active sites. SIR-SNC stores SDF samples within its prescribed truncation band around the surface, whereas DV-Edge stores only surface-intersecting cells. The reported 2.7×2.7\times site count in Table 4 of our paper reflects this comparison under SIR-SNC's official settings.

W2c/Q7. Topology metrics and Hausdorff distance. By "topology-flexible", we mean that DV-Edge accommodates open surfaces and non-manifold configurations through edge intersections, without requiring the inside/outside signs used by the standard SDF/MC pipeline. The original manuscript evaluates reconstruction fidelity using CD, F-score, NC, and visual comparisons, following standard practice in lossy mesh compression.

To address your concerns about holes, non-manifold edges, and flipped orientations, Table 2 of the revision, which is also shown below reports Boundary CD, component and hole counts, non-manifold-edge ratio, orientation consistency, and Hausdorff distance at matched bitrates. It is worth noting that Boundary CD and Hausdorff distance better reflect topology preservation in lossy compression, as they capture misplaced boundaries and large local errors, while component and hole counts are neither size- nor position-aware. SIEVE achieves the lowest Boundary CD and Hausdorff distance, demonstrating that it better preserves boundaries and local structures.

Table 2: Topology-aware metrics at matched rates. Draco and V-DMC are omitted because their lowest rates exceed this target rate on some test sets. We also report the topology statistics of ground truth for reference.

MethodB-CD (10⁻²)↓HD (10⁻²)↓OC (%)↑NM (%)#CC#H
GT––99.990.39330.5435.2
G-PCC6.264.38100.00.04191.343.1
SparsePCGC6.366.13100.00.02118.373.7
SIR-SNC5.995.43100.00.00162.5345.2
SIEVE4.772.7999.825.27116.7247.5
SIEVE + hole filling4.372.7899.815.24116.7218.6

Q2--Q4. Representation details.

(Q2) A voxel is active if any of its 12 edges is crossed; all four voxels incident to a crossed edge are therefore active. Each edge flag is stored only once, among the three outgoing edges owned by the voxel at its origin. xix_i is stored in local coordinates normalized to [0,1]3[0,1]^3 within its voxel. During encoding, the QEF is minimized under the constraint that the vertex lies inside the voxel, and the decoded coordinates are clamped to the voxel. For degenerate QEFs (e.g., planar configurations), our implementation adds a small Tikhonov term that pulls the solution toward the mean of the intersection points. This term was omitted from Eq. (3) of the revision for brevity. No QEF solve is required during decoding.

(Q3) The unsigned flags do not encode face orientation. Each quad is formed from the four voxels around a crossed edge in a fixed cyclic order determined by the edge axis, and is split into two triangles along its shorter diagonal. Consequently, face orientation is not guaranteed to be globally consistent. We improve orientation consistency by propagating a consistent winding over manifold edges after decoding (Table 2 of the revision). This only reverses the vertex order of some faces, leaving vertex positions and surface distances unchanged.

(Q4) If any of the four incident voxels is predicted inactive, the quad is skipped. This can create holes in the reconstructed mesh. These holes can be reduced by the hole-filling postprocessing shown in Fig. 15 of the revision.

Q6. Thresholds. No. We fixed both thresholds at 0.5 before evaluation, without tuning them on the test set. Table 4 (Table 5 in the revision) reports a post-hoc sensitivity analysis and was not used to select the thresholds.

Q8. Rate breakdown and watertight results based on tight-band SDF. We have added the coordinate/feature breakdown in Appendix C, Table 8 of the revision, which is also shown below. The G-PCC-coded latent coordinates take 14--21% of SIEVE's rate across datasets and resolutions, and the entropy-coded latent features take the rest.

Table 8: Rate breakdown of SIEVE (KB) into coordinate and feature bitstreams.

