acceptodds
Under review as a conference paper at ICLR 2027

Recovering Sub-Voxel Geometry: Learned Implicit Quadrics for Large-Scale 3D Scenes

Abstract

Large-scale 3D scene representations based on discrete voxelization inherently discard sub-voxel surface structure, limiting geometric fidelity and the accuracy of downstream physics-aware sensor simulation. We introduce QuadricUNet, a neural architecture that represents scene surfaces using compact, continuous second-order polynomial surfaces (quadrics) instead of discrete voxels. Given only a sparse semantic occupancy grid, QuadricUNet predicts the 10 coefficients of a quadric surface for each occupied voxel, without requiring dense point clouds, surface normals, or explicit meshes as input. Because each surface is a low-degree implicit polynomial, geometric operations act directly on the predicted coefficients. We develop a vectorized Newton-Raphson projection for dense surface sampling that is more than 11× faster than Marching Cubes sampling the same field, and an analytical ray-quadric intersection for LiDAR simulation that eliminates the need for mesh extraction. On large-scale outdoor scenes, QuadricUNet achieves a Chamfer distance of 4.91 cm and 90.25% F1@10 cm on SemanticKITTI, and 4.93 cm and 93.37% F1@10 cm on the Waymo Open Dataset, improving on Marching Cubes and NKSR when these baselines operate on the same voxel-only input. For LiDAR simulation, the analytical representation achieves a 0.0814 m depth MAE, and an SPVCNN segmentation model trained on its simulated sweeps reaches 0.48 macro-mIoU on real sweeps, compared with 0.37 and 0.33 for sweeps simulated from Marching Cubes and NKSR meshes (0.64 when trained on real sweeps). These results show that continuous quadric-based geometry supports both higher-fidelity reconstruction and sensor simulation from sparse scene representations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Deep Learning, LiDAR Simulation, Point Clouds, Sim-to-Real Gap, Autonomous Perception, 3D Scene Reconstruction

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_AC4rOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the loss of sub-voxel geometric information in discrete semantic occupancy representations. The authors propose QuadricUNet, which predicts local implicit quadric surfaces from semantic voxel grids, enabling continuous surface reconstruction and analytical LiDAR ray intersection. Experiments on SemanticKITTI and Waymo demonstrate improved geometric reconstruction over the evaluated baselines, while simulated LiDAR sweeps yield better sim-to-real segmentation performance.

Critical strengths
  1. The paper addresses the loss of geometric detail caused by coarse voxelization, which is particularly relevant to 3D scene reconstruction and physics-aware sensor simulation. Representing sub-voxel surfaces using learned quadrics provides a meaningful connection between data-driven geometric inference and analytical operations such as ray intersection.

  2. The method achieves substantial improvements over the evaluated baselines on SemanticKITTI and Waymo. Importantly, the improvements extend beyond geometric metrics to sim-to-real LiDAR segmentation, where the macro-mIoU increases from 0.3672 with Marching Cubes to 0.4806 with QuadricUNet, providing evidence of practical downstream benefits.

Critical weaknesses
  1. The main baselines are Marching Cubes and NKSR, neither of which is trained specifically to recover sub-voxel geometry from the given semantic occupancy representation. However, existing approaches such as Neural Dual Contouring can learn surface geometry directly from binary voxel grids without requiring input surface normals. Although their output representations differ, the absence of such learning-based comparisons makes it difficult to establish whether the reported reconstruction gains arise specifically from the proposed quadric representation or more generally from learned geometric priors.

  2. Appendix A states that the input and supervised voxel sets are restricted to voxels for which reference quadrics can be fitted from aggregated LiDAR observations. This reference-dependent filtering introduces an additional condition into the evaluation inputs, while all reported experiments use ground-truth rather than predicted occupancy grids. Consequently, the results demonstrate effectiveness under an idealized and filtered occupancy setting, but do not yet establish robustness to independently generated or imperfect occupancy representations.

Questions for authors
  1. Could the authors provide comparisons with learning-based voxel-to-surface methods, particularly Neural Dual Contouring or a voxel-conditioned implicit surface decoder, under comparable input and supervision settings? A learned planar representation would also help isolate the contribution of second-order geometry.

  2. Could the authors clarify how many occupied voxels are excluded by the reference-quadric fitting requirement in Appendix A, and whether this filtering uses geometric information unavailable in a realistic inference setting? Results using occupancy inputs constructed independently of the reference LiDAR geometry would help establish the generality of the voxel-only claim.

