acceptodds
Under review as a conference paper at ICLR 2027

CoVoGen: Coarse Voxels for Compositional 3D Scene Generation

Abstract

Compositional 3D scene generation from sparse observations requires complete geometry and coherent spatial arrangement. We introduce CoVoGen, a scene-first framework that extracts coarse object voxels from a completed shared scene to guide component refinement. Given posed RGB-D views with instance masks, a scene predictor completes coarse occupancy in world coordinates. A learned membership classifier extracts each target's coarse voxels from this shared prediction. Object structure refinement combines these voxels with neighboring occupancy and aligned observations to predict finer occupancy. A frozen pretrained geometry generator converts the refined occupancies into meshes in the shared scene. On four benchmarks comprising 1,129 scenes and 6,880 target objects, CoVoGen achieves the highest F@5 and Mean F for objects and scenes in our main comparison using the same input views, improving scene F@5 by up to 4.57 percentage points over the best baseline. These results support compositional scene generation through scene-derived coarse structure and object refinement.

Primary area:
generative models
Keywords:
compositional 3D scene generation, sparse-view reconstruction, scene completion, geometric conditioning, flow matching

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_XYrQOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CoVoGen reconstructs separate object meshes from sparse posed RGB-D views and instance masks matched across views. It first predicts shared coarse scene occupancy using a depth-conditioned adaptation of GenRecon. A learned membership classifier selects coarse voxels for each object. These voxels, neighboring geometry, and aligned observations guide local object refinement. A frozen TRELLIS.2 geometry generator then produces meshes in world coordinates. The main two-view experiments report improved object and scene F-scores, while ablations suggest that completed scene structure is particularly helpful for heavily occluded objects.

Critical strengths
  1. A useful connection between scene completion and object generation. The method combines shared scene structure with pretrained object priors to generate separate meshes in world coordinates. The visibility-based ablation is particularly informative: scene-derived conditioning improves object Mean F by 17.75 percentage points for objects with less than 25% visible surface, supporting its value under heavy occlusion.
  2. Broad evaluation with useful checks. The paper evaluates four datasets, including held-out real scenes, compares several view budgets against ShapeR, and reports paired scene-bootstrap confidence intervals. F-score evaluation includes failed outputs, and additional captured-depth measurements help assess consistency with real observations.
Critical weaknesses
  1. The key technical advance over prior work remains unclear. HIVE-3D already follows a coarse-scene generation, component extraction, and voxel-conditioned refinement pipeline. GenRecon provides the scene-generation foundation, while 3D-Fixer already performs coarse-to-fine object completion in place. CoVoGen mainly adds depth-aware conditioning, learned membership, and local spatial refinement, but the discussion and experiments do not sufficiently establish what these designs contribute beyond the closest prior work.
  2. Robustness and spatial consistency in complex scenes are insufficiently evaluated. Object extraction depends on predicted scene occupancy, and no mesh is generated when an object receives no supporting voxels. The paper provides limited analysis of recovery when small objects, thin structures, or substantial parts are missing. Although refinement uses neighboring geometry, it does not directly evaluate whether shape changes introduce intersections, lost contacts, or floating objects. A reasonable coarse layout therefore does not establish the physical plausibility of the final scene. Multi-room examples are provided, but background geometry is excluded from the final evaluation, and complex backgrounds and larger scenes lack systematic validation.
Questions for authors
  1. Please clarify the paper’s main technical innovation.
  2. How does performance vary with object size, crowding, and coarse-support recall? When valid masks and depth observations are available but the predicted support is empty or severely incomplete, is there a fallback mechanism? Can the method recover thin structures or substantially missing geometry?
  3. Refinement and mesh generation change individual object shapes. How are non-penetration, contact, and support relationships preserved? Are there before-and-after measurements of intersections, lost contacts, or floating objects?
  4. The architectural backgrounds in the multi-room examples use reference geometry. What is the current scope of the method for complex backgrounds and larger scenes? What are the main limitations and possible extensions regarding background generation, scene chunking, and computational cost?
5 comments
AuthorsOpenReview ↗

Response to Reviewer XYrQ

W1/Q1:

Our contribution is the conditioning interface from completed shared occupancy to instance-specific, spatially aligned object refinement:

