We thank the reviewer for the constructive comments and positive feedback on our paper. Regarding the concerns of the reviewer GNa2, we provide the following responses.

W&Q1: A key concern is that the observed reduction in primitives may stem more from the aggressive pruning-and-splitting module rather than the intrinsic properties of superquadrics.

• Thanks for your valuable suggestion! To directly evaluate the advantage of superquadrics, we performed an ablation experiment where we replaced the superquadrics in QuadricFormer (with 1,600 primitives) with Gaussians, while keeping all other aspects of the model exactly the same.

• The results in the table below show that the Gaussian-based model achieves much lower performance. This is because, without a heavy depth backbone and with a limited number of primitives, the Gaussian-based method cannot model the scene as effectively as superquadrics. In contrast, the superquadric-based approach maintains strong performance and efficiency.

• These results demonstrate that the core improvements in performance and efficiency stem from the intrinsic representational power of superquadrics, rather than from architectural choices or the pruning strategy alone. We have included these findings in the revised manuscript for greater clarity. Thank you for prompting us to further validate this important point.

W&Q2: The rendered superquadrics often appear as simple cuboids, quite similar in shape to what Gaussians would produce, just at a larger scale. They do not seem to wrap around object corners or capture complex surfaces any better than their Gaussian counterparts in the visualizations provided.

• Thanks for your insightful suggestion. We acknowledge that superquadrics cannot capture precise object boundaries or highly complex surfaces in the occupancy prediction task. This is because the task uses 0.5-meter voxel labels, which only represent approximate object geometry rather than detailed surfaces.

• However, superquadrics do possess clear advantages over Gaussians in modeling typical scene structures. For example, modeling large flat regions (such as road surfaces) would require many overlapping Gaussian ellipsoids, whereas only a small number of cuboid superquadrics are needed.

• While images cannot be included in the NIPS rebuttal, we have added more detailed visualizations in the revised manuscript to better demonstrate the structural advantages of superquadrics in capturing geometric details within the limitations of current occupancy benchmarks.

W&Q3: The evaluation is currently limited to the nuScenes dataset. To demonstrate the generalizability and robustness of the proposed method, it is crucial to include results on other standard benchmarks for this task.

• In response to the reviewer’s concerns about robustness and generalizability of our method, we have conducted additional experiments on the SSCBench-KITTI-360 [1] dataset. The results are summarized in the table below.

• Compared to previous dense voxel-based and sparse Gaussian-based methods, our method achieves comparable performance with much fewer scene primitives. This highlights the efficiency of our method, as it can represent complex scenes with fewer expressive superquadrics without sacrificing much accuracy.

• We would like to clarify a key difference with GaussianFormer-2, which employs a heavy depth backbone to initialize Gaussians for improved performance. While this strategy can yield better results, it comes at the cost of significantly reduced efficiency (i.e., high inference latency and memory usage), as evidenced by Table 2 in the paper. In contrast, to preserve the efficiency advantage of our method, we did not incorporate such a heavy depth-initialized module.

• These results demonstrate both the effectiveness and the efficiency of our approach. We hope this addresses your concerns and further highlights the practical advantages of our method.

• Due to the limited time during the rebuttal period, we only conducted the experiments with 12800 superquadrics on SSCBench-KITTI-360. We will provide more experiments on Occ3D and SSCBench-KITTI-360 later.

W&Q4: Table 2 presents a confusing picture regarding efficiency. It is counter-intuitive that a model with N superquadrics would be faster and require less memory than a model with N Gaussians, as the table seems to imply.

• The key reason for the efficiency difference is not the number of parameters per primitive, but rather the architectural choices. Specifically, GaussianFormer-2 relies on a heavy depth backbone (ResNet-101) to initialize the Gaussians, which becomes the main bottleneck for inference speed and memory usage. In contrast, QuadricFormer adopts cost-free random initialization for the superquadrics, which substantially reduces the computational and memory overhead.

• We appreciate your suggestion, which helped us improve Table 2 to avoid any confusion. In the revised table, we have now indicated whether depth-backbone-based initialization is used and provided model parameters, making it clear where the efficiency gains originate.

W&Q5: The comparison with GaussianFormer-2 appears to be against a weaker variant (Ch=128, mIoU=20.02) rather than its best-performing published model (Ch=192, mIoU=20.82), which has a higher mIoU than the proposed method.