  3. Table 6 reports a substantial speedup in surface sampling from the same fitted quadric field, but excludes network inference and shows higher memory consumption. Could the authors provide end-to-end computational comparisons, including geometry prediction and LiDAR simulation, to better establish the practical efficiency advantages of the complete framework?

4 comments
AuthorsOpenReview ↗

Response to Reviewer AC4r [1/2]

Thank you for the detailed and valuable feedback.

W1 and Q1. Could the authors provide comparisons with learned baselines under comparable input and supervision settings?

We added two learned models with our exact input. QuadricUNet (SDF head) is the voxel-conditioned implicit decoder as suggested. It keeps our U-Net and replaces the quadric head with a query MLP. UniScene is a published occupancy-to-LiDAR model [1], trained on our data.

MethodKITTI Chamfer / F1@10LiDAR per-ray error / F1@10
QuadricUNet4.97 ± 0.04 cm / 87.8 ± 0.4%14.8 ± 0.2 cm / 39.6 ± 0.5%
QuadricUNet (SDF head)5.01 cm / 81.8%22.4 cm / 27.8%
UniScene5.47 cm / 81.7%22.0 cm / 35.9%

This comparison separates the two factors. In Table 1, the SDF head is on par with QuadricUNet in Chamfer distance but has a clearly lower F1@10. It predicts surfaces in most of the voxel volume which lowers the Chamfer distance (good completeness) but also ends up predicting incorrect surfaces (low accuracy), which F1 penalises. Learning-based setups overall have an edge in reconstruction performance. Our representation has the edge on accuracy and LiDAR results back this: on the same backbone, the query decoder's LiDAR error is much higher. Rendering QuadricUNet through UniScene's slower volume renderer gives the same per-ray error as our analytical intersection (14.6 cm with both), so the gap is not due to the renderer.

We thank the reviewer for pointing to Neural Dual Contouring [2]. QuadricUNet (SDF head) addresses the same question and is trained on the same reference points. We do not train a learned planar representation, as fitted planes reach 4.53 cm Chamfer and 61.8% F1@5 (Table 1), against 3.92 cm and 66.9% for fitted quadrics. This shows that the representational capacity of planes is already lower than quadrics for our setting.

W2 and Q2. Could the authors clarify the effect of filtering and the results in a realistic inference setting? Results using occupancy inputs constructed independently of the reference LiDAR geometry would help establish the generality of the voxel-only claim.

The filter was used to clean up noisy voxels during preprocessing (10.5% of the static voxels on sequence 08, which hold 0.011% of the reference points). It therefore uses LiDAR points, but only at dataset creation. We tested and found that a model trained without it performs similarly (4.98 cm Chamfer without the filter against 4.97 ± 0.04 cm with it), as the removed voxels hold almost no reference points. Inference requires only the semantic voxel grid. We agree that all results are uniformly evaluated on ground-truth grids. We avoid evaluation on predicted occupancy grids for two reasons. First, the ground-truth grids are built from the LiDAR-mapped surfaces that serve as our reference, whereas a predicted grid need not align with the reference, so the reference cannot score it meaningfully. Second, ground-truth grids keep the errors of upstream occupancy prediction models out of our evaluation. Predicted grids therefore need their own evaluation protocol, which we leave to future work.

The end-to-end cost comparison (Q3) follows in the next comment.

References

[1] B. Li et al., "UniScene: Unified Occupancy-centric Driving Scene Generation," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025.

[2] Z. Chen et al., "Neural Dual Contouring," ACM Transactions on Graphics (TOG), vol. 41, no. 4, 2022.

AuthorsOpenReview ↗

Response to Reviewer AC4r [2/2]

Q3. Could the authors provide an end-to-end cost comparison?

We revised the cost tables to time every method end-to-end on the GPU.

Table 1. Dense sampling, per frame. Means over 97 frames of sequence 08 (SemanticKITTI) on one NVIDIA H200. NKSR's network time includes its normal estimation and solver.

MethodNetwork (ms)Field queries (ms)Mesh extraction (ms)Surface sampling (ms)End to end (ms)GPU memory (GB)
Reference quadrics (upper bound)–––77.377.31.3
QuadricUNet25.9––79.0104.91.5
QuadricUNet (SDF head)25.8182.743.818.9271.18.1
UniScene11.9219.436.016.4283.76.5
NKSR1964.8–396.814.32375.93.4
Marching Cubes––1.610.612.11.0

Table 2. One LiDAR sweep, per frame.