DatasetRCoordinatesFeaturesTotalCoordinate share (%)
ShapeNet1920.461.882.3419.8
ShapeNet2560.743.324.0618.2
ShapeNet3841.467.528.9816.3
ShapeNet5122.3613.3315.6815.0
ShapeNet7684.6929.6634.3513.7
ScanNet1920.762.893.6520.9
ScanNet2561.184.996.1619.1
ScanNet3842.2710.8513.1217.3
ScanNet5123.6118.8522.4716.1
ScanNet7687.1440.8347.9614.9
AOM + RWTT1920.733.213.9418.5
AOM + RWTT2561.225.696.9117.6
AOM + RWTT3842.6312.5815.2117.3
AOM + RWTT5124.3922.0126.4016.6
AOM + RWTT7688.8847.8856.7615.6
Objaverse1920.532.232.7619.1
Objaverse2560.903.884.7818.8
Objaverse3841.888.6410.5217.9
Objaverse5123.2215.1118.3317.6
Objaverse7686.6933.4640.1416.7

Regarding a tight-band SDF codec, the SIR-SNC paper shows that narrowing the band below approximately 0.75 voxel diagonals increases distortion (Sec. 4.3, Fig. 8). We therefore use its official band setting for the main comparisons. To further assess a tighter band, we retrained SIR-SNC with a band width of 0.75 voxel diagonals ( denoted as SIR-SNC*) and evaluated it on the watertight subset of AOM+RWTT (Appendix F, Fig. 10 of the revision). Compared with the original SIR-SNC, SIR-SNC* achieves similar compression performance on the test set, while SIEVE performs better than both SDF-based compression methods.

Q9. Future work. We thank the reviewer for these suggestions. Hyperprior or context entropy models and flag-weighted losses are natural extensions of SIEVE, and we will explore them in future work.

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

This paper proposes SIEVE, a mesh compression framework based on surface–grid intersections. Its DV-Edge representation stores binary edge-intersection flags and precomputed dual vertex positions on a sparse regular grid. A sparse convolutional autoencoder compresses these attributes into quantized latent features, with feature values and coordinates encoded into separate bitstreams. The decoder reconstructs occupancy, intersection flags, and vertex positions for mesh extraction. Experiments across five datasets report improved rate–distortion performance and efficient inference, particularly for open surfaces and thin structures.

Critical strengths

Relevant practical motivation. Compressing open surfaces, thin structures, and imperfect scanned meshes is an important problem. Avoiding global inside–outside classification is a reasonable design choice for these inputs.

Promising reported results and useful controls. The paper reports substantial bitrate savings across several datasets. The parameter-matched SIR-SNC-32 comparison and representation-only reconstruction experiments help separate model capacity and representation-induced distortion.

Clear accounting of compression and runtime costs. Reported bitrates include both latent features and coordinates. The runtime breakdown includes representation construction and mesh extraction, and the alternative sparse-convolution backend provides an additional efficiency control.

Critical weaknesses

The novelty of the central representation is insufficiently established. DV-Edge closely resembles existing unsigned dual-contouring representations. In particular, GenUDC already represents surfaces using grid-edge intersection flags and cellwise vertex coordinates, with vertices fitted through a QEF and connectivity determined by the intersection flags (Section 3.2, Equations 1–5). The manuscript mentions Neural Dual Contouring but does not adequately distinguish DV-Edge from its unsigned formulation or discuss GenUDC. The claimed representation contribution therefore requires more precise positioning.

The compression architecture also follows the established sparse-autoencoder and entropy-coding pipeline of SIR-SNC. Applying a related representation to compression may be useful, but the paper needs to establish the specific methodological advance beyond this adaptation.

The topology-related claims lack direct evaluation. CD, F-score, and normal consistency measure geometric similarity but do not establish preservation of connectivity, boundaries, or topology. The appendix acknowledges that decompression can introduce holes. Moreover, one intersection flag per edge and one vertex per cell impose limitations when multiple surface sheets occupy the same local grid region. Controlled experiments with known topology and explicit boundary or mesh-validity measures are needed to substantiate the claimed flexibility.

The BD-BR calculation requires clarification. In Figure 4, the ShapeNet curves for SIEVE and G-PCC appear to have no overlapping quality range for either CD or F-score, yet Table 1 reports BD-BR savings of 93.2% and 92.1%, respectively. The manuscript does not explain the integration intervals, interpolation or extrapolation procedure, or whether BD-BR is computed per mesh and then averaged. Raw operating points and calculation details are necessary to verify these substantial savings.

Some baseline comparisons conflate compression with reconstruction and remeshing choices. Draco retains the input connectivity, whereas SIEVE reconstructs a new mesh. A simplification-plus-Draco baseline would better assess lossy geometric compression under comparable objectives. Likewise, the point-cloud baselines rely on estimated normals and Poisson reconstruction, so their results partly reflect the limitations of this reconstruction pipeline on open and thin surfaces.