• Thanks for your suggestion. In Table 1, we only reported the results of GaussianFormer-2 (Ch=128, mIoU=20.02) with a feature dimension of 128 to ensure a fair comparison, since our method is under the same setting. However, we agree that the best results should also be included to provide a more complete picture. In the revised version, we have added the strongest results of GaussianFormer-2 (Ch=192, mIoU=20.82) and clarified this in the manuscript.

• Notably, even when compared with the best configuration of GaussianFormer-2, our QuadricFormer (with 12,800 superquadrics) still achieves superior performance with clear efficiency advantages. We appreciate your suggestion, which has improved the completeness and transparency of our experimental comparisons.

W&Q6: Several key hyperparameters are missing.

Thanks for your detailed comments and suggestions. We address your questions as follows:

1. Number of Quadrics before Pruning-and-Splitting: The number of superquadrics remains unchanged before and after the pruning-and-splitting stage, since each pruned superquadric is replaced by a newly added one.
2. Definition of a Block and Value of B in the Decoder: A "quadric-encoder block" refers to a module that consists of a sparse convolutional layer, a cross-attention layer, and a residual update layer. Our model employs $B=4$ blocks to encode superquadrics.
3. Small-scale and Large-scale Quadrics: The distinction between small-scale and large-scale superquadrics is determined by the product of their scale values. More details can be found in our general response (GR3).

We appreciate your valuable suggestions. We will include these clarifications in the revised version of our manuscript.

W&Q7: Recent relevant LiDAR-based methods are missing: PaSCo (CVPR'24), SCPNet (CVPR'23), and CAL (ICML'25).

• Thanks for your suggestion. We will include these works and discuss them in the revised version of our paper.

References

[1] SSCBench: A Large-Scale 3D Semantic Scene Completion Benchmark for Autonomous Driving, IROS 2024.

[2] LMSCNet: Lightweight Multiscale 3D Semantic Completion, 3DV 2020.

[3] Semantic Scene Completion From a Single Depth Image, CVPR 2017.

[4] MonoScene: Monocular 3D Semantic Scene Completion, CVPR 2022.

[5] VoxFormer: Sparse Voxel Transformer for Camera-Based 3D Semantic Scene Completion, CVPR 2023.

[6] Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction, CVPR 2023.

[7] OccFormer: Dual-path Transformer for Vision-based 3D Semantic Occupancy Prediction, ICCV 2023.

[8] GaussianFormer: Scene as Gaussians for Vision-Based 3D Semantic Occupancy Prediction, ECCV 2024.

[9] GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction, CVPR 2025.

Thanks the authors for their detailed response and the considerable effort put into conducting experiments, especially within such a short timeframe.

Regarding W&Q1: Is reduction in primitives stem from the pruning-and-splitting module?

I am convinced by the authors' response that the performance gains stem from the use of superquadrics rather than network architecture components like the "aggressive pruning-and-splitting module." Additionally, the fact that superquadrics don't require depth initialization is a significant advantage of the method.

Regarding W&Q2: The rendered superquadrics appear as simple cuboids, similar to Gaussians?

I agree that "modeling large flat regions (such as road surfaces) would require many overlapping Gaussian ellipsoids, whereas only a small number of cuboid superquadrics are needed." The results in Table 2 convincingly show that QuadricFormer performs best for "large flat regions" classes. However, Figure 2a is somewhat misleading.

Regarding W&Q3 (also raised by #N2Bk, #cJXu, #GNa2): Experiments are done only on NuScenes?

The authors' addition of results on SSCBench-KITTI-360 addresses this concern adequately.

Regarding W&Q5 (also shared by #N2Bk, #cJXu): Why not compare against the strongest GaussianFormer-2 variant?

The authors state that "In Table 1, we only reported the results of GaussianFormer-2 (Ch=128, mIoU=20.02) with a feature dimension of 128 to ensure a fair comparison, since our method is under the same setting." Could you elaborate on what "the same setting" means here? Typically, we compare the best against the best. If GaussianFormer-2 has a more advantageous setting, why not place QuadricFormer in the same setting for comparison?

Clarification on pruning-and-splitting module

If the number of superquadrics remains unchanged (since each pruned superquadric is replaced by a newly added one), isn't the name "Pruning-and-splitting" misleading?

Thank you for the clarifications on "Definition of a Block and Value of B in the Decoder" and "Small-scale and Large-scale Quadrics," as well as confirming that relevant works will be included in the revised paper.

My current stand

I find the comments from other reviewers compelling:

• SOTA claim issues (#cJXu, #GNa2, #N2Bk)
• Insufficient experiments (#uPb2)
• Confusing pruning-and-splitting module (#N2Bk)

I would like to see these discussions resolved before making my final decision.