MethodNetwork (ms)Mesh extraction and build (ms)Ray casting or rendering (ms)End to end (ms)GPU memory (GB)
Reference quadrics (upper bound)––24.024.00.5
QuadricUNet25.5–23.148.61.4
QuadricUNet (SDF head), volume rendering25.6–35.461.01.9
QuadricUNet (SDF head), sphere tracing26.2–1799.61825.91.4
UniScene, volume rendering10.6–39.650.22.2
UniScene, sphere tracing10.7–951.1961.80.9
NKSR1964.8401.739.42405.96.5
Marching Cubes–2.423.826.22.2

Table 3. Training and storage.

MethodTraining (min)Stored per frame (MB)
QuadricUNet924.0, 10 numbers per voxel
QuadricUNet (SDF head)129256, dense volume
UniScene12512.7 sparse, 32 numbers per voxel; 256 dense
NKSRn/a50.9, mesh
  1. Dense sampling: QuadricUNet is the fastest learned method and uses the least GPU memory of all learned baselines. It is more than twice as fast as both learned baselines and more than an order of magnitude faster than NKSR. Its sampling step costs about as much as sampling the reference quadrics, so the predicted coefficients add no overhead. Marching Cubes is faster because it involves no learned model, but its surfaces are far less accurate.
  2. One LiDAR sweep: QuadricUNet matches the speed of UniScene's volume rendering at a clearly lower error, and is far faster than sphere tracing either implicit model. Its ray casting is the cheapest step of all learned methods. Marching Cubes is faster because its sweep is a mesh ray cast, but its error is several times higher.
  3. Same U-Net with a query head: QuadricUNet (SDF head) has the same network cost, so the whole gap is the head. It is slower on both outputs and needs several times more GPU memory for dense sampling and far more storage. The quadric head, not the U-Net, is what makes QuadricUNet cheap.
  4. Storage: QuadricUNet stores fewer numbers per voxel than any learned baseline. Its output is several times smaller than UniScene's sparse features and an order of magnitude smaller than NKSR's mesh.

The savings come from needing no network at query time.

Thank you for the detailed rebuttal and additional experiments. The new learning-based baselines and end-to-end efficiency comparisons adequately address my main concerns and strengthen the paper. I appreciate the authors' efforts and maintain my positive recommendation (3: Weak Acceptance).

AuthorsOpenReview ↗

Response to Reviewer AC4r (follow-up)

We thank the reviewer for their response and for maintaining the positive score. We are glad we could address your concerns. Your constructive feedback helped strengthen our submission.

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

This paper proposes QuadricUNet, a method that takes as input a set of occupancy voxels with semantic labels, and outputs quadrics that implicitly represent the underlying surface. This prediction is performed by a 3D UNet.

Critical strengths
  • The setting and method are well described and illustrated. The paper is clear.
  • The method is evaluated in depth, with qualitative, quantitative, and ablation results.
Critical weaknesses
  • The setting (predicting surfaces from sparse voxels with semantic labels) appears quite artificial, or at least it is not very well motivated. In most settings where one would have an occupancy grid, one would also have the RGB images that were used to generate it, and potentially depth. It therefore seems a bit convoluted to assume a setting where this grid is available, but no image or any other modality. Similarly, this seems like a very indirect way to generate LiDAR training data.
  • The proposed method is very straightforward, with a 3D UNet converting the occupancy grid into per-voxel quadric coefficients.
  • The produced surfaces are never visualized in the paper or supplementary.
  • The two baselines used, Marching Cubes and NKSR, are not entirely meaningful. Their input is a completely different modality. For example, the authors have converted the voxel grid into a point-cloud by sampling a point at the center of each voxel. While there may be no better direct alternative methods in this setting, the authors could at least compare alternative architectures.
Questions for authors
  • Will the boundaries match across different voxels since their quadrics are independent?
4 comments
AuthorsOpenReview ↗

Response to Reviewer XTwW [1/2]

Thank you for the review. We have carefully addressed your concerns below.

W1. The setting is artificial, semantic grids are available alongside images or other modalities, and the method is an indirect way to generate LiDAR training data.