Important attribution controls and reproducibility details are missing. Comparisons against an existing UDC representation using the same codec, and against direct quantization and entropy coding of DV-Edge, would clarify the contribution of the representation and neural compressor. Evaluation details such as geometric normalization, CD sampling, and the F-score threshold are insufficiently specified. Runtime is reported at resolution 256, while the RD operating points span resolutions from 192 to 768.

Questions for authors

What are the precise mathematical and algorithmic differences between DV-Edge and the unsigned dual-contouring representations used in UNDC and GenUDC?

How are the BD-BR values computed when the plotted curves have little or no overlapping quality range?

How does the representation handle multiple intersections along one edge or multiple disconnected surface sheets within one cell?

How do the results change with simplification-plus-Draco and alternative reconstruction procedures for decoded point clouds?

Is hole filling used in any reported quantitative results or runtime measurements?

6 comments

We thank the reviewer for the careful reading and the constructive suggestions. Revisions are marked in red in the updated PDF.

W1/Q1/W6. Relation to UNDC and GenUDC. DV-Edge shares the local flags-and-vertices parameterization of UNDC and GenUDC, but adopts sparse storage for mesh compression. Whereas these methods target reconstruction and generation, SIEVE compresses meshes by coding both sparse support and surface attributes into a bitstream.

(i) Sparse storage. Dense UDC stores all N3N^3 cells, whereas DV-Edge stores only surface-intersecting cells, allowing finer grids with fewer stored sites and scaling to 7683768^3. In Table 4 of the revision, DV-Edge at 2563256^3 uses 127k sites versus 2097k for UDC at 1283128^3, while achieving lower surface CD (0.29 vs. 0.80). Dense UDC would change the support processed by our sparse codec; sparsifying it with identical attributes would yield the same input as DV-Edge. For convenience, we copy the relevant rows from Table 4 of the revision below.

Representation fidelity and storage comparison. CD and Surf. CD are in units of 10⁻³.

RepresentationCD↓Surf. CD↓F-score↑NC↑# Sites↓
UDC, R = 644.842.120.650.88262k
UDC, R = 1283.940.800.740.912097k
DV-Edge, R = 2563.680.290.780.92127k
DV-Edge, R = 5123.620.100.780.93530k

(ii) Compression framework. SIEVE codes both the sparse support and the flags-and-vertices attributes rate--distortion optimization. UNDC and GenUDC use this parameterization for reconstruction and generation; they do not provide a mesh codec with entropy-coded bitstreams and an explicit rate--distortion objective. We have clarified this relationship in Sec. 3.2 of the revision.

W2. Our novelty lies in the mesh compression framework. SIEVE establishes a mesh compression framework that codes both sparse support and surface attributes of a sign-free representation under learned rate--distortion optimization. We do not claim new network modules. Two comparisons substantiate its compression benefits: Fig. 7 isolates the parameterization choice relative to LSF under the same codec, while Table 1 reports 50--61% BD-BR savings in CD over the SDF-based SIR-SNC baseline. These gains are achieved with a standard sparse autoencoder and an established entropy model, without requiring more elaborate codec modules.

W3. Topology flexibility and topology-aware evaluation. By "topology-flexible", we mean that DV-Edge accommodates open surfaces and non-manifold configurations through edge intersections, without the inside/outside signs required by the standard SDF/MC pipeline used in SDF-based mesh codecs. The original manuscript evaluates reconstruction fidelity using CD, F-score, NC, and visual comparisons, following standard practice in lossy mesh compression.

To address your concerns about boundaries, connectivity, and mesh validity, we additionally report Boundary CD, component and hole counts, non-manifold-edge ratio, Hausdorff distance, and orientation consistency at matched bitrates (Table 2 in the revision, which is also shown below). It is worth noting that Boundary CD and Hausdorff distance better reflect topology preservation in lossy compression, as they capture misplaced boundaries and large local errors, while component and hole counts are neither size- nor position-aware. SIEVE achieves the lowest Boundary CD and Hausdorff distance, demonstrating that it better preserves boundaries and local structures

Table 2: Topology-aware metrics at matched rates. Draco and V-DMC are omitted because their lowest rates exceed this target rate on some test sets. We also report the topology statistics of ground truth for reference.