However, I disagree with reviewer #uPb2's assertion that "The idea lacks sufficient impact." Research impact is difficult to judge and cannot be dismissed simply due to lower performance metrics. Demonstrating that an idea works and showing its benefits compared to counterparts, even if not SOTA, is already sufficient contribution.

I currently remain overall positive about the paper.

Great results! Thank you for the detailed response. I truly appreciate the additional experiments you conducted, especially over the weekend.

Overall, I find that all concerns, including those raised by other reviewers, have been addressed convincingly:

• The use of superquadrics as a new representation for 3D occupancy prediction is both novel and non-trivial. While previously explored in indoor scenes, this is the first time they are applied to urban environments.

• The advantages of this method are clear: it matches or even outperforms GaussianFormer-2 (very recent SOTA), while operating under fewer constraints (e.g., no need for depth initialization). The improvements on large object classes are particularly noteworthy and observed on nuScenes. Not sure about SCCBench-KITTI360, maybe the authors should report per-class performance on this dataset too, I hope it doesn't require much additional work?

• Regarding benchmarks, I believe Occ-3D and SurroundOcc were released around the same time, roughly between the CVPR’23 and ICCV’23 deadlines. I'm not convinced that Occ-3D is “more widely adopted,” as suggested by reviewer #uPb2. Evaluating on one benchmark seems sufficient, especially since the authors follow the standard benchmarks used by TPVFormer, SurroundOcc, GaussianFormer, and GaussianFormer-2, arguably the most influential works in this domain.

• Finally, the pruning-and-splitting module is now much clearer, and I believe it can be revised with minimal effort.

In summary, I’m convinced by the method’s merits and would support accepting the paper (will raise score to 5), unless other reviewers raise any critical points.

We thank the reviewer for the constructive comments. Regarding the concerns of the reviewer uPb2, we provide the following responses.

W1 & Q1 & Q2: The proposed idea lacks sufficient impact. The efficiency comparison requires further clarification.

• On the fairness and validity of efficiency comparison:

  1. We would like to clarify that our efficiency comparison is fair and meaningful. As noted, model efficiency (inference speed and memory usage) can be affected by various factors. To ensure a fair comparison with Gaussian-based methods, we made the model architectures and evaluation settings as consistent as possible.
  2. The only difference is that GaussianFormer-2 requires an additional depth backbone for Gaussian initialization, while our method dose not. Instead, we employ a lightweight pruning-and-splitting module to enhance performance.
  3. Since the depth backbone in GaussianFormer-2 is a heavy ResNet-101, it becomes the main bottleneck for inference speed and memory usage. As a result, even with the same number of primitives, our method achieves better performance and efficiency. This highlights the effectiveness and efficiency of our method.

• On performance gains and broader impact:

  1. We respectfully disagree that the performance improvements over GaussianFormer-2 are limited. While GaussianFormer-2 achieves strong results by leveraging a heavy depth backbone and increasing feature dimensions, this comes at the cost of significant efficiency loss (see Table 2).
  2. Without relying on these additional tricks, QuadricFormer (with 12800 superquadrics) surpasses GaussianFormer-2 in mIoU (21.11 vs 20.82) with superior efficiency (179 ms latency vs 451 ms). This demonstrates that our method not only exceeds the performance of existing baselines but also sets a new standard for efficiency, which is valuable for real-world applications.

• In summary, our results demonstrate that the proposed superquadric representations leads to meaningful advances in both performance and efficiency, supporting the impact and relevance of our work. We thank you for your suggestions, which have encouraged us to further clarify these points in the revised manuscript.

W2 & Q3 & Q4: The experiments are insufficient in terms of benchmark diversity and baseline comparisons. The evaluation metrics are inadequate and need to be updated.

• Experiments on additional benchmarks:

  • Thanks for your valuable suggestion. Due to the limited time of the rebuttal period, we were only able to provide additional results on SSCBench-KITTI-360 [1], following the experimental protocol of GaussianFormer-2. The results are summarized in the table below.

  • Compared to previous dense voxel-based and sparse Gaussian-based methods, our method achieves comparable performance with much fewer scene primitives. This highlights our method’s efficiency, as it is able to represent complex scenes with fewer expressive superquadrics without sacrificing much accuracy.

  • We would like to clarify a key difference with GaussianFormer-2, which employs a heavy depth backbone to initialize Gaussians for improved performance. While this strategy can yield better results, it comes at the cost of significantly reduced efficiency (i.e., high inference latency and memory usage), as evidenced by Table 2 in the paper. In contrast, to preserve the efficiency advantage of our method, we did not incorporate such a heavy depth-initialized module.