Many recent works produce semantic occupancy without images or depth. For example, generative models that synthesize scenes directly as semantic voxels (SemCity [1], XCube [2], DynamicCity [3]), and world models that predict occupancy (OccWorld [4]). There are many works that terminate at voxels as the final output. Our work positions itself as a method to improve upon these voxel outputs and bring them geometrically closer to LiDAR-mapped surfaces (lowering the sim-to-real gap) that would find greater utility in downstream applications, as we have demonstrated. Where images exist, occupancy predicted from them can be used without change. Generating LiDAR training data from semantic occupancy is indirect, but it is an established route. UniScene [5] first generates occupancy and then renders LiDAR from it. We also primarily evaluate against LiDAR because it is the most geometrically accurate data available for outdoor scenes, so a comparison with real scans tests the recovered surfaces directly.

W2. The proposed method is very straightforward.

Our primary contribution is in motivating the representation and its benefits. We show this in the representation's efficiency in memory, surface point sampling, lidar simulation, reconstruction performance, and we also quantified sim-2-real gap with segmentation performance. We have also added more competitive baselines to our evaluation and the simple 3D UNet matches or outperforms them on every task. We believe that the simplicity of the neural architecture is an advantage for implementation rather than a critical weakness.

W3. Surfaces are never visualized.

We agree that our work would benefit from added visualizations. The initial submission shows the predicted surfaces in two forms. The densely sampled surfaces on SemanticKITTI and LiDAR sweeps (simulated and real). In the revised PDF, we include meshes and wireframes to aid surface visualization. Figure 2 now shows meshes of all five methods, a street view and a close-up are included. The close-up of the mesh wireframe is added to the appendix as well, with the remaining figures.

W4 and the question on voxel boundaries are in the next comment.

AuthorsOpenReview ↗

Response to Reviewer XTwW [2/2]

W4. The two baselines used, Marching Cubes and NKSR, are not entirely meaningful.

We agree and have acknowledged the shortcomings of Marching Cubes [6] and NKSR [7] in our initial submission. We therefore add two learned models that take exactly our input and are trained on the same data: UniScene [5] and QuadricUNet (SDF head). QuadricUNet remains the most accurate method. QuadricUNet (SDF head) matches its Chamfer distance on SemanticKITTI, but its F1@10 is lower and its LiDAR error higher. This also isolates the effect of the backbone and representation.

MethodKITTI Chamfer / F1@10Waymo Chamfer / F1@10LiDAR per-ray errorSegmentation performance (SPVCNN mIoU)
QuadricUNet4.97 ± 0.04 cm / 87.8 ± 0.4%5.43 cm / 84.1%14.8 ± 0.2 cm34.7 ± 0.7%
UniScene5.47 cm / 81.7%5.89 cm / 77.1%22.0 cm29.8 ± 1.2%
QuadricUNet (SDF head)5.01 cm / 81.8%n/a22.4 cmn/a
NKSR8.47 cm / 65.2%8.92 cm / 63.2%18.0 cm27.6%
Marching Cubes12.99 cm / 30.5%11.78 cm / 34.0%57.3 cm25.5%

Q. Do the surfaces match across voxel boundaries?

While the predicted quadrics are independent, meshing shows that the recovered surfaces show minimal discontinuities despite no surface continuity constraints enforced during training. We show this in the wireframe close-up in the last appendix. The continuity in predicted surfaces likely comes from two sources: the quadric fitting that uses a neighbourhood of points (including adjacent voxels) and the UNet inherently learning it from spatial context. Enforcing surface continuity in quadrics during training is left open for now.

References

[1] J. Lee et al., "SemCity: Semantic Scene Generation with Triplane Diffusion," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.

[2] X. Ren et al., "XCube: Large-Scale 3D Generative Modeling using Sparse Voxel Hierarchies," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.

[3] H. Bian et al., "DynamicCity: Large-Scale 4D Occupancy Generation from Dynamic Scenes," in Proceedings of the International Conference on Learning Representations (ICLR), 2025.

[4] W. Zheng et al., "OccWorld: Learning a 3D Occupancy World Model for Autonomous Driving," in Proceedings of the European Conference on Computer Vision (ECCV), 2024.

[5] B. Li et al., "UniScene: Unified Occupancy-centric Driving Scene Generation," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025.

[6] W. E. Lorensen and H. E. Cline, "Marching Cubes: A High Resolution 3D Surface Construction Algorithm," ACM SIGGRAPH Computer Graphics, vol. 21, no. 4, pp. 163–169, 1987.

[7] J. Huang et al., "Neural Kernel Surface Reconstruction," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023.