MethodB-CD (10⁻²)↓HD (10⁻²)↓OC (%)↑NM (%)#CC#H
GT––99.990.39330.5435.2
G-PCC6.264.38100.00.04191.343.1
SparsePCGC6.366.13100.00.02118.373.7
SIR-SNC5.995.43100.00.00162.5345.2
SIEVE4.772.7999.825.27116.7247.5
SIEVE + hole filling4.372.7899.815.24116.7218.6

Q3. Multiple edge intersections and surface sheets. DV-Edge stores one binary intersection flag per grid edge and one dual vertex per cell, following standard dual contouring.

(1) Multiple intersections per edge. If an edge is crossed several times, it is marked as intersected, and its intersection data are aggregated (positions averaged, one representative normal kept) before the QEF solve, so the edge produces a single quad.

(2) Multiple sheets per cell. If several disconnected sheets pass through one cell, their edge intersections contribute to a single QEF, collapsing the sheets to one shared dual vertex within that cell.

Both cases can be alleviated with finer grids. Table 4 shows the overall fidelity benefit: doubling the resolution from 256 to 512 reduces DV-Edge's surface CD by 66%.

W4/Q2. BD-BR calculation between SIEVE and G-PCC. For readability, each panel of Fig. 4 only shows the low-rate range, so higher-rate operating points fall outside the plotted range. We have added the full-range RD curves of all methods on all datasets in Fig. 8 and Appendix D of the revision, including the overlapping quality interval used for integration.

W5/Q4. Comparison with V-DMC. To address your suggestion of simplification followed by Draco coding, we add V-DMC, an MPEG standard incorporating mesh simplification, support for Draco base-mesh coding, and established reference configurations. As shown in Fig. 4 and Table 1 in the revision, SIEVE achieves 11--80% BD-BR savings for CD and 7--75% for F-score over V-DMC across the evaluated datasets. For your convenience, we copy the V-DMC comparison below. This comparison further demonstrates SIEVE's performance advantages.

BD-BR (%) of SIEVE relative to V-DMC, reported as CD/F-score. Negative values indicate bitrate savings.

BaselineShapeNetScanNetAOM + RWTTObjaverse
V-DMC−80.4/−74.7−11.0/−7.2−44.3/−42.7−59.2/−54.8

Q5. Use of hole filling in evaluation. In the original submission, hole filling was used only for the qualitative illustration in Fig. 13 of the supplementary material. By default, all quantitative results and runtime measurements exclude this step.

Evaluation details. Before compression, each mesh is centered at its bounding-box center and uniformly scaled into [−1,1]3[-1,1]^3. Following common practice, to compute CD and F-score, we sample 10510^5 points uniformly by area on both the ground-truth and the decompressed mesh and compute nearest-neighbor distances in both directions. CD is the L1 Chamfer distance, and the F-score uses a distance threshold of 0.0050.005. These details are now given in Sec. 4.1 of the revision. SIEVE's runtime at all RD resolutions (R=192R = 192--768768) is reported in Table 7 of the revision.

AuthorsOpenReview ↗

Looking forward to your further assessment

Dear Reviewer Njor,

Thank you for taking the time to review our manuscript and for your valuable feedback and recognition. We have carefully addressed the comments and concerns raised, as reflected in our detailed responses and the revised manuscript.

We are looking forward to your further assessment.

Best regards,
The Authors

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

This paper proposes SIEVE, a learned mesh geometry codec designed to preserve open boundaries and thin structures better than SDF-based methods. Building on the line segment field (LSF) of SALS, the authors first propose a regularly structured LSF by restricting LSF queries to grid edges and retaining only active grid locations. Then, a compact DV-Edge parameterization is derived by solving the QEF, resulting in dual vertex positions together with the intersection flags (six-dimensional). A sparse convolutional autoencoder is built for DV-Edge, including a sparse convolutional encoder, an entropy model, and a symmetric sparse convolutional decoder. Experiments on real-world scans and meshes demonstrate that SIEVE achieves better rate–distortion performance than related methods.

Critical strengths

(1) A well-motivated choice of a sign-free surface representation. The paper does not compress point clouds, which need a separate surface reconstruction step such as Poisson. It also does not compress SDFs, which depend on inside/outside sign labels. Instead, it adopts a representation built from surface–edge intersections and dual vertices. Mesh connectivity can then be recovered without sign estimation, and open surfaces can in principle be represented.