  • These results demonstrate both the effectiveness and the efficiency of our approach. We hope this addresses your concerns and further highlights the practical advantages of our method.

  • Due to the limited time during the rebuttal period, we only conducted the experiments with 12800 superquadrics on SSCBench-KITTI-360. We will provide more experiments on Occ3D and SSCBench-KITTI-360 later.

  Method	Input Modality	Number of Primitives	IoU	mIoU
  LMSCNet [2]	LiDAR	-	47.53	13.65
  SSCNet [3]	LiDAR	-	53.58	16.95
  MonoScene [4]	Camera	262144 (128x128x16)	37.87	12.31
  Voxformer [5]	Camera	262144 (128x128x16)	38.76	11.91
  TPVFormer [6]	Camera	81920 (256x256+256x32x2)	40.22	13.64
  OccFormer [7]	Camera	262144 (128x128x16)	40.27	13.81
  GaussianFormer [8]	Camera	38400	35.38	12.92
  GaussianFormer-2 [9]	Camera	38400	38.37	13.90
  Ours	Camera	12800	36.81	12.86

• Evaluation metrics:

  • While SparseOcc [10] introduced the RayIoU metric, the majority of recent and leading works on the SSCBench-KITTI-360 continue to report results primarily with mIoU and IoU as evaluation metrics. Therefore, for consistency and comparability with prior art, we continue to evaluate our model using these established metrics.
  • Nevertheless, we acknowledge the value of more comprehensive evaluation. In future work, we will conduct additional experiments on the Occ3D dataset and adopt metrics such as RayIoU to enable a more thorough comparison with more recent occupancy prediction methods.

References

[1] SSCBench: A Large-Scale 3D Semantic Scene Completion Benchmark for Autonomous Driving, IROS 2024.

[2] LMSCNet: Lightweight Multiscale 3D Semantic Completion, 3DV 2020.

[3] Semantic Scene Completion From a Single Depth Image, CVPR 2017.

[4] MonoScene: Monocular 3D Semantic Scene Completion, CVPR 2022.

[5] VoxFormer: Sparse Voxel Transformer for Camera-Based 3D Semantic Scene Completion, CVPR 2023.

[6] Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction, CVPR 2023.

[7] OccFormer: Dual-path Transformer for Vision-based 3D Semantic Occupancy Prediction, ICCV 2023.

[8] GaussianFormer: Scene as Gaussians for Vision-Based 3D Semantic Occupancy Prediction, ECCV 2024.

[9] GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction, CVPR 2025.

[10] Fully Sparse 3D Occupancy Prediction, ECCV 2024.

Thanks to the AC and the reviewer for their valuable comments and feedback.

We would like to clarify that Occ-3D and SurroundOcc were released around the same time and provide additional 3D occupancy annotations to the nuScenes data. Subsequent methods usually chose one of them as the occupancy evaluation benchmark on nuScenes.

Typically, voxel-based 3D occupancy prediction methods targeting improving the model architecture (e.g., SparseOcc, FB-Occ, and CVT-Occ) and occupancy generation methods (e.g., OccSora) usually adopt Occ-3D. On the other hand, methods targeting developing new 3D representations to replace the voxel representation (e.g., GaussianFormer [1], GaussianFormer-2 [2]) usually adopt SurroundOcc.

We followed this convention, and as our main goal is to validate the effectiveness and advantage of superquadrics as a new scene representation, we adopted SurroundOcc to ensure fair comparisons with the most relevant counterparts (e.g., the Gaussian representation in GaussianFormer, GaussianFormer-2). This choice is also supported by Reviewer #VvnD ("the authors follow the standard benchmarks used by TPVFormer, SurroundOcc, GaussianFormer, and GaussianFormer-2, arguably the most influential works in this domain.").

Therefore, even though we agree that additional evaluation on Occ3D is a valuable enhancement, we respectfully disagree that the lack of it is a fundamental problem. We are currently trying our best to adapt our method to Occ3D and will provide the results once we have them. Once again, we appreciate the time and effort the AC and reviewer devoted to reviewing and improving our work. We hope the above response can help address your concerns. We are happy to answer any additional questions you may have.

References

[1] GaussianFormer: Scene as Gaussians for Vision-Based 3D Semantic Occupancy Prediction, ECCV 2024.

[2] GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction, CVPR 2025.