  • HIVE-3D — How voxels are assigned and used. HIVE-3D assigns voxels by thresholding the generator’s attention scores aggregated over an image mask. CoVoGen predicts whether each occupied 3D cell belongs to a specified instance, using multi-view mask, depth, and scene evidence. During refinement, HIVE-3D feeds encoded voxels through a cross-attention adapter. CoVoGen combines the assigned voxels with RGB-D and mask evidence at each object-grid location and injects the resulting feature at that location.
  • GenRecon — What measured depth changes. At each projected 3D location, CoVoGen uses measured depth to distinguish free space, the observed surface, and space behind it. This distinction changes both the feature supplied by each view and the weight given to that view. Observed surfaces also initialize the scene flow. The resulting occupancy is then partitioned by instance to guide separate object refinements.
  • 3D-Fixer — Where the object’s structural support comes from. 3D-Fixer starts object completion from visible fragments cropped from an estimated point cloud. CoVoGen starts object refinement with voxels assigned to the target after completing the shared scene. Predicted occluded parts can therefore already be present in the target’s support before object refinement begins.

The hierarchy and backbones are inherited; the contribution lies in constructing and transferring these conditions.

W2/Q2:

We paired CoVoGen and ShapeR on all 1,129 scenes/6,880 targets, retaining unknown groups and all eight generation failures scored as F=0. Differences below are CoVoGen minus ShapeR in object Mean F percentage points. Size bins are within-dataset tertiles of reference-box diagonal:

DatasetSmallMediumLarge
3D-FRONT+1.48+2.03+0.01
ARSG+0.89+2.64+4.28
SAGE-2.36+1.89+1.67
ScanNet-0.77+1.54+0.73

There is no automatic fallback for empty support: no mesh is generated, affecting 4/6,613 synthetic targets. Nonempty support undergoes ordinary refinement, but missing parts or an inaccurate frame can persist.

W2/Q3:

Shared coordinates and target/neighbor occupancy provide spatial conditioning, without hard non-penetration, contact or support constraints. The two-view diagnostic covers the shared 40 scenes/211 targets/876 pairs. “Coarse” decodes membership occupancy; “final” uses saved final meshes. Both receive postprocessing, so differences include decoding/resampling effects as well as refinement.

DatasetScenes / targetsPairsCoarse crossingsFinal crossingsNew crossings
3D-FRONT24 / 901461020
ARSG10 / 78308910
SAGE1 / 294061020
ScanNet proxy5 / 1416630
All40 / 2118763580

Crossings measure triangle-surface intersections, not penetration volume. Contact is not guaranteed: at 1% of the reference-object diagonal, four of eight initially valid FRONT floor contacts remain near and four become crossing; one of seven initially near-floor SAGE targets moves below it.

W2/Q4:

We generate target-object meshes in shared world coordinates. In the rebuttal visual supplement (p. 1, Fig. Q4.1), colored objects are generated; gray walls, floors and stairs are reference geometry. Architectural RGB-D observations can condition generation, but complete architectural meshes are outside the current output scope.

Both two-storey/four-room cases retain all 32 targets at their original positions, using 364/387 RGB-D views. They demonstrate dense-view feasibility.

Overlapping chunk predictions are merged on a common lattice before instance assignment. Figure Q4.2 (p. 1) shows the vertical partition using 2.85 m cubes and 1.425 m stride:

CaseViewsRetained / candidate chunksVertical layersCore runtime (min)
136436 / 45336.15
238763 / 100438.76

Timings use one RTX PRO 5000 72 GB. Architecture generation and adaptive chunk scheduling are possible extensions.

The cited visual supplement is REBUTTAL_VISUAL_SUPPLEMENT.pdf in the updated supplementary ZIP.

Thanks for the additional experiments and explanations. They answer some of my questions. I now see this as a borderline paper, between 2 and 3.