Reviewer_XTwWOpenReview ↗

Clarifications

I thank the authors for their response.

W1 Setting. This is a fair point. Also, other reviewers did not criticize this aspect, so I stand corrected.

W2 Straightforward. Even excluding the model, the representation itself with quadrics is very classical and not particularly outstanding. I fully agree that this kind of work has value, but it reads more like a technical report than an ICLR paper with a clear novel contribution.

W3 Surface visualization. Thank you for this addition. The resulting surfaces are qualitatively of very poor quality, which is not that surprising. Some form of surface prior could be useful, like the new SDF baseline added, but it is clearly not trained well.

W4 Baselines. The task of this paper represents a small part of UniScene's pipeline, however the improvement is clear and it is now a fair comparison. This is a good addition.

Q Surface continuity. It is hard to interpret this from the wireframe without access to a 3D mesh in supplementary, and also because the voxels are not displayed so we don't know where to look. This could be measured with a metric easily. Overall, there are many surface patches that are disconnected from the rest, showing that per-voxel-quadrics only bring a very weak surface prior.

I am still of the opinion that the methodological contribution of this paper is too narrow for ICLR. The representation and models are both very standard, and a more ambitious submission could have included a better surface prior. That said, the clarification on motivation and the new baseline are valuable, and the paper and rebuttal are both very clear and well written, so I will improve my rating from clear reject to weak reject.

AuthorsOpenReview ↗

Response to Reviewer XTwW (follow-up)

We thank the reviewer for the response.

W1. Setting.

We are glad that the reviewer is convinced of the validity of the setting.

W2. The representation is classical.

We agree that the quadric representation is classical, and we do not claim it as a contribution. Our first response addressed the concern that our work uses a simple 3D U-Net, so it did not discuss our other contributions. Besides predicting quadrics, we develop two mesh-free operators designed for this representation: a vectorised Newton-Raphson projection for dense surface sampling and an analytical ray-quadric intersection for LiDAR simulation. Unlike standard methods and our baselines, both act on the coefficients directly, with no mesh extraction or query-time network evaluation. Our cost analysis clearly outlines their benefits.

W3. The surfaces are of poor quality, and the SDF baseline is not trained well.

The surface quality shows that recovering accurate surfaces from an occupancy grid alone is a difficult problem, because the voxel grid carries little geometric information. A learned surface prior is tested with our QuadricUNet with an SDF head.

We would like to assure the reviewer that the models are trained fairly. QuadricUNet (SDF head) keeps our U-Net and replaces the quadric head with UniScene's SDF field, samplers and NeuS renderer, adapted from UniScene's official implementation to our setting. Training details are documented in Appendix G. It matches QuadricUNet in Chamfer distance (5.01 cm with SDF against 4.97 cm) and the F1@10cm is 6.0 points lower. One reason is that our data setting limits what an SDF prior can gain. The aggregated LiDAR point clouds are affected by range noise of up to a few centimetres, pose errors and drift. Neither dataset provides meshes, so the clean signed distances that supervised SDF training would benefit from are not available in our setting. Our quadric targets are robust fits, which reduce this noise in the supervision (Appendix B).

W4. Baselines.

We thank the reviewer for the positive feedback on the added baselines. We would like to clarify that our paper and UniScene both contain semantic occupancy-to-LiDAR pipelines as part of their tasks but the primary focus differs. Our focus is recovering sub-voxel surface geometry, whereas UniScene primarily focuses on BEV layouts to semantic occupancy, LiDAR and video generation.

Q. Surface continuity could be measured with a metric.

We quantify continuity by measuring surfaces at voxel boundaries directly (sequence 08, 20 frames). We assume a face between two occupied voxels is crossed when aggregated LiDAR points (reference) on both sides of the face form a surface across the face within a 2 cm threshold. On 4,000 crossed faces (200 per frame, sampled at random from 89,460 total across 20 frames), we compare predicted and reference points across the face. A face is considered linked when the surfaces of its two voxels are within 5 or 10 cm.

QuadricUNet is the most accurate at voxel boundaries, with the lowest Chamfer distance and highest F1. The SDF head, a single continuous field (same as UniScene), links the most faces. Of the faces QuadricUNet does not link at 5 cm, about half link at 10 cm, so many of its gaps are steps of a few centimetres. We already state this surface continuity shortcoming in the limitations section.

