- 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
-
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.
-
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
-
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.
-
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
- (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?
- (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?
- (Clarification) With unsigned flags and no normals, how are face orientation and quad triangulation determined?
- (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?
- (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?
- (Clarification) Were the thresholds τᵤ and τₑ selected on the test set?
- (Experiment) Please report topology metrics (hole/boundary-edge count, non-manifold edges, orientation consistency) and Hausdorff distance.
- (Experiment) Please provide a rate breakdown (coordinate vs. feature bits), and compare against a tight-band SDF codec on a watertight subset.
- (Future work) Consider stronger entropy models (hyperprior or context models), and loss weighting that prioritizes flag accuracy over vertex positions.
7 comments
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.
| Method | B-CD (10⁻²)↓ | HD (10⁻²)↓ | OC (%)↑ | NM (%) | #CC | #H |
|---|---|---|---|---|---|---|
| GT | – | – | 99.99 | 0.39 | 330.5 | 435.2 |
| G-PCC | 6.26 | 4.38 | 100.0 | 0.04 | 191.3 | 43.1 |
| SparsePCGC | 6.36 | 6.13 | 100.0 | 0.02 | 118.3 | 73.7 |
| SIR-SNC | 5.99 | 5.43 | 100.0 | 0.00 | 162.5 | 345.2 |
| SIEVE | 4.77 | 2.79 | 99.82 | 5.27 | 116.7 | 247.5 |
| SIEVE + hole filling | 4.37 | 2.78 | 99.81 | 5.24 | 116.7 | 218.6 |
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. is stored in local coordinates normalized to 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.
| Dataset | R | Coordinates | Features | Total | Coordinate share (%) |
|---|---|---|---|---|---|
| ShapeNet | 192 | 0.46 | 1.88 | 2.34 | 19.8 |
| ShapeNet | 256 | 0.74 | 3.32 | 4.06 | 18.2 |
| ShapeNet | 384 | 1.46 | 7.52 | 8.98 | 16.3 |
| ShapeNet | 512 | 2.36 | 13.33 | 15.68 | 15.0 |
| ShapeNet | 768 | 4.69 | 29.66 | 34.35 | 13.7 |
| ScanNet | 192 | 0.76 | 2.89 | 3.65 | 20.9 |
| ScanNet | 256 | 1.18 | 4.99 | 6.16 | 19.1 |
| ScanNet | 384 | 2.27 | 10.85 | 13.12 | 17.3 |
| ScanNet | 512 | 3.61 | 18.85 | 22.47 | 16.1 |
| ScanNet | 768 | 7.14 | 40.83 | 47.96 | 14.9 |
| AOM + RWTT | 192 | 0.73 | 3.21 | 3.94 | 18.5 |
| AOM + RWTT | 256 | 1.22 | 5.69 | 6.91 | 17.6 |
| AOM + RWTT | 384 | 2.63 | 12.58 | 15.21 | 17.3 |
| AOM + RWTT | 512 | 4.39 | 22.01 | 26.40 | 16.6 |
| AOM + RWTT | 768 | 8.88 | 47.88 | 56.76 | 15.6 |
| Objaverse | 192 | 0.53 | 2.23 | 2.76 | 19.1 |
| Objaverse | 256 | 0.90 | 3.88 | 4.78 | 18.8 |
| Objaverse | 384 | 1.88 | 8.64 | 10.52 | 17.9 |
| Objaverse | 512 | 3.22 | 15.11 | 18.33 | 17.6 |
| Objaverse | 768 | 6.69 | 33.46 | 40.14 | 16.7 |
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.
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.
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 , 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× site count in Table 4 of our paper reflects this comparison under SIR-SNC's official settings.