Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

QuadricFormer: Scene as Superquadrics for 3D Semantic Occupancy Prediction

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read QuadricFormer claims superquadrics should replace ellipsoidal Gaussians as 3D occupancy primitives, achieving better accuracy with 1,600 primitives than 12,800 Gaussians at a third of the latency.

desk verdict Plausible superquadric extension of GaussianFormer-2 with real efficiency gains, but the paper's own tables contradict its headline claim and the central shape equation has a typo. read the letter →

arxiv 2506.10977 v1 pith:XWHHTOVA submitted 2025-06-12 cs.CV

Sicheng Zuo , Wenzhao Zheng , Xiaoyong Han , Longchao Yang , Yong Pan , Jiwen Lu This is my paper
classification cs.CV
keywords superquadrics3Dsemanticoccupancypredictionobject-centricrepresentationprobabilisticmixturemodelautonomousdrivingsparsescenenuScenesefficiency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the object-centric representation used in recent 3D occupancy models is wasteful because Gaussians impose an ellipsoidal silhouette on every object, forcing many tightly packed primitives to approximate a cube, a slab, or a road surface. It replaces each Gaussian with a superquadric, a parametric surface with three scale factors and two shape exponents that can represent box-like, cylindrical, spherical, and irregular geometries. A probabilistic mixture of these superquadrics converts them into voxel occupancy and semantic probabilities. The resulting model, QuadricFormer, reports mIoU around 20.1 with 1,600 primitives and 162 ms latency, surpassing a 12,800-Gaussian baseline at 451 ms. The paper's central claim is that geometric expressiveness, not raw primitive count, is what makes an object-centric scene representation efficient.

What carries the argument

The central object is the superquadric surface function $f(\mathbf{x}) = \left(\left|\frac{x}{s_x}\right|^{2/\epsilon_2} + \left|\frac{y}{s_y}\right|^{2/\epsilon_2}\right)^{\epsilon_2/\epsilon_1} + \left|\frac{z}{s_z}\right|^{2/\epsilon_2}$, whose two shape exponents $\epsilon_1$ and $\epsilon_2$ control whether the iso-surface is ellipsoidal, box-like, cylindrical, or star-shaped. Occupancy at a point is $\exp(-f(\mathbf{x}_Q))$ in the superquadric's local coordinate frame, the final occupancy is $1 - \prod_i (1-p_i)$ over all primitives, and semantics are a weighted average of per-primitive semantic probabilities. The shape exponents are the load-bearing novelty: they let one primitive cover geometry that would require many packed ellipsoids. A pruning-and-splitting module then reallocates small primitives from empty regions to occupied regions, keeping the total budget fixed while concentrating it where it matters.

What would settle it

Take a region where two or more learned superquadrics overlap and compare the model's Eq. 8 occupancy probability with dense LiDAR-derived occupancy for that region; if the overlap region consistently reads more occupied than the underlying geometry, the independence assumption fails. A clean synthetic check is to model one solid box first with a single superquadric and then with two identical overlapping superquadrics: the correct occupancy is the same in both cases, so any change in predicted occupancy reveals the mixture bias.

Watch

Extended reading notes

Core claim

QuadricFormer's central claim is that superquadrics are the right primitive for vision-based 3D semantic occupancy prediction. Where a Gaussian's iso-probability surfaces are ellipsoids, a superquadric's surface is defined by raising each scaled coordinate to a learned power, so the same primitive can collapse into a cuboid, a cylinder, a sphere, or a rounded star-like shape. The paper turns this surface into an occupancy distribution through $\exp(-f(\mathbf{x}_Q))$ in the primitive's local frame and aggregates all primitives with a probabilistic mixture. In the paper's reported comparison, 1,600 superquadrics reach 20.12 mIoU at 162 ms while the 12,800-Gaussian baseline reaches 20.02 mIoU at 451 ms on the nuScenes benchmark.

Load-bearing premise

The final occupancy probability multiplies the 'empty' probabilities of all superquadrics as if they were independent, so if several primitives overlap the same space, the predicted occupancy can be systematically wrong.