  1. I do not find the overall approach particularly elegant, with several stages that need to be trained separately. CoVoGen is faster than ShapeR, but the gains are not enough to change my view. Since the methods use different base models and training setups, it is hard to tell how much of the improvement comes from the proposed design. Texturing has not been evaluated, and it is unclear how much extra adaptation it would require.
  2. I also remain concerned about the quality of mesh. The real-scene example looks better than ShapeR, but still has visible fragments and other shape errors. Limited observations and a gap between the training data and real scenes may partly explain this, but producing reliable geometry under these conditions is an important part of the task.
  3. Some of the limitations discussed earlier also remain.
AuthorsOpenReview ↗

Follow-up on attribution, texturing, and geometry quality

Thank you for the reassessment and for clarifying the remaining concerns.

Modularity and contribution. The scene stage provides structural support for occluded objects, while local refinement combines this support with spatially aligned observations and pretrained object priors. Our contribution is this conditioning interface between completed scene occupancy and instance-specific refinement. The modular formulation connects global scene completion with local object priors while keeping the final geometry generator frozen. Efficiency is an additional practical benefit of this design.

Evidence under a fixed model. To examine the contribution of scene-derived target support with the backbone fixed, we use a shared-checkpoint ablation on 40 scenes/211 targets. Replacing visible-depth target support with completed-scene target support improves final Mean F from 79.91% to 86.05%, with the same neighboring support, trained models, two RGB-D inputs, initialization, sampling and postprocessing. This provides diagnostic evidence for the value of scene-derived structure within the same trained model, complementing the system-level comparison with ShapeR.

Texturing without additional training. We have additionally conducted a qualitative texturing test using frozen TRELLIS.2 on both two-storey examples, covering 64 objects. It operates on the existing postprocessed meshes using one observed image crop per object, with no additional training or fine-tuning. Figures T1–T2 in TEXTURING_SUPPLEMENT.pdf, included in the updated supplementary ZIP, show the full-scene views and local texture details. These retain the original dense-view geometry settings and object positions; gray architecture is reference geometry. The examples demonstrate integration without learned adaptation.

The real-scene comparison shows improved scene structure under matched observations, with fine-surface fidelity remaining an important direction for further improvement.

Reviewer_2DSKOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces CoVoGen, an instance-centric method to generate 3D scenes made of aligned object meshes from input RGB-D views and instance masks. The generation happens in two stages. First, the partial input geometry of the scene is refined globally. To do that, the coarse scene occupancy coming from depth frames is obtained by back projection and voxelization and then refined using the pretrained backbone from GenRecon at a resolution of 64^3. Then, a membership classifier predicts the instance segmentation (grounded to the original instance masks) of the refined scene occupancy. Single objects occupancies are then first refined by using a Lora adaptation of the first stage of Trellis2 and meshes are extracted using a frozen Trellis2 second stage. The method outperforms several relevant object-centric 3D scene reconstruction baselines both on synthetic (3D-FRONT, ARSG, SAGE) and real-world (ScanNet) datasets in terms of geometric metrics as Chamfer Distance and F-score.

Critical strengths

Soundness of the idea and improvement over baselines. The idea of relying on a coarse global refinement followed by a detailed local step is interesting and allows to reduce the complexity of 3D scene generation while still obtaining high quality instance mesh predictions leveraging Trellis2’s prior.

Critical weaknesses
  • Limited method novelty and unclear explanation. While the idea of combining a coarse global stage of generation with a fine-grained local one is very interesting, the paper does not propose an ad-hoc end-to-end method to solve the problem, but proposes a pipeline made of existing components (GenRecon for global scene refinement and Trellis2 for object-level refinement) which results in modest improvements over the best baseline (ShapeR). Also, the presentation of the method in the main paper is not clear, specifically for the object-level refinement, where only vague details are provided. Reading Appendix B.6 and C.1 becomes necessary to understand this missing part.
  • Missing qualitative results on real-world datasets. I believe this evaluation is crucial to understand the generalization of the method on real scenes as the provided numbers depends on the quality of the annotations of 3D ground truths coming from MetaScenes.
Questions for authors
  • Would it be possible to include more details on the main paper about the spatial encoder mentioned in L216-217 and about how depth, image features and instance masks are combined and processed by the spatial encoder and injected in the structure flow (L220-226)?
  • Have you tried removing the single-view conditioning via cross-attention? Could you explain why it is needed in addition to the aggregated multi-view features?
  • Is it possible to provide qualitative evaluation on real-world scenes (e.g. on ScanNet)?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 2DSK

W1:

Our contribution is transferring completed shared occupancy into instance-specific, observation-aligned conditions for object refinement. We build on GenRecon and TRELLIS.2, adding measured-depth scene conditioning, learned membership and local fusion of support with RGB-D/mask evidence. The stages are trained separately; the final geometry generator is frozen. We do not claim the coarse-to-fine hierarchy itself as new.