(2) Broad evaluation covering both curated assets and real-world scans. The model is trained on AMA, DT4D, and Thingi10K and evaluated on five different datasets. These include synthetic CAD models (ShapeNet), artist-created assets (Objaverse), and real scanned data (ScanNet, AOM, RWTT), so the evaluation also tests cross-dataset generalization.

(3) Careful consideration of practical runtime. The paper reports end-to-end encoding and decoding times. These include not only neural inference but also mesh-to-representation conversion, bitstream coding, and final mesh extraction (Table 5). It also gives a variant on the same sparse-convolution backend as the baseline (SIEVE*), which makes the speed comparison fairer. Encoding is several times faster than SIR-SNC (0.39 s vs. 2.51 s), and decoding speed is comparable, which makes the method practical for real deployment.

Critical weaknesses

(1) The narrative built on LSF does not hold, and the claimed derivation of DV-Edge is overstated. The paper presents SALS's LSF as a distinctive surface representation and extends it to a regularly structured LSF. It then solves the QEF in advance and derives the final DV-Edge, which contains dual vertex positions and intersection flags. It seems that the final DV-Edge retains none of the properties specific to LSF, and the preceding motivation becomes largely ineffective. Under these circumstances, DV-Edge can hardly be described as a representation "derived" from LSF. Moreover, DV-Edge closely resembles the Flexible Dual Grid (O-Voxel) of TRELLIS.2, which is also a structured grid with attributes including intersection flags and others.

(2) The incremental novelty of the network design is very limited. The codec architecture closely mirrors the sparse compression VAE of TRELLIS.2: the encoder uses sparse convolutional residual blocks with spatial-to-channel rearrangement, and the decoder uses channel-to-spatial rearrangement with eight-child occupancy prediction for pruning. The entropy model is not original either, as it is borrowed from SIR-SNC. As a result, the technical contribution reduces mainly to combining existing components.

(3) The paper's central selling point, topology flexibility, is not demonstrated quantitatively. Evaluation relies solely on CD, F-score, and NC, which are largely insensitive to topological errors such as small holes, spurious edges, or collapsed thin sheets. Topology-aware metrics, such as boundary preservation, the number of connected components, the ratio of non-manifold edges, and hole counts, are absent. Without such metrics, the reported gains cannot be attributed to the claimed topological advantage.

Questions for authors

Regarding the critical weaknesses above, I would like the authors to address the following concerns:

(1) Please clarify how DV-Edge remains connected to LSF. Since it no longer appears to retain properties specific to LSF, it instead seems highly similar to the Flexible Dual Grid (O-Voxel) of TRELLIS.2.

(2) Please clarify which parts of the proposed network architecture are new compared with the sparse compression VAE of TRELLIS.2 and with SIR-SNC.

(3) Please clarify what other evidence, beyond CD, F-score, and NC, demonstrates the claimed topology flexibility, which is the central selling point of the paper.

4 comments

We thank the reviewer for the careful review and for recognizing the motivation of our sign-free representation, the breadth of our cross-dataset evaluation, and our end-to-end runtime analysis. We address each concern below. Revisions are marked in red in the updated PDF.

W1/Q1. Connection to LSF and relation to O-Voxel.

a) LSF provides the source data for DV-Edge. We disagree with the assessments that DV-Edge can hardly be described as "derived" from LSF and that "the preceding motivation becomes largely ineffective". We believe there is a misunderstanding about the relationship between the source intersection data and the dual vertices computed by DC. In the original Dual Contouring paper (SIGGRAPH 2002, Sec. 2.2), grid-edge intersections and normals (Hermite data) are the QEF inputs, and the dual vertex is its output. SALS (ICLR 2025, Sec. 3.2) obtains these data from LSF and uses DC for mesh extraction. In our framework, DV-Edge retains LSF's intersection flags and stores the dual vertices computed from its intersection points and normals. Our ablation (Fig. 7) demonstrates the compression benefit of this choice.

b) DV-Edge and O-Voxel serve different tasks. We disagree that the resemblance to O-Voxel invalidates our motivation for the representation. O-Voxel (CVPR 2026, Sec. 3.1.1) likewise computes dual vertices from edge intersections and normals by minimizing a QEF. The resemblance to O-Voxel reflects their shared DC foundation; it does not negate DV-Edge's derivation from LSF's intersection data. By determining connectivity from intersection events rather than inside/outside signs, both support open surfaces and thin structures. Nor does this resemblance invalidate our contribution: whereas O-Voxel targets 3D generation, SIEVE develops a mesh compression framework under rate constraints. We do not claim the shared flags-and-vertices structure itself as new. We have added the discussion in Sec. 3.2 of the revision.