MethodAcc. (cm)Comp. (cm)Chamfer (cm)F1@5/10 (%)Linked@5/10 cm (%)
QuadricUNet3.622.653.1385.1/99.382.6/90.6
QuadricUNet (SDF head)4.342.263.3080.4/97.395.7/96.5
UniScene4.462.573.5277.3/97.389.4/89.9
NKSR4.527.045.7854.7/83.463.9/63.9
Marching cubes10.8414.4612.656.4/32.358.0/58.6

We thank the reviewer again for the constructive discussion.

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

This paper proposes QuadricUNet, a method for recovering sub-voxel 3D geometry from semantic occupancy grids. The method employs a 3D U-Net to predict an implicit quadric, parameterized by ten coefficients, for each occupied voxel, and introduces a curvature-weighted loss to improve local geometric prediction. The authors further develop Newton projection for dense surface sampling and analytical ray intersection for LiDAR simulation. Experiments on SemanticKITTI and the Waymo Open Dataset show that the proposed method outperforms Marching Cubes and NKSR in surface reconstruction metrics. In addition, semantic segmentation models trained on simulated LiDAR sweeps generated by the proposed method achieve better performance on real data. This work provides a potentially useful approach to recovering continuous geometry from semantic voxel representations.

Critical strengths

The method proposes a solution to the specific problem of geometric information loss within voxels. The representation of the quadratic surface is concise, and the analytical sampling and ray intersection have practical significance. The surface reconstruction results on two datasets, as well as the simulation-to-reality transfer experiments on SemanticKITTI, provide preliminary and relatively consistent evidence for this design.

Critical weaknesses

The basis for some of the experimental conclusions is not yet sufficient.

Questions for authors
  1. The paper claims that only a semantic occupancy grid is required at inference time. However, Appendix A indicates that the SemanticKITTI input voxels are filtered according to the support provided by aggregated LiDAR points, while the Waymo input grids are constructed by voxelizing aggregated labeled point clouds. Could the authors clarify whether constructing the input grids at inference time depends on raw LiDAR point clouds and their observation coverage?

  2. In Section 4.2, the authors compare QuadricUNet's F1@5 cm of 60.09% with the 73.59% achieved by the reference quadric fitting method in Table 4, concluding that the gap mainly reflects prediction error rather than insufficient representational capacity. However, Appendix J explicitly states that the fitting domain and evaluation protocol of Table 4 differ from those of Table 1, including voxels that are unoccupied or have sparse point support. Therefore, this comparison does not sufficiently support the stated conclusion. Could the authors reevaluate the reference fitting results and network predictions using the same voxel set, reference point clouds, and evaluation protocol?

  3. The paper states that all experimental results are obtained from single training runs, while the improvements attributed to the curvature loss are relatively small. Could the authors report the mean and standard deviation of the main results across multiple random seeds to establish the stability of these improvements?

2 comments
AuthorsOpenReview ↗

Response to Reviewer EeGo [1/2]

Thank you for the constructive review and feedback. We have addressed your questions below.

Q1 Could the authors clarify whether constructing the input grids at inference time depends on raw LiDAR point clouds and their observation coverage?

At inference time, the network input is strictly a semantic voxel grid and does not access LiDAR points or their coverage. The voxel grids provided in the SemanticKITTI dataset [1] are voxelized from aggregated LiDAR. The Waymo Open Dataset [2] does not provide such grids, so we build them following standard practice [1,3]. During dataset creation, we filtered the grids with LiDAR to improve data quality. These filters are applied once, during dataset creation. We designed the method to avoid reliance on other data modalities at inference (normals, images, LiDAR points) and work purely on semantic voxel grid inputs that could be created by existing techniques [4–7]. Our method can be applied to bridge the quality gap between such grids and LiDAR-mapped 3D environments.

Q2 Could the authors reevaluate the reference fitting results and network predictions using the same voxel set, reference point clouds, and evaluation protocol?

We agree that the original comparison was not consistent, and we have corrected it. We made sure that all evals are now consistent across the methods. On SemanticKITTI, the predicted quadrics nearly match the fitted quadrics at 10 cm, and a gap remains only at the tighter 5 cm threshold. We attribute this difference to prediction error and not a limitation of the representation. Section 4.2 now uses this comparison.

Table 1. Fitted reference quadrics, QuadricUNet predictions and the baselines on the same voxel set, SemanticKITTI sequence 08 (Table 1).