Editorial extensions

If this is right

  • The same accuracy as a 12,800-Gaussian baseline is reachable with roughly 8x fewer primitives (1,600 superquadrics), so occupancy prediction becomes viable at much lower latency and memory on embedded driving hardware.
  • Per-class gains concentrate on shapes that are poorly approximated by ellipsoids—bicycles, motorcycles, trucks, drivable surfaces, and sidewalks—so the representation's benefit is directly tied to geometric diversity.
  • Because each superquadric is a compact, interpretable object-like unit, the representation carries an explicit structure that could support simulation, editing, or downstream reasoning without dense grids.
  • The pruning-and-splitting module improves mIoU as more primitives are reallocated from empty space to occupied regions, implying that further gains are available by making the allocation of primitives denser where the scene actually is.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The independence assumption behind Eq. 8 is inherited from the Gaussian baseline without scrutiny: if several superquadrics overlap the same voxel, the product-of-complements rule will inflate occupancy estimates, so a normalized or locally truncated mixture is the natural next experiment.
  • Because superquadrics come with explicit pose and shape parameters, the same representation could be extended to instance-level perception, where occupancy primitives are grouped into detected objects with editable geometry.
  • The efficiency comparison is measured on nuScenes, whose annotated space is mostly empty; on denser scenes the 1,600-vs-12,800 primitive gap could shrink or reverse, so the headline efficiency gain should be tested across scene densities.
  • The fixed range of shape exponents (0.1, 2) leaves power-law and flatter-than-box shapes unexplored; letting the exponents be unbounded or conditioned on semantic class is a direct testable extension.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes QuadricFormer, a 3D semantic occupancy prediction model that represents scenes with superquadric primitives instead of ellipsoidal Gaussians. Each superquadric carries position, scale, rotation, opacity, semantic probability, and two shape exponents; occupancy is obtained through a probabilistic mixture that combines per-primitive occupancy probabilities, and semantics are aggregated by occupancy-weighted averaging. The model is built on the GaussianFormer-2 architecture, with a pruning-and-splitting module to reallocate primitives toward occupied regions. Experiments on nuScenes report state-of-the-art mIoU with far fewer primitives and lower latency/memory than Gaussian-based baselines, e.g., 1600 superquadrics achieving 20.12 mIoU and 162 ms latency versus 12800 Gaussians at 20.02 mIoU and 451 ms.

Significance. If the empirical claims hold, the paper makes a useful contribution by showing that a more expressive geometric primitive than an ellipsoid can reduce the number of primitives needed for object-centric occupancy prediction while improving accuracy. The efficiency numbers in Table 2 are striking, and the qualitative visualizations support the claim of more adaptive shapes. The paper also provides an ablation of the epsilon range and of the pruning-and-splitting module, and it releases code, which strengthens reproducibility. However, the central efficiency-accuracy claim is currently entangled with an untested probabilistic aggregation rule, so the significance cannot be fully assessed until that rule is validated and the equations are made internally consistent.