HIVE-3D assigns voxels through thresholded image-mask attention and uses voxel cross-attention for refinement, whereas CoVoGen learns instance membership from multi-view masks, depth, and scene features, then fuses support with observations at corresponding 3D locations. Building on GenRecon, CoVoGen adds measured-depth corrections to both view features and aggregation weights, and initializes the scene flow from observed surfaces. While 3D-Fixer completes visible point-cloud fragments, CoVoGen refines objects using support extracted from a completed shared scene, which can already contain predicted occluded parts.

Q1:

The spatial encoder operates at each cell of the object-local 16³ latent grid:

  1. Position, cell size and target/context occupancy indicators form a query.
  2. Its projected footprint samples frozen DINOv3 features, camera-Z depth and target masks on a 3 × 3 stencil in each selected full-scene view.
  3. Projected image features are added to an MLP embedding of geometric/mask descriptors, including depth offsets, ray-surface relation and depth validity. Missing depth is explicitly encoded.
  4. Query-conditioned attention aggregates samples within and across views into g(u)g(u). After each flow Transformer block, a zero-initialized affine map injects this condition at the same cell:

Hn+1(u)=Bn(Hn,t,c)(u)+An(g(u)).H_{n+1}(u)=B_n(H_n,t,c)(u)+A_n(g(u)).

The block retains primary-view crop cross-attention; observed target depth separately initializes the flow. Thus, spatial fusion, crop conditioning and initialization are distinct paths.

Q2:

We disable all 30 crop cross-attention residuals in Refine, including LoRA contributions, with trained weights, inputs, frames, noise and decoding fixed. The primary view remains in the spatial encoder, and the final geometry generator receives the same crop in both settings.

The shared 40-scene/211-target cohort uses up to eight selected views per target. Final Mean F (%) is:

DatasetScenes / targetsFullNo crop CAGain (pp)
All40 / 21189.5587.69+1.86
3D-FRONT24 / 9088.8286.40+2.42
ARSG10 / 7895.1093.57+1.53
SAGE1 / 2991.3788.86+2.52
ScanNet proxy5 / 1459.5960.79-1.20

Crop attention supplies object-centered appearance tokens, while spatial conditioning aggregates observations at aligned 3D locations. The overall gain supports this complementary path under the current weights.

W2/Q3:

The rebuttal visual supplement (p. 2, Fig. R2.1) shows ScanNet, retaining all five targets. Both methods use the same eight RGB-D inputs and matched display cameras/scale; ShapeR* uses masked RGB, while CoVoGen uses direct RGB. CoVoGen preserves the seating-and-table layout and avoids ShapeR*'s large extraneous structures, although local chair defects remain.

The cited visual supplement is REBUTTAL_VISUAL_SUPPLEMENT.pdf in the updated supplementary ZIP.

I thank the authors for the additional experiments and clarifications. However, I still have two main concerns:

  1. Technical contribution. The proposed method combines GenRecon and Trellis2, which are trained separately and integrated for a common downstream task. While I appreciate the effectiveness of this combination, I remain unconvinced that it constitutes a sufficiently novel technical contribution.

  2. Performance on real-world scenes. I appreciate the additional experiment on ScanNet, although results on a single scene provide limited evidence of real-world performance. Moreover, the reconstruction quality appears rather poor across all methods, and I would appreciate further clarification on the experimental setup and results. More specifically:

    • How are the depth maps and instance masks obtained?
    • Why does the proposed method produce so many reconstruction artifacts?
    • Why does ShapeR appear to fail almost entirely at reconstructing the scene?

Thank you!

AuthorsOpenReview ↗

Follow-up on technical contribution and ScanNet reconstruction

Thank you for clarifying your concerns. We explain below what CoVoGen adds to prior methods and what the ScanNet example reveals about the two pipelines.