Since the introduction of the Occ3D benchmark, it has been cited nearly 300 times, and numerous milestone methods have been evaluated on it. Therefore, comparison with these works (a small part of them are as below) is essential to meet the high standards of NeurIPS.

I found this comment is a bit bias. Regarding adoption metrics, SurroundOcc acctually demonstrates higher adoption:

• Citations: 323 (SurroundOcc) vs 289 (Occ3D)
• GitHub stars: 937 (SurroundOcc) vs 486 (Occ3D)

We thank the reviewer for the constructive comments and positive feedback on our paper. Regarding the concerns of the reviewer GNa2, we provide the following responses.

W1: The experimental evaluation in the current version is insufficient. I strongly encourage the authors to include additional evaluations to compare with other sparse occupancy prediction methods.

• Thanks for your valuable suggestion. Due to the limited time of the rebuttal period, we were only able to provide additional results on SSCBench-KITTI-360 [1], following the experimental protocol of GaussianFormer-2. The results are summarized in the table below.

• Compared to previous dense voxel-based and sparse Gaussian-based methods, our method achieves comparable performance with much fewer scene primitives. This highlights our method’s efficiency, as it is able to represent complex scenes with fewer expressive superquadrics without sacrificing much accuracy.

• We would like to clarify a key difference with GaussianFormer-2, which employs a heavy depth backbone to initialize Gaussians for improved performance. While this strategy can yield better results, it comes at the cost of significantly reduced efficiency (i.e., high inference latency and memory usage), as evidenced by Table 2 in the paper. In contrast, to preserve the efficiency advantage of our method, we did not incorporate such a heavy depth-initialized module.

• These results demonstrate both the effectiveness and the efficiency of our approach. We hope this addresses your concerns and further highlights the practical advantages of our method.

• Due to the limited time during the rebuttal period, we only conducted the experiments with 12800 superquadrics on SSCBench-KITTI-360. We will provide more experiments on Occ3D and SSCBench-KITTI-360 later.

W2.1: QuadricFormer achieves lower IoU and mIoU than SurroundOcc and OccFormer. Highlighting QuadricFormer as the best-performing model is potentially misleading to readers.

• We apologize for the lack of clarity in Table 1, where we only reported the performance of QuadricFormer with 1600 superquadrics. This presentation may have caused confusion about our SOTA claim. Our model actually can be configured to use different numbers of superquadrics, and using more superquadrics yields better accuracy with only a slight reduction in efficiency. As shown in Table 2, with 6400 and 12800 superquadrics, QuadricFormer achieves mIoUs of 20.79 and 21.11 and IoUs of 31.89 and 32.13, respectively, surpassing all previous methods and achieving SOTA results.

• Moreover, unlike GaussianFormer-2, we do not use the heavy depth-initialized module, so our model maintains high efficiency even with 6400 or 12800 superquadrics (see Table 2).

• To avoid confusion, we have updated Table 1 in the revised manuscript to clarify the SOTA claim and clearly indicate the results for different numbers of superquadrics.

W2.2: The reported numbers for GaussianFormer-2 (30.56 IoU, 20.02 mIoU) are not its best-performing results.

• Thanks for your suggestion. In Table 1, we only reported the results of GaussianFormer-2 (Ch=128, mIoU=20.02) with a feature dimension of 128 to ensure a fair comparison, since our method is under the same setting. However, we agree that the best results should also be included to provide a more complete picture. In the revised version, we have added the strongest results of GaussianFormer-2 (Ch=192, mIoU=20.82) and clarified this in the manuscript.

• Notably, even when compared with the best configuration of GaussianFormer-2, our QuadricFormer (with 12,800 superquadrics) still achieves superior performance with clear efficiency advantages. We appreciate your suggestion, which has improved the completeness and transparency of our experimental comparisons.

Q1: Could the authors provide more details on the pruning-and-splitting module?

We provide implementation details of the prunning-and-splitting module. To clarify, we take the QuadricFormer with $N$ superquadrics as an example and describe the process as follows:

1. Initial Training: We first train a QuadricFormer with $B=4$ quadric-encoder blocks and without the prunning-and-splitting module. The model starts from $N$ randomly initialized superquadrics $\mathbf{Q}_{init}$ and predicts adjusted superquadrics $\mathbf{Q}$.