major comments (5)
  1. [Sec. 3.2, Eq. (8) and Table 2] The final occupancy is computed as a noisy-OR product over all P superquadrics, with no local truncation or overlap handling specified. Under Eq. (8), every primitive contributes at every voxel, so if the conditional-independence assumption is violated the aggregation bias grows with the number of primitives. This directly affects the headline comparison in Table 2: the 1600-superquadric model is compared with a 12800-Gaussian baseline under an aggregation rule whose bias may scale with P, so the reported margin may partly reflect aggregation behavior rather than superquadric expressiveness. Please report per-voxel active-primitive counts, overlap statistics, or calibration curves, and provide an ablation that replaces Eq. (8) with a truncated-support or overlap-aware aggregation to show that the relative ranking is robust.
  2. [Sec. 3.2, Eq. (4) and Appendix A] Equation (4) places the exponent 2/epsilon_2 on the z-term, but the standard superquadric definition and the paper's own Appendix A description both require the z-exponent to be 2/epsilon_1, with epsilon_1 controlling the profile in planes containing the z-axis and epsilon_2 controlling the xy cross-section. As written, Eq. (4) makes the two shape exponents play roles different from those described in the text, which undermines the interpretation of the shape prior. Please correct Eq. (4) and Eq. (7) and verify that the implementation matches the corrected formula.
  3. [Sec. 3.2, Eqs. (7)-(9)] The opacity attribute a_i appears in the semantic aggregation of Eq. (9) but is absent from the occupancy probability in Eqs. (7)-(8). If opacity is intended to scale each primitive's contribution to occupancy, it should appear in Eq. (7) or in the aggregation; if it is only a semantic weighting factor, that role should be stated explicitly. As written, the equations are internally inconsistent because opacity affects semantics but not occupancy, despite the text describing opacity as a geometric attribute.
  4. [Sec. 3.2, Eqs. (4) and (7)] The coordinate ratios in the superquadric equation are not enclosed in absolute values. For arbitrary real-valued x_Q, expressions such as (x_Q/s_x)^(2/epsilon_2) are undefined or sign-dependent when 2/epsilon_2 is not an integer, which is the generic case for the epsilon ranges studied in Table 3. The standard superquadric formulation uses absolute values; without them the probability formula is not well-defined over the full voxel grid.
  5. [Tables 1 and 2 and Sec. 4.2] The same configuration is reported inconsistently across tables: QuadricFormer reports 20.12 mIoU in Table 1 but 20.04 mIoU in Table 2 for 1600 primitives, and GaussianFormer-2 reports 20.02 mIoU in Table 1 but 19.69 mIoU in Table 2 for 12800 primitives. The text states that 1600 superquadrics is the main configuration, but the differing numbers are not explained. Please clarify the protocol differences (e.g., depth initialization, evaluation split, or checkpoint selection) and report the main configuration consistently.
minor comments (6)
  1. [Sec. 2.3] The section heading contains a typo: "Superquarics" should be "Superquadrics".
  2. [Eq. (7)] The scale vector is written as s = (sx, yx, zs)^T; this should be s = (sx, sy, sz)^T.
  3. [Table 4] The IoU value of 39.77 in the row with "Crop & Split Number = 0" appears inconsistent with the surrounding values (30.35-31.22) and is likely a typo; please correct it.
  4. [Sec. 4.4 and Table 4] The parameter is referred to as "Crop & Split Number" in the table and "crop& split number" in the text, but the method is called a pruning-and-splitting module; please unify the terminology and define the parameter precisely.
  5. [Sec. 4.2] The sentence "The numbers of Superquarics are set to 1600" contains a typo and should read "The number of superquadrics is set to 1600".
  6. [Appendix C and Figure 6] Figure 6 is described as a sampled image from a video demo, but the caption and text do not clarify whether it is a static frame or a composite; please state this explicitly.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the superquadric representation and its probabilistic aggregation are specified in-paper and validated empirically; the only self-citation is a non-load-bearing inheritance of the Gaussian mixture framework.

full rationale

The central derivation is self-contained. QuadricFormer explicitly replaces the Gaussian ellipsoidal iso-surface of Eq. 3 with the superquadric surface of Eq. 4, defines each primitive's occupancy probability in Eq. 7, and aggregates primitives via Eq. 8; all equations are stated in the manuscript rather than imported by citation. The efficiency and accuracy claims are empirical comparisons on nuScenes (Tables 1 and 2), and the shape-expressiveness advantage follows from the classical superquadric parameterization of Barr 1981 rather than from any fitted quantity in this paper. The only self-citation is 'Inspired by GaussianFormer-2 [12]' for the probabilistic mixture model; Refs. [12] and [15] share authors with this paper (Zheng and Lu), but the mixture rule is restated explicitly and the novel superquadric geometry prior is independently testable through ablations such as the epsilon-range study in Table 3. The skeptical concern about the noisy-OR independence assumption in Eq. 8 is a legitimate calibration and correctness risk, not a circularity: the assumption is not fitted to the occupancy labels it is used to predict, and no equation or fitted parameter reduces to the paper's own target result. The paper's stated limitation about random initialization leaving superquadrics in empty regions is likewise an acknowledgment of imperfect optimization, not a derivation loop. No circular step can be exhibited, so the paper receives no significant circularity score.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. Superquadrics are established mathematical shapes, and the probabilistic mixture treatment is a modeling choice rather than a new entity. The ledger instead shows that the core representation depends on a couple of untested modeling assumptions and on hyperparameters selected by ablation.