W2/Q2. Our novelty lies in the mesh compression framework. We do not claim novelty in the individual network modules. Our contribution is a mesh compression framework coupling DC-derived DV-Edge with learned rate--distortion optimization. Sparse convolutions, hierarchical down/upsampling, and entropy models are established components; TRELLIS.2 and SIR-SNC likewise use established components for their respective tasks. Shared modules alone do not invalidate this framework-level contribution. Fig. 7 validates the compression benefit of reducing LSF from 15 to 6 attributes, while Table 1 reports 50--61% BD-BR savings in CD over parameter-matched SIR-SNC-32.

W3/Q3. Clarifying topology flexibility and its empirical support. By "topology-flexible", we mean that DV-Edge accommodates open surfaces and non-manifold configurations through edge intersections, without the inside/outside signs required by the standard SDF/MC pipeline used in SDF-based mesh codecs. This provides a basis for compressing meshes beyond watertight surfaces. Following standard practice in lossy mesh compression, the original manuscript evaluates reconstruction fidelity using CD, F-score, NC, and visual comparisons.

SIEVE performs lossy geometry compression; its representational flexibility does not guarantee exact preservation of the input topology. To address your concerns about topology-specific diagnostics, we additionally report Boundary CD, component and hole counts, non-manifold-edge ratio, Hausdorff distance, and orientation consistency at matched bitrates (Table 2 in the revision, which is also shown below). It is worth noting that Boundary CD and Hausdorff distance better reflect topology preservation in lossy compression, as they capture misplaced boundaries and large local errors, while component and hole counts are neither size- nor position-aware. SIEVE achieves the lowest Boundary CD and Hausdorff distance, demonstrating that it better preserves boundaries and local structures

Table 2: Topology-aware metrics at matched rates. Draco and V-DMC are omitted because their lowest rates exceed this target rate on some test sets. We also report the topology statistics of ground truth for reference.

MethodB-CD (10⁻²)↓HD (10⁻²)↓OC (%)↑NM (%)#CC#H
GT––99.990.39330.5435.2
G-PCC6.264.38100.00.04191.343.1
SparsePCGC6.366.13100.00.02118.373.7
SIR-SNC5.995.43100.00.00162.5345.2
SIEVE4.772.7999.825.27116.7247.5
SIEVE + hole filling4.372.7899.815.24116.7218.6
AuthorsOpenReview ↗

Looking forward to your further assessment

Dear Reviewer TvRC,

Thank you for taking the time to review our manuscript and for your feedback. We have carefully addressed all the comments and concerns raised, as reflected in our detailed responses and the revised manuscript.

We are looking forward to your further assessment.

Best regards,

The Authors

Thank you for the detailed responses and revisions. The additional baseline comparisons, topology-aware diagnostics, and clarified evaluation details address several of my concerns and strengthen the empirical evidence.

I accept that latent compression for 3D generation and entropy-coded bitstream compression have different objectives. However, the core compression strategy appears to closely follow SC-VAE’s hierarchical encoder–decoder design. The down/upsampling strategy directly influences latent size and reconstruction fidelity, and therefore the achievable rate–distortion performance. These are central compression design choices, rather than incidental network components. Adding a bitstream-coding stage is necessary for an operational codec, but, when combined with established entropy modeling and rate–distortion training, does not by itself sufficiently establish the incremental methodological contribution.

In recognition of the substantive revisions and demonstrated practical value, I am willing to raise my rating to 2. I retain a rejection recommendation because the contribution beyond adapting the closest existing representation and compression design remains insufficiently established.

AuthorsOpenReview ↗

Clarification of SIEVE’s Contribution

Thank you for recognizing the additional evidence. However, we respectfully disagree with your assessment of our contribution and would like to clarify our position.