2. Prunning-and-Splitting Module: During experiments, we observed that some superquadrics in $\mathbf{Q}$ contribute little to scene modeling, which are usually located in empty regions with very small scales. To address this, we introduce the prunning-and-splitting module:

   • We divide all superquadrics in $\mathbf{Q}$ into two groups based on the product of their scales: the $N_{crop}$ superquadrics $\mathbf{C}$ with the smallest scales and the remaining $N_{valid}$ superquadrics $\mathbf{V}$, where $N=N_{crop}+N_{valid}$.
   • We discard the smaller superquadrics $\mathbf{C}$ as they are most likely to contribute little to scene modeling.
   • We randomly sample $N_{crop}$ superquadrics from $\mathbf{V}$ to form $\mathbf{S}$. The features of $\mathbf{S}$ remain unchanged, and only their positions are slightly adjusted.

3. Further Refinement: Finally, $\mathbf{S}$ and $\mathbf{V}$ are passed through two additional quadric-encoder blocks to further refine their attributes, resulting in the final superquadrics $\mathbf{Q}_{final}$. At this stage, we load the pretrained model parameters and continue training for 10 more epochs.

M1 & M2

1. Thanks for pointing out this editing error. We have now removed the incorrect "Ch" notation from the caption of Table 1.
2. We acknowledge that there was a mismatch between the text and results in Table 1. This was because Table 1 did not include model performance under settings of more than 1600 superquadrics, which were included in Table 2. We have updated Table 1 and revised the corresponding statements to ensure full consistency between our description and the reported results.

References

[1] SSCBench: A Large-Scale 3D Semantic Scene Completion Benchmark for Autonomous Driving, IROS 2024.

[2] LMSCNet: Lightweight Multiscale 3D Semantic Completion, 3DV 2020.

[3] Semantic Scene Completion From a Single Depth Image, CVPR 2017.

[4] MonoScene: Monocular 3D Semantic Scene Completion, CVPR 2022.

[5] VoxFormer: Sparse Voxel Transformer for Camera-Based 3D Semantic Scene Completion, CVPR 2023.

[6] Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction, CVPR 2023.

[7] OccFormer: Dual-path Transformer for Vision-based 3D Semantic Occupancy Prediction, ICCV 2023.

[8] GaussianFormer: Scene as Gaussians for Vision-Based 3D Semantic Occupancy Prediction, ECCV 2024.

[9] GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction, CVPR 2025.

I disagree with the reviewer. The claim that "works remain incomparable due to diverse evaluation benchmarks" is flawed. Diverse benchmarks actually enhance comparability by revealing method strengths and limitations across multiple perspectives.

Each benchmark captures unique aspects of occupancy prediction - different sensor modalities, environmental conditions, and task formulations. Even benchmarks using the same dataset differ significantly: although Occ3D and SurroundOcc both use nuScenes, their approaches to aggregating 3D occupancy ground-truth are different. Furthermore, SurroundOcc open-sources their ground-truth generation pipeline while Occ3D doesn't provide theirs. This open-source approach not only ensures transparency and reproducibility but also enables other researchers to generate occupancy ground-truth for additional datasets, fostering broader community development rather than restricting progress to a single closed benchmark.

Multiple benchmarks reveal method robustness and identify failure modes that single-benchmark evaluation would miss. The goal should be developing methods that perform well across diverse settings, not excelling on one standardized test.

Of course, I agree that it would be more beneficial if the authors provide results on more benchmarks, however this also requires extensive computational resources that may not be available to all research groups

We thank the reviewer for the constructive comments and positive feedback on our paper. Regarding the concerns of the reviewer cJXu, we provide the following responses.

Q1: In Table 1, the authors claim to achieve state-of-the-art performance. However, both the IoU and mIoU scores are lower than those of existing methods.

• We apologize for the lack of clarity in Table 1, where we only reported the performance of QuadricFormer with 1600 superquadrics. This presentation may have caused confusion about our SOTA claim. Our model actually can be configured to use different numbers of superquadrics, and using more superquadrics yields better accuracy with only a slight reduction in efficiency. As shown in Table 2, with 6400 and 12800 superquadrics, QuadricFormer achieves mIoUs of 20.79 and 21.11 and IoUs of 31.89 and 32.13, respectively, surpassing all previous methods and achieving SOTA results.

• Moreover, unlike GaussianFormer-2, we do not use the heavy depth-initialized module, so our model maintains high efficiency even with 6400 or 12800 superquadrics (see Table 2).

• To avoid confusion, we have updated Table 1 in the revised manuscript to clarify the SOTA claim and clearly indicate the results for different numbers of superquadrics.