free parameters (3)
  • Shape exponent range for eps1 and eps2 = (0.1, 2)
    Chosen by ablation in Table 3; it directly controls the shapes the superquadrics can express and is central to the representation.
  • Number of superquadrics P = 1600 main, up to 12800
    A hyperparameter swept in Table 2; performance grows with P, so the efficiency claim is tied to this choice.
  • Surface level k in Eq. 4 = Not reported
    Eq. 4 defines an implicit surface family with level k, but Eq. 7 computes occupancy as exp(-f(x_Q)) with no k appearing. Either k is implicitly fixed to 1 or the equations are inconsistent, which changes the resulting occupancy probabilities.
assumptions (4)
  • domain assumption The superquadric implicit surface equation in Eq. 4 is a valid shape model for the scene primitives.
    The entire representation rests on Eq. 4 as the geometry prior, but as written it deviates from the standard superquadric definition in the z term, making the implemented model uncertain.
  • domain assumption Occupancy probabilities of different superquadrics are conditionally independent, as used in Eq. 8.
    The final occupancy is 1 minus the product of complement probabilities, which assumes independence across all primitives; no validation of this assumption is provided.
  • domain assumption Semantic probabilities can be obtained by an occupancy-weighted average of primitive class distributions, as in Eq. 9.
    This mixture rule is inherited from GaussianFormer-2 and is neither derived nor ablated in the paper.
  • standard math Gaussians are the eps1 = eps2 = 1 special case of superquadrics, so superquadrics strictly generalize the ellipsoidal prior.
    This motivates the method, but it only holds for the standard superquadric formula, which Eq. 4 may not actually implement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QuadricFormer: Scene as Superquadrics for 3D Semantic Occupancy Prediction." pith.science (2026). https://pith.science/paper/XWHHTOVA

@misc{pith2026250610977,
  author       = {Pith},
  title        = {Pith review of: QuadricFormer: Scene as Superquadrics for 3D Semantic Occupancy Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XWHHTOVA}},
  note         = {Machine review of arXiv:2506.10977}
}
read the original abstract

3D occupancy prediction is crucial for robust autonomous driving systems as it enables comprehensive perception of environmental structures and semantics. Most existing methods employ dense voxel-based scene representations, ignoring the sparsity of driving scenes and resulting in inefficiency. Recent works explore object-centric representations based on sparse Gaussians, but their ellipsoidal shape prior limits the modeling of diverse structures. In real-world driving scenes, objects exhibit rich geometries (e.g., cuboids, cylinders, and irregular shapes), necessitating excessive ellipsoidal Gaussians densely packed for accurate modeling, which leads to inefficient representations. To address this, we propose to use geometrically expressive superquadrics as scene primitives, enabling efficient representation of complex structures with fewer primitives through their inherent shape diversity. We develop a probabilistic superquadric mixture model, which interprets each superquadric as an occupancy probability distribution with a corresponding geometry prior, and calculates semantics through probabilistic mixture. Building on this, we present QuadricFormer, a superquadric-based model for efficient 3D occupancy prediction, and introduce a pruning-and-splitting module to further enhance modeling efficiency by concentrating superquadrics in occupied regions. Extensive experiments on the nuScenes dataset demonstrate that QuadricFormer achieves state-of-the-art performance while maintaining superior efficiency.

Figures

Figures reproduced from arXiv: 2506.10977 by the authors.