Our contribution is surface-intersection coding for mesh compression, an alternative to explicit-mesh and SDF-based codecs. It combines structured coding with support for open surfaces and diverse topology. We do not claim a new standalone representation or compression module.

Compact DV-Edge attributes and a lightweight autoencoder already suffice to outperform leading mesh codecs, without a more elaborate representation, network, or entropy model. This simplicity is a strength, not evidence against its contribution. Similar representations in generation, such as those in TRELLIS.2, support the geometric rationale; they do not negate our contribution to mesh compression.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for the Area Chair

Dear Area Chair,

We sincerely thank the reviewers for their time and constructive feedback. The reviews noted the well-motivated sign-free design for open surfaces and thin structures, the consistent rate–distortion gains with cross-dataset generalization across test sets, and the practical encoding and decoding efficiency of SIEVE (Reviewers TvRC, Njor, 2xtW). Reviewers Njor and 2xtW further noted the parameter-matched SIR-SNC-32 comparison as a useful control.

The main concerns raised by the reviewers were:

  • Novelty relative to prior DC-based representations and standard codec components, including DV-Edge's derivation from LSF and its relationship to existing methods (Reviewers TvRC, Njor, 2xtW).

  • Requests for topology-aware evaluation to assess topology flexibility (Reviewers TvRC, Njor, 2xtW).

  • Requests for additional mesh-codec comparisons and clarification of SIR-SNC's settings (Reviewers Njor, 2xtW).

  • Further clarification on BD-BR computation, representation details, and rate breakdown (Reviewers Njor, 2xtW).

Our responses and manuscript revisions address these concerns as follows:

  • Our core contribution is surface-intersection-based mesh compression. Unlike existing mesh codecs based on explicit meshes (Draco and V-DMC) or SDFs (SIR-SNC), SIEVE codes sparse surface intersections through DV-Edge, supporting diverse mesh topologies, including open surfaces and non-manifold configurations. With established codec components, SIEVE achieves 50–61% BD-BR savings in CD over the evaluated SDF-based baseline and 11–80% over V-DMC. Sharing a DC foundation and standard codec components does not negate this contribution.

    DV-Edge retains LSF's intersection flags and replaces its intersection positions and normals with precomputed dual vertices; Fig. 7 validates this compression-oriented choice. Sec. 3.2 of the revision clarifies this derivation and its relationship to prior DC-based representations, including O-Voxel, UNDC/GenUDC, Faithful Contouring, and NDC. Table 4 compares storage efficiency and reconstruction fidelity with dense UDC.

  • Topology-aware evaluation: We clarified that topology flexibility means representing open surfaces and diverse topology without requiring watertightness. We added Boundary CD, connected-component and hole counts, non-manifold edge ratio, orientation consistency, and Hausdorff distance at matched bitrates (Table 2). SIEVE achieves the lowest Boundary CD and Hausdorff distance, indicating better preservation of boundaries and local surface geometry.

  • Additional baseline comparisons: We added V-DMC, against which SIEVE achieves 11–80% BD-BR savings for CD (Fig. 4 and Table 1). We further added a comparison on the watertight subset with a tight-band SIR-SNC (Appendix F, Fig. 10). The tight-band and original SIR-SNC achieve similar RD performance, while SIEVE outperforms both on this subset.

  • Additional experimental and implementation details: We added full-range RD curves, BD-BR calculation details (Fig. 8 and Appendix D), and a coordinate/feature rate breakdown (Appendix C, Table 8). We also clarified voxel activation, face orientation, multi-sheet handling, and decoder thresholds.

We are encouraged by the positive feedback from the discussion. Reviewer Njor raised the rating to 3 (Weak acceptance) "based on the additional results and explanation." Reviewer TvRC stated that our additional baseline comparisons, topology-aware diagnostics, and clarified evaluation details "address several of my concerns and strengthen the empirical evidence," and raised the rating accordingly. Reviewer 2xtW did not respond during the discussion period; we have provided all experiments and clarifications requested in this review, including topology-aware metrics, the V-DMC comparison, the tight-band SDF comparison on the watertight subset, and the coordinate/feature rate breakdown.

In summary, we have addressed the reviewers' concerns through substantial new experiments and clarifications, and the remaining point raised in the discussion concerns the scope of novelty, which we have clarified in the revised manuscript. We sincerely appreciate the reviewers for their insightful comments and the Area Chair for the effort and time.

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.