Q2: I would like to see results on additional datasets such as SemanticKITTI. This would be necessary to better demonstrate the generalizability and practical effectiveness of the proposed method.

• In response to the reviewer’s concerns about robustness and generalizability of our method, we have conducted additional experiments on the SSCBench-KITTI-360 [1] dataset. The results are summarized in the table below.

• Compared to previous dense voxel-based and sparse Gaussian-based methods, our method achieves comparable performance with much fewer scene primitives. This highlights our method’s efficiency, as it is able to represent complex scenes with fewer expressive superquadrics without sacrificing much accuracy.

• We would like to clarify a key difference with GaussianFormer-2, which employs a heavy depth backbone to initialize Gaussians for improved performance. While this strategy can yield better results, it comes at the cost of significantly reduced efficiency (i.e., high inference latency and memory usage), as evidenced by Table 2 in the paper. In contrast, to preserve the efficiency advantage of our method, we did not incorporate such a heavy depth-initialized module.

• These results demonstrate both the effectiveness and the efficiency of our approach. We hope this addresses your concerns and further highlights the practical advantages of our method.

• Due to the limited time during the rebuttal period, we only conducted the experiments with 12800 superquadrics on SSCBench-KITTI-360. We will provide more experiments on Occ3D and SSCBench-KITTI-360 later.

Q3: Would increasing this number further lead to better performance? Additionally, how does the Crop & Split Number affect inference speed and memory usage?

• We provide more results to quantitatively analyze the impact of the pruning-and-splitting module on both model performance and efficiency. All experiments used a fixed total number of 1600 superquadrics.

• The results show that this module clearly improves performance with only a minor efficiency loss. Moreover, changes in $N_{crop}$ have only a minor effect on inference speed and memory usage, allowing the model to achieve better performance while maintaining high efficiency.

• The best performance is achieved when $N_{crop}=800$ (i.e., the crop & split number in the table). Increasing $N_{crop}$ beyond 800 actually leads to a slight drop in performance. We believe this is because excessive cropping removes too many superquadrics, including some valid ones, which negatively impacts the performance.

  Crop & Split Number	mIoU	IoU	Latency (ms)	Memory (MB)
  0	19.41	29.77	158	2554
  200	19.65	30.35	161	2554
  400	19.90	30.67	161	2554
  800	20.12	31.22	162	2554
  1200	19.91	30.41	164	2554

References

[1] SSCBench: A Large-Scale 3D Semantic Scene Completion Benchmark for Autonomous Driving, IROS 2024.

[2] LMSCNet: Lightweight Multiscale 3D Semantic Completion, 3DV 2020.

[3] Semantic Scene Completion From a Single Depth Image, CVPR 2017.

[4] MonoScene: Monocular 3D Semantic Scene Completion, CVPR 2022.

[5] VoxFormer: Sparse Voxel Transformer for Camera-Based 3D Semantic Scene Completion, CVPR 2023.

[6] Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction, CVPR 2023.

[7] OccFormer: Dual-path Transformer for Vision-based 3D Semantic Occupancy Prediction, ICCV 2023.

[8] GaussianFormer: Scene as Gaussians for Vision-Based 3D Semantic Occupancy Prediction, ECCV 2024.

[9] GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction, CVPR 2025.

We thank the reviewer for the constructive comments and positive feedback on our paper. Regarding the concerns of the reviewer N2Bk, we provide the following responses.

W1: Issue with SOTA Claim in Table 1

• We apologize for the lack of clarity in Table 1, where we only reported the performance of QuadricFormer with 1600 superquadrics. This presentation may have caused confusion about our SOTA claim. Our model actually can be configured to use different numbers of superquadrics, and using more superquadrics yields better accuracy with only a slight reduction in efficiency. As shown in Table 2, with 6400 and 12800 superquadrics, QuadricFormer achieves mIoUs of 20.79 and 21.11 and IoUs of 31.89 and 32.13, respectively, surpassing all previous methods and achieving SOTA results.

• Moreover, unlike GaussianFormer-2, we do not use the heavy depth-initialized module, so our model maintains high efficiency even with 6400 or 12800 superquadrics (see Table 2).

• To avoid confusion, we have updated Table 1 in the revised manuscript to clarify the SOTA claim and clearly indicate the results for different numbers of superquadrics.

W2: The evaluation is limited to nuScenes.

• In response to the reviewer’s concerns about robustness and generalizability of our method, we have conducted additional experiments on the SSCBench-KITTI-360 [1] dataset. The results are summarized in the table below.