MethodChamfer (cm) ↓HD95 (cm) ↓F1@5 (%) ↑F1@10 (%) ↑
Fitted quadrics3.9214.366.988.4
QuadricUNet (predicted quadrics)4.97 ± 0.0413.6 ± 0.255.6 ± 0.187.8 ± 0.4
QuadricUNet (SDF head)5.0128.948.281.8
UniScene5.4715.345.881.7
NKSR8.4731.634.265.2
Marching Cubes12.9924.46.330.5

The numbers are different (lower) than in the submission due to the following updates:

  1. Every input voxel with reference points is scored, including voxels with too few points to fit a reference quadric. The submission previously ignored voxels without a quadric fit (too few points to fit a quadric). Tables 1 and 4 now use the same fitter and the same voxel sets, without dropping voxels.
  2. We found and fixed a bug in the evaluation code that caused a class absent from a prediction but present in GT to be skipped. It now scores F1 = 0. (Table 1)

Q3 (seed runs and the curvature term) follows in the next comment.

References

[1] J. Behley et al., "SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences," in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019.

[2] P. Sun et al., "Scalability in Perception for Autonomous Driving: Waymo Open Dataset," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020.

[3] X. Tian et al., "Occ3D: A Large-Scale 3D Occupancy Prediction Benchmark for Autonomous Driving," in Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2023.

[4] A.-Q. Cao and R. de Charette, "MonoScene: Monocular 3D Semantic Scene Completion," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.

[5] Y. Li et al., "VoxFormer: Sparse Voxel Transformer for Camera-based 3D Semantic Scene Completion," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023.

[6] J. Lee et al., "SemCity: Semantic Scene Generation with Triplane Diffusion," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.

[7] B. Li et al., "UniScene: Unified Occupancy-centric Driving Scene Generation," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025.

AuthorsOpenReview ↗

Response to Reviewer EeGo [2/2]

Q3 Could the authors report the main results across multiple random seeds?

We carried out the 3 seed runs and report results for QuadricUNet on reconstruction, LiDAR simulation and segmentation performance. The spread across seeds is much smaller than the gaps to the baselines, so the comparisons in the paper hold.

Table1. QuadricUNet over three seeds, SemanticKITTI sequence 08 (Tables 1 to 3).

RunChamfer (cm) ↓HD95 (cm) ↓F1@5 (%) ↑F1@10 (%) ↑Per-ray median (cm) ↓Ray F1@10 (%) ↑Segmentation performance (SPVCNN mIoU, %) ↑
Seed 05.0113.555.587.514.639.234.91
Seed 14.9613.955.787.614.940.233.91
Seed 24.9313.455.688.214.939.635.17
Mean ± std4.97 ± 0.0413.6 ± 0.255.6 ± 0.187.8 ± 0.414.8 ± 0.239.6 ± 0.534.66 ± 0.67

Our initial submission claimed that the curvature term offered a small improvement over the Huber-only loss function. However, with the 3 seed runs, we find that overall performance is similar and training with the curvature term enabled offers minor benefit to a few of the classes as described in the table below. We made the necessary corrections in the revision. Section 4.4 and Tables 8 and 9 (Appendix K) now state these effects.

Table 2. Curvature term on and off (Table 8), mean ± std over three seeds.

SettingChamfer (cm) ↓F1@5 (%) ↑F1@10 (%) ↑
Curvature term off4.98 ± 0.0355.5 ± 0.287.8 ± 0.3
Curvature term on (ours)4.97 ± 0.0455.6 ± 0.187.8 ± 0.4

Table 3. Per-class surface accuracy with the curvature term on and off, mean ± std over three seeds. Better: with the term on, the mean Chamfer distance is lower and the mean F1@5 higher, both by more than the seed std of either setting.