Technical contribution. CoVoGen addresses how to turn a completed scene into spatial guidance for refining individual objects. GenRecon provides a shared scene prediction. The next step requires deciding which target owns each predicted voxel, including voxels on surfaces that neither input view observes, and where that prediction should influence object refinement. HIVE-3D uses thresholded image-mask attention to assign coarse voxels and voxel cross-attention to condition refinement. CoVoGen learns membership directly on the completed 3D occupancy using multi-view masks, depth, and scene features. It then fuses the assigned support with observations at corresponding locations on the object’s refinement grid. This lets inferred geometry in occluded regions guide the relevant object, while measurements provide local evidence where surfaces are visible. The technical contribution lies in these two operations at the interface between the separately trained stages.

1. How were the depth maps and masks obtained? We used ScanNet’s captured sensor depth and official filtered 2D instance annotations. We converted the depth to meters and aligned the color images and masks to the depth grid using ScanNet calibration. The illustrated scene uses eight shared RGB-D frames and includes all five annotated targets. Our formal two-view ScanNet evaluation covers 24 scenes and 267 targets, using MetaScenes meshes as approximate references. Page 1 of SCANNET_DIAGNOSTICS.pdf shows the actual inputs.

2. Why does CoVoGen produce these artifacts? The end table illustrates two input problems. First, parts of the tabletop lack depth measurements. In one view, about 22% of the pixels inside the table mask have no valid depth. Second, the mask includes part of a neighboring chair arm. Those pixels have valid depth, so back-projection assigns chair geometry to the table. In the frame shown on p. 2, 1,147 points pass the 1st–99th percentile checks for camera depth and horizontal coordinates, yet lie at least 12 cm horizontally from the approximate table bounding box. The RGB crop and top-down point cloud show where those points come from and where they land. CoVoGen uses masks and depth to assign scene voxels to objects and to condition refinement, so missing readings reduce evidence for the tabletop and misassigned readings supply evidence outside the table. The matched views on p. 2 show stray structure already present in the refined voxels and fragments remaining in the final mesh.

3. Why does ShapeR reconstruct the scene poorly? ShapeR receives the same imperfect observations described above. In our ScanNet adapter, the full masked point cloud also determines each instance’s center and scale. The table’s input cloud spans 2.37 m horizontally. The isolated predictions on p. 3 show that the excess geometry is already present at the object level and overlaps neighboring objects after assembly. To test the influence of image background, we whitened pixels outside the instance mask while holding the masks, depth points, cameras, normalization, and sampling fixed. This reduced the chair-like geometry in the table prediction and shrank its projected area from 21,257 to 8,087 pixels. The resulting mesh still spans 2.32 m horizontally, compared with 0.47 m for the approximate reference. The control establishes an effect of image background. The contributions of point-cloud contamination and normalization to the remaining errors still require separate controls.

As the discussion period draws to a close, we wanted to check whether our follow-up clarified the technical contribution and the ScanNet results. We’d be happy to clarify either point further if helpful.

Reviewer_q6PZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CoVoGen, a method for compositional 3D scene generation from sparse RGB-D observations. CoVoGen first completes a shared scene occupancy representation, extracts coarse object voxels, and then refines the structure of each object. The method is clearly motivated and evaluated on multiple synthetic and real-world datasets. In particular, the low-visibility experiments show that scene-derived coarse voxels substantially improve the generation quality of heavily occluded objects, providing strong support for the paper’s central claim.

Critical strengths
  1. The scene-first design is well motivated and effectively leverages shared scene structure to support object completion.
  2. The experiments cover 1,129 scenes and 6,880 objects across multiple benchmarks, providing a relatively comprehensive evaluation.
  3. The controlled ablations, paired bootstrap analysis, and oracle experiments provide substantial diagnostic evidence.
  4. The paper is clearly organized and provides detailed descriptions of its training and evaluation configurations.
Critical weaknesses
  1. CoVoGen and competing methods such as ShapeR use different geometry generators and training conditions. Consequently, the improvements in the main comparison tables cannot be attributed entirely to scene-derived conditioning. A stronger comparison would include an independent object-completion baseline using the same TRELLIS.2 generator, observation inputs, and training data as CoVoGen.
  2. Learned membership is primarily compared with a bounding-box selection rule. Stronger geometry-based baselines, such as mask- and depth-guided propagation, would better establish the necessity of the learned classifier. In addition, neighboring occupancy is presented as an important condition for object refinement, but its independent contribution is not ablated.
  3. Figure 5 uses 11 or 30 input views, whereas the main quantitative setting uses only two views. Qualitative comparisons under the same two-view budget, together with representative failure cases, would make the evaluation more consistent and informative.
Questions for authors

Please see the weakness.