Figure 1
Figure 1. Considering the ellipsoidal shape prior of Gaussians, we propose leveraging expressive [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparisons between different representations. (a) Quadric-based method represents the same object with a smaller number of primitives and greater shape expressiveness. (b) Quadric-based representation outperforms existing methods in both accuracy and speed with far fewer primitives. 2.2 Object-centric scene representations Existing 3D scene representations primarily use voxel-based frameworks for fine-grained volum… view at source ↗
Figure 3
Figure 3. Overall Framework of QuadricFormer.We use several quadric-encoder blocks to update superquadrics, and employ a pruning-and-splitting module to further enhance modeling efficiency. superquadrics allow for continuous and diverse shape variations as the shape parameters change. This inherent parameter efficiency and geometric expressiveness enable superquadrics to model diverse shapes without being densely packed. Cons… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: 3D Superquadrics and occupancy visualizations on nuScenes. Our model is able to predict high-fidelity shapes and achieves comprehensive occupancy results. numbers of Superquarics are set to 1600 in our main results for nuScenes. For optimization, we train our model usi…
Figure 5
Figure 5. Figure 5: Qualitative comparisons. QuadricFormer predicts more flexible and adaptive shapes. Effect of the pruning-splitting module. We conduct ablation studies on the effect of the pruning￾splitting module, as shown in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Visualizations of the proposed QuadricFormer compared to GaussianFormer-2 [12] for 3D semantic occupancy prediction on the nuScenes [3] validation set. We visualize the six surrounding camera inputs, the corresponding occupancy prediction results, and the primitive rep…
Figure 7
Figure 7. Figure 7: Superquadrics of different shape parameters. The figure illustrates how varying ϵ1 and ϵ2 produces a wide range of shapes, from star-like and rounded shapes to square-like structures. Such diversity enables superquadrics to flexibly model complex object geometries in 3…
Figure 8
Figure 8. Figure 8: Visualizations of primitive position distributions learned by different methods. Our approach produces well-structured spatial arrangements while using significantly fewer primitives. C Video Demonstration [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. VISA: VLM-Guided Instance Semantic Auditing for 3D Occupancy World Models

    cs.CV 2026-06 unverdicted novelty 7.0 of 10

    VISA improves closed-set 3D occupancy mIoU on nuScenes by using VLM instance audits as reliability-weighted semantic supervisors during training of existing world models.

  2. DVGT: Driving Visual Geometry Transformer

    cs.CV 2025-12 conditional novelty 6.0 of 10

    DVGT predicts metric-scaled global 3D point maps and ego poses from unposed multi-view driving video, beating prior geometry models on several driving benchmarks.

Reference graph

Works this paper leans on

47 extracted references · 30 canonical work pages · cited by 2 Pith papers

  1. [1]

    Superquadrics and angle-preserving transformations

    Alan H Barr. Superquadrics and angle-preserving transformations. IEEE Computer graphics and Applications, 1(01):11–23, 1981

    work page 1981

  2. [2]

    The lovász-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks

    Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovász-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, pages 4413–4421, 2018

    work page 2018

  3. [3]

    nuscenes: A multimodal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020

    2020

  4. [4]

    Monoscene: Monocular 3d semantic scene completion

    Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022

    work page 2022

  5. [5]

    Scenerf: Self-supervised monocular 3d scene recon- struction with radiance fields

    Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene recon- struction with radiance fields. In ICCV, pages 9387–9398, 2023

    work page 2023

  6. [6]

    3d sketch-aware semantic scene completion via semi-supervised structure prior

    Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193–4202, 2020

    work page 2020

  7. [7]

    Multi-view 3d object detection network for autonomous driving

    Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017

    work page 2017

  8. [8]

    3d semantic scene completion from a single depth image using adversarial training

    Yueh-Tung Chen, Martin Garbade, and Juergen Gall. 3d semantic scene completion from a single depth image using adversarial training. In 2019 IEEE International Conference on Image Processing (ICIP), pages 1835–1839. IEEE, 2019

    work page 2019

Show all 47 references
  1. [9]