• Compared to previous dense voxel-based and sparse Gaussian-based methods, our method achieves comparable performance with much fewer scene primitives. This highlights our method’s efficiency, as it is able to represent complex scenes with fewer expressive superquadrics without sacrificing much accuracy.

• We would like to clarify a key difference with GaussianFormer-2, which employs a heavy depth backbone to initialize Gaussians for improved performance. While this strategy can yield better results, it comes at the cost of significantly reduced efficiency (i.e., high inference latency and memory usage), as evidenced by Table 2 in the paper. In contrast, to preserve the efficiency advantage of our method, we did not incorporate such a heavy depth-initialized module.

• These results demonstrate both the effectiveness and the efficiency of our approach. We hope this addresses your concerns and further highlights the practical advantages of our method.

• Due to the limited time during the rebuttal period, we only conducted the experiments with 12800 superquadrics on SSCBench-KITTI-360. We will provide more experiments on Occ3D and SSCBench-KITTI-360 later.

W3: Pruning-and-Splitting Module Is Somewhat Confusing.

• We provide implementation details of the prunning-and-splitting module. To clarify, we take the QuadricFormer with $N$ superquadrics as an example and describe the process as follows:

  1. Initial Training: We first train a QuadricFormer with $B=4$ quadric-encoder blocks and without the prunning-and-splitting module. The model starts from $N$ randomly initialized superquadrics $\mathbf{Q}_{init}$ and predicts adjusted superquadrics $\mathbf{Q}$.

  2. Prunning-and-Splitting Module: During experiments, we observed that some superquadrics in $\mathbf{Q}$ contribute little to scene modeling, which are usually located in empty regions with very small scales. To address this, we introduce the prunning-and-splitting module:

     • We divide all superquadrics in $\mathbf{Q}$ into two groups based on the product of their scales: the $N_{crop}$ superquadrics $\mathbf{C}$ with the smallest scales and the remaining $N_{valid}$ superquadrics $\mathbf{V}$, where $N=N_{crop}+N_{valid}$.
     • We discard the smaller superquadrics $\mathbf{C}$ as they are most likely to contribute little to scene modeling.
     • We randomly sample $N_{crop}$ superquadrics from $\mathbf{V}$ to form $\mathbf{S}$. The features of $\mathbf{S}$ remain unchanged, and only their positions are slightly adjusted.

  3. Further Refinement: Finally, $\mathbf{S}$ and $\mathbf{V}$ are passed through two additional quadric-encoder blocks to further refine their attributes, resulting in the final superquadrics $\mathbf{Q}_{final}$. At this stage, we load the pretrained model parameters and continue training for 10 more epochs.

• We provide more results to quantitatively analyze the impact of the pruning-and-splitting module on both model performance and efficiency. All experiments used a fixed total number of 1600 superquadrics. The results show that this module clearly improves performance with only a minor efficiency loss. The best performance is achieved when $N_{crop}=800$ (i.e., the crop & split number in the table). Increasing $N_{crop}$ further actually reduces the number of valid superquadrics and leads to a drop in performance. Moreover, changes in $N_{crop}$ have only a minor effect on inference speed and memory usage, allowing the model to achieve better performance while maintaining high efficiency.

  Crop & Split Number	mIoU	IoU	Latency (ms)	Memory (MB)
  0	19.41	29.77	158	2554
  200	19.65	30.35	161	2554
  400	19.90	30.67	161	2554
  800	20.12	31.22	162	2554
  1200	19.91	30.41	164	2554

• Regarding the puzzlingly high IoU (39.77 for number=0) in Table 4, we apologize for this typographical error. The correct value is 29.77 (not 39.77), and this typo has been fixed in the revised manuscript.

References

[1] SSCBench: A Large-Scale 3D Semantic Scene Completion Benchmark for Autonomous Driving, IROS 2024.

[2] LMSCNet: Lightweight Multiscale 3D Semantic Completion, 3DV 2020.

[3] Semantic Scene Completion From a Single Depth Image, CVPR 2017.

[4] MonoScene: Monocular 3D Semantic Scene Completion, CVPR 2022.

[5] VoxFormer: Sparse Voxel Transformer for Camera-Based 3D Semantic Scene Completion, CVPR 2023.

[6] Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction, CVPR 2023.

[7] OccFormer: Dual-path Transformer for Vision-based 3D Semantic Occupancy Prediction, ICCV 2023.

[8] GaussianFormer: Scene as Gaussians for Vision-Based 3D Semantic Occupancy Prediction, ECCV 2024.

[9] GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction, CVPR 2025.