1 comment
AuthorsOpenReview ↗

Response to Reviewer q6PZ

W1/Q1:

We vary explicit structural support at inference using the same trained Refine checkpoint and frozen TRELLIS.2 generator. The two RGB-D inputs, original support-derived frames, image conditioning, initialization, sampling and postprocessing are fixed:

  1. RCN — completed target + neighbors. Learned membership extracts target cells from completed scene occupancy; surrounding cells provide context.
  2. RVN — visible target + the same neighbors. Target cells are replaced by the voxelized, deduplicated union of target depth back-projected from both inputs. Neighbor cells are unchanged.
  3. RN0 — no explicit target or neighbor support. Both cell sets are emptied. The observation encoder, projected RGB-D/mask evidence, image conditioning and visible-depth initialization remain active.

Final Mean F is averaged over all 211 targets in 40 scenes:

SettingExplicit target supportNeighbor supportFinal Mean F (%)
Ours (RCN)From the completed sceneRetained86.05
Visible-depth support (RVN)From observed target depthSame as RCN79.91
No explicit support (RN0)RemovedRemoved84.16

W2/Q2:

We compare four extraction rules on the same predicted scene occupancy, using frozen Refine and TRELLIS.2 models. The three geometric controls use no reference geometry:

  1. Box. Back-project valid target-masked depth from the inputs, merge the points in world coordinates, and expand their axis-aligned bounding box by one voxel on each side. Select occupied cells whose centers lie inside.
  2. Mask. Project each occupied cell center into every input view. Each in-front, in-image projection votes positive inside the target mask and negative otherwise, including views where the target is absent. Select cells with at least one positive vote and at least 50% positive valid votes. No depth-based visibility filtering is applied.
  3. Depth. Attach observed, instance-labeled depth points to their nearest occupied cells within 1.5 voxel widths. Propagate labels over the occupied 26-neighbor graph using shortest paths weighted by Euclidean cell-center distances, including seed-attachment distances. Cells seeded exclusively by another instance block propagation. In voxel units, assign a cell to target kk if its finite cost satisfies dk≤min⁡jdj+1d_k \leq \min_j d_j + 1, over all observed labels jj. Similar costs can produce overlapping assignments; unseeded components remain unassigned, and propagation cannot cross missing cells.

On the shared 40 scenes/211 targets, the protocol selects up to eight views per target. Values are percentages; IoU is surface-cell IoU on a common world grid.

MethodExtracted Mean FExtracted IoURefine Mean FRefine IoUFinal Mean FFinal IoU
Learned62.0936.3478.1552.6789.6258.24
Box60.5335.8678.1352.6589.3958.18
Mask61.8034.3871.0545.6986.3451.00
Depth58.3534.5874.2549.5686.7955.41

Neighbor-only ablation. On 30 ARSG scenes/243 targets with two inputs, fixing target support and removing neighbors changes Mean F from 87.47 to 87.28. The unrounded gain is 0.20 points, with paired 95% CI [-0.02, 0.41]. This does not establish a reliable independent benefit; we treat neighboring occupancy as auxiliary.

W3/Q3:

The rebuttal visual supplement (p. 2, Fig. W3.1) compares CoVoGen and ShapeR using the same two raw inputs, native preprocessing, and matched display cameras and scale. CoVoGen better preserves cabinet surfaces and handles in (a); thin-structure errors remain in (b).

Figure W3.2 (p. 2) shows two failures: spurious wardrobe surface geometry and bedcover tears. Broad structure is present at Refine, but additional surface artifacts appear during TRELLIS.2 geometry generation and persist after postprocessing. Coherent structural support therefore does not guarantee faithful fine geometry.

The cited visual supplement is REBUTTAL_VISUAL_SUPPLEMENT.pdf in the updated supplementary ZIP.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.