    2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network

    Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547–12556, 2021

    2021

  2. [10]

    Su- perdec: 3d scene decomposition with superquadric primitives

    Elisabetta Fedele, Boyang Sun, Leonidas Guibas, Marc Pollefeys, and Francis Engelmann. Su- perdec: 3d scene decomposition with superquadric primitives. arXiv preprint arXiv:2504.00992, 2025

    2025

  3. [11]

    Planning-oriented autonomous driving

    Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, pages 17853–17862, 2023

    2023

  4. [12]

    Probabilistic gaussian superposition for efficient 3d occupancy prediction

    Yuanhui Huang, Amonnut Thammatadatrakoon, Wenzhao Zheng, Yunpeng Zhang, Dalong Du, and Jiwen Lu. Probabilistic gaussian superposition for efficient 3d occupancy prediction. arXiv preprint arXiv:2412.04384, 2024

    2024 arXiv

  5. [13]

    Selfocc: Self- supervised vision-based 3d occupancy prediction

    Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self- supervised vision-based 3d occupancy prediction. In CVPR, pages 19946–19956, 2024

    2024

  6. [14]

    Tri-perspective view for vision-based 3d semantic occupancy prediction

    Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, pages 9223–9232, 2023

    2023

  7. [15]

    Gaussian- former: Scene as gaussians for vision-based 3d semantic occupancy prediction

    Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussian- former: Scene as gaussians for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2405.17429, 2024

    2024 arXiv

  8. [16]

    Segmentation and recovery of superquadrics, volume 20

    Ales Jaklic, Ales Leonardis, and Franc Solina. Segmentation and recovery of superquadrics, volume 20. Springer Science & Business Media, 2000

    2000

  9. [17]

    Pointpillars: Fast encoders for object detection from point clouds

    Alex H Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019

    2019

  10. [18]

    Anisotropic convolutional networks for 3d semantic scene completion

    Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351–3359, 2020. 12

    2020

  11. [19]

    V oxformer: Sparse voxel transformer for camera-based 3d semantic scene completion

    Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. V oxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In CVPR, pages 9087–9098, 2023

    2023

  12. [20]

    Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers

    Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022

    2022

  13. [21]

    Fb-occ: 3d occupancy prediction based on forward-backward view transformation

    Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023

    2023 arXiv

  14. [22]

    Amvnet: Assertion-based multi-view fusion network for lidar semantic segmentation

    Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. Amvnet: Assertion-based multi-view fusion network for lidar semantic segmentation. arXiv preprint arXiv:2012.04934, 2020

    2012 arXiv

  15. [23]

    Octreeocc: Efficient and multi-granularity occupancy prediction using octree queries

    Yuhang Lu, Xinge Zhu, Tai Wang, and Yuexin Ma. Octreeocc: Efficient and multi-granularity occupancy prediction using octree queries. arXiv preprint arXiv:2312.03774, 2023

    2023 arXiv

  16. [24]

    Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction

    Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. InCVPR, pages 19936–19945, 2024

    2024

  17. [25]

    Driveworld: 4d pre-trained scene understanding via world models for autonomous driving

    Chen Min, Dawei Zhao, Liang Xiao, Jian Zhao, Xinli Xu, Zheng Zhu, Lei Jin, Jianshu Li, Yulan Guo, Junliang Xing, et al. Driveworld: 4d pre-trained scene understanding via world models for autonomous driving. In CVPR, pages 15522–15533, 2024

    2024

  18. [26]

    Atlas: End-to-end 3d scene reconstruction from posed images

    Zak Murez, Tarrence van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In ECCV, pages 414–431, 2020

    2020

  19. [27]

    Lmscnet: Lightweight multiscale 3d semantic completion

    Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119, 2020

    2020

  20. [28]

    Occupancy as set of points

    Yiang Shi, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Xinggang Wang. Occupancy as set of points. arXiv preprint arXiv:2407.04049, 2024

    2024 arXiv

  21. [29]