ClassChamfer (cm) ↓, on / offF1@5 (%) ↑, on / offEffect of the term
car4.60 ± 0.02 / 4.59 ± 0.0660.9 ± 0.2 / 61.2 ± 0.5Not significant
bicycle5.25 ± 0.05 / 5.25 ± 0.1048.8 ± 0.4 / 48.9 ± 0.8Not significant
motorcycle4.79 ± 0.02 / 4.80 ± 0.0758.3 ± 0.3 / 58.4 ± 0.6Not significant
truck5.53 ± 0.04 / 5.61 ± 0.1245.6 ± 0.2 / 45.5 ± 0.6Not significant
other-vehicle4.77 ± 0.02 / 4.79 ± 0.1058.8 ± 0.3 / 58.7 ± 0.9Not significant
person5.09 ± 0.03 / 5.10 ± 0.0552.6 ± 0.4 / 52.9 ± 0.2Not significant
bicyclist8.45 ± 0.79 / 8.26 ± 1.5815.1 ± 1.8 / 18.6 ± 9.0Not significant
motorcyclist4.74 ± 0.06 / 4.72 ± 0.1760.0 ± 1.5 / 56.1 ± 4.9Not significant
road3.51 ± 0.01 / 3.56 ± 0.0276.9 ± 0.2 / 76.4 ± 0.3Better
parking3.98 ± 0.03 / 4.03 ± 0.0269.2 ± 0.5 / 68.6 ± 0.1Better
sidewalk3.65 ± 0.02 / 3.69 ± 0.0374.7 ± 0.2 / 74.2 ± 0.4Better
other-ground6.35 ± 0.10 / 6.30 ± 0.0738.8 ± 0.6 / 39.2 ± 0.2Not significant
building4.65 ± 0.02 / 4.66 ± 0.0757.9 ± 0.1 / 58.2 ± 0.9Not significant
fence4.74 ± 0.02 / 4.75 ± 0.0757.9 ± 0.1 / 57.9 ± 0.7Not significant
vegetation5.34 ± 0.18 / 5.55 ± 0.4248.6 ± 1.3 / 47.6 ± 2.4Not significant
trunk4.76 ± 0.03 / 4.77 ± 0.0858.1 ± 0.4 / 58.3 ± 1.0Not significant
terrain4.35 ± 0.01 / 4.40 ± 0.0362.1 ± 0.1 / 61.5 ± 0.3Better
pole4.75 ± 0.06 / 4.72 ± 0.0858.2 ± 1.1 / 59.0 ± 0.9Not significant
traffic-sign5.07 ± 0.02 / 5.05 ± 0.0653.6 ± 0.3 / 54.1 ± 0.8Not significant

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal general response

We thank the reviewers for their time and for their constructive and detailed feedback. We apologize for the delayed response. We address the questions with as many supporting experiments as possible within the rebuttal phase.

We have also uploaded a revised PDF. The updates are described below:

Major updates

  1. New learned baselines on the same input setting. We add two new learned baselines. UniScene's occupancy-to-LiDAR model [1] and QuadricUNet (SDF head), which is our own U-Net with a voxel-conditioned implicit decoder in place of the quadric head. Both train on the same input setting as our proposed model (QuadricUNet). (Response to: XTwW W4, AC4r Q1).
  2. We unified the evaluation protocol. The same protocol now scores all methods on the same voxels, frames and sampling density. We also corrected a half-voxel offset in the Marching Cubes baseline, and NKSR runs with its best official outdoor settings. (Response to: EeGo Q2, XTwW W4).
  3. Three seed runs. QuadricUNet and the curvature ablation are reported over 3 seed runs as mean ± std. With three seeds, the curvature term leaves the overall scores unchanged and gives small gains on the ground classes; Section 4.4 is corrected accordingly. (Response to: EeGo Q3).
  4. End-to-end cost on the GPU, including the network pass (Response to: AC4r Q3).

Minor updates

  1. Visualisations. Figure 2 now shows meshes of all five methods on the same SemanticKITTI scene, as a street view and a close-up. The simulated LiDAR sweeps, the densely sampled surfaces from the submission, and a wireframe close-up of the QuadricUNet mesh are added to the appendix. (Response to: XTwW W3).
  2. Evaluation bug fix. A class present in the ground truth but absent from a prediction was skipped in the F1 evaluation of surface point sampling. It now scores F1 = 0.
  3. Segmentation protocol. Segmentation performance (SPVCNN mIoU) is lower than in the submission for every method. We now report the standard dataset-level mIoU, computed over all test points, instead of the per-scan mean used in the submission.

Summary

  1. We show that under the updated evaluation settings, QuadricUNet has the highest F1 of all learned and classical methods on both datasets and the lowest LiDAR error, and its simulated LiDAR transfers best to real scans.
  2. Our new experiments also support that the gains come from the quadric representation, not the backbone. On the same U-Net, an implicit SDF head variant is trained to verify this.
  3. The corrected protocol and new baselines change several numbers, but the conclusions stated in our initial submission still hold. However, our claim on the curvature term's significance was overstated and is corrected in the revised paper.

References

[1] B. Li et al., "UniScene: Unified Occupancy-centric Driving Scene Generation," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.