    Searching efficient 3d architectures with sparse point-voxel convolution

    Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In ECCV, pages 685–702, 2020

    2020

  22. [30]

    Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction

    Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In CVPR, pages 15035–15044, 2024

    2024

  23. [31]

    Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving

    Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023

    2023 arXiv

  24. [32]

    Scene as occupancy

    Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, pages 8406–8415, 2023

    2023

  25. [33]

    Opus: occupancy prediction using a sparse set

    Jiabao Wang, Zhaojiang Liu, Qiang Meng, Liujiang Yan, Ke Wang, Jie Yang, Wei Liu, Qibin Hou, and Ming-Ming Cheng. Opus: occupancy prediction using a sparse set. arXiv preprint arXiv:2409.09350, 2024

    2024 arXiv

  26. [34]

    Occsora: 4d occupancy generation models as world simulators for autonomous driving

    Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024

    2024 arXiv

  27. [35]

    Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception

    Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 13

    2023 arXiv

  28. [36]

    Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving

    Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, pages 21729–21740, 2023

    2023

  29. [37]

    Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion

    Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, volume 35, pages 3101–3109, 2021

    2021

  30. [38]

    Renderworld: World model with self-supervised 3d label

    Ziyang Yan, Wenzhen Dong, Yihua Shao, Yuhang Lu, Liu Haiyang, Jingwen Liu, Haozhe Wang, Zhe Wang, Yan Wang, Fabio Remondino, et al. Renderworld: World model with self-supervised 3d label. arXiv preprint arXiv:2409.11356, 2024

    2024 arXiv

  31. [39]

    Lidarmultinet: Towards a unified multi-task network for lidar perception

    Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022

    2022 arXiv

  32. [40]

    Drinet++: Efficient voxel-as-point point cloud segmentation

    Maosheng Ye, Rui Wan, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Drinet++: Efficient voxel-as-point point cloud segmentation. arXiv preprint arXiv: 2111.08318 , 2021

    2021 arXiv

  33. [41]

    Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin

    Zichen Yu, Changyong Shu, Jiajun Deng, Kangjie Lu, Zongdai Liu, Jiangyong Yu, Dawei Yang, Hui Li, and Yan Chen. Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058, 2023

    2023 arXiv

  34. [42]

    Occformer: Dual-path transformer for vision- based 3d semantic occupancy prediction

    Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision- based 3d semantic occupancy prediction. In ICCV, pages 9433–9443, 2023

    2023

  35. [43]

    Occworld: Learning a 3d occupancy world model for autonomous driving

    Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In ECCV, 2024

    2024

  36. [44]

    Gaussianad: Gaussian-centric end-to-end autonomous driving

    Wenzhao Zheng, Junjie Wu, Yao Zheng, Sicheng Zuo, Zixun Xie, Longchao Yang, Yong Pan, Zhihui Hao, Peng Jia, Xianpeng Lang, et al. Gaussianad: Gaussian-centric end-to-end autonomous driving. arXiv preprint arXiv:2412.10371, 2024

    2024 arXiv

  37. [45]

    V oxelnet: End-to-end learning for point cloud based 3d object detection

    Yin Zhou and Oncel Tuzel. V oxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, pages 4490–4499, 2018

    2018

  38. [46]

    Pointocc: Cylin- drical tri-perspective view for point-based 3d semantic occupancy prediction

    Sicheng Zuo, Wenzhao Zheng, Yuanhui Huang, Jie Zhou, and Jiwen Lu. Pointocc: Cylin- drical tri-perspective view for point-based 3d semantic occupancy prediction. arXiv preprint arXiv:2308.16896, 2023

    2023 arXiv

  39. [47]

    Gaussianworld: Gaussian world model for streaming 3d occupancy prediction.arXiv preprint arXiv:2412.10373, 2024

    Sicheng Zuo, Wenzhao Zheng, Yuanhui Huang, Jie Zhou, and Jiwen Lu. Gaussianworld: Gaussian world model for streaming 3d occupancy prediction.arXiv preprint arXiv:2412.10373, 2024. 14

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.