Pith. sign in

REVIEW 5 major objections 5 minor 36 references

GaussianFusionOcc: A Seamless Sensor Fusion Approach for 3D Occupancy Prediction Using 3D Gaussians

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read GaussianFusionOcc claims the first multi-modal 3D Gaussian occupancy predictor, reporting 45.20 IoU and 30.37 mIoU on nuScenes using camera, LiDAR, and radar with fewer parameters than leading fusion baselines.

desk verdict Solid multi-modal occupancy architecture with real efficiency gains, but the SOTA claim rests on baseline numbers that may not be protocol-matched. read the letter →

arxiv 2507.18522 v1 pith:M35FU5WP submitted 2025-07-24 cs.CV

Tomislav Pavković , Mohammad-Ali Nikouei Mahani , Johannes Niedermayer , Johannes Betz This is my paper
classification cs.CV
keywords 3DsemanticoccupancypredictionGaussiansplattingmulti-sensorfusioncamera-LiDAR-radardeformableattentionnuScenesautonomousdrivingGaussian-to-voxel
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 sets out to show that 3D semantic occupancy prediction for autonomous driving can be both more accurate and cheaper when the scene is represented as a set of semantic 3D Gaussians rather than a dense voxel grid, and when camera, LiDAR, and radar features are fused per Gaussian. It claims that its model, GaussianFusionOcc, is the first multi-modal 3D Gaussian occupancy method and that it outperforms existing voxel-, planar-, and Gaussian-based models on the nuScenes validation set, with particular gains in rainy and nighttime conditions. It also claims large efficiency advantages: 79.96 million parameters and 2.90 GB memory for the three-sensor model, versus 114.97 million and 5.78 GB for the strongest fusion baseline. If these claims hold, a perception stack could get higher-quality occupancy at a fraction of the compute and memory cost, which matters for real-time autonomous driving.

What carries the argument

The central object is the semantic 3D Gaussian, a primitive $G = (m, s, r, a, c)$ carrying mean, scale, rotation, opacity, and semantic logits, refined over four Gaussian fusion blocks. Each block uses a modality-agnostic Gaussian encoder, built on deformable attention, that samples 3D reference points around each Gaussian center, projects them onto each sensor's feature maps, and aggregates per-Gaussian features. The fusion module concatenates those per-modality feature vectors, passes them through a multilayer perceptron, and adds sparse-convolution context computed from the voxelized Gaussian means; the resulting unified vector drives an MLP that predicts refined Gaussian properties. At the end, probabilistic Gaussian-to-voxel splatting converts the refined Gaussians into the occupancy grid. This machinery is what carries both the accuracy and efficiency claims, because computation concentrates on occupied Gaussians rather than on every voxel.

What would settle it

Retrain OccFusion and M-CONet under exactly the SurroundOcc dense-label recipe used for GaussianFusionOcc, with the same voxel size, training schedule, and without LiDAR-initialized Gaussian means in the camera-only row; if OccFusion C+L+R then matches or exceeds 45.20 IoU and 30.37 mIoU, the central superiority claim falls. A simpler check is to remove LiDAR initialization from the camera-only GaussianFusionOcc row and see whether its 37.05 IoU in Table 1 still holds.

Watch

Extended reading notes

Core claim

The paper reports that GaussianFusionOcc with camera, LiDAR, and radar reaches 45.20 IoU and 30.37 mIoU on the nuScenes validation set using SurroundOcc dense occupancy labels, exceeding OccFusion's 44.66 IoU and 27.30 mIoU and M-CONet's 39.20 IoU and 24.70 mIoU while using fewer parameters and less memory. On the rainy subset it reports 44.36 IoU and 29.86 mIoU for the full three-sensor model, and on the nighttime subset 42.51 IoU and 18.45 mIoU, with the largest category-level gains coming from adding LiDAR and radar. The paper interprets these results as showing that an object-centric Gaussian representation combined with a modality-agnostic fusion mechanism is a new and better operating point on the accuracy-efficiency curve for multi-modal occupancy prediction.

Load-bearing premise

The paper's state-of-the-art claim assumes the baseline numbers in Table 1 were obtained under training and evaluation conditions comparable enough to GaussianFusionOcc; if the dense-label protocol, voxel resolution, or LiDAR initialization differ materially, the reported margins may not be real.

Editorial extensions

If this is right

  • If the reported numbers hold, GaussianFusionOcc sets a new high-water mark for multi-modal 3D semantic occupancy prediction on nuScenes, at lower parameter and memory cost than grid-based fusion methods.
  • The three-sensor configuration would be the first Gaussian-based occupancy model to use camera, LiDAR, and radar together, with the paper showing that adding radar helps most for dynamic-object classes at night.
  • The efficiency results imply that Gaussian representations can bring multi-modal occupancy prediction closer to real time, with 480 ms inference for the full sensor set and 179 ms for LiDAR-only on an RTX A6000.
  • The ablations suggest that raising the number of Gaussians or feature channels improves accuracy with only mild efficiency loss, giving a direct capacity knob for different deployment budgets.

Reading between the lines

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

  • Editorial inference: the state-of-the-art comparison would be cleaner if every baseline were retrained under the same SurroundOcc dense-label protocol and if the camera-only GaussianFusionOcc row did not use LiDAR-initialized Gaussian means; the 0.54-point IoU gap over OccFusion could be sensitive to those protocol differences.
  • Editorial inference: the radar contribution is uneven, since the full three-sensor model's nighttime mIoU is slightly below the camera-plus-LiDAR model's, so radar fusion may need a mechanism that better exploits its sparse velocity information for dense prediction.
  • Editorial inference: the fixed number of Gaussians, which the authors name as a limitation, is the natural next target; dynamic pruning and densification would test whether representing sparse scenes more flexibly removes the remaining efficiency-accuracy trade-off.
  • Editorial inference: the claimed leadership is tied to the nuScenes/SurroundOcc evaluation, and transferring the same recipe to other dense-occupancy benchmarks would show whether the method generalizes or overfits to this label distribution.
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 / 5 minor

Summary. The paper proposes GaussianFusionOcc, a 3D semantic occupancy prediction framework that represents the scene with semantic 3D Gaussians and fuses camera, LiDAR, and radar features via modality-agnostic deformable attention. On the nuScenes validation set, the camera+LiDAR+radar model is reported to achieve 45.20 IoU and 30.37 mIoU, which the authors claim surpasses existing methods such as OccFusion (44.66 IoU, 27.30 mIoU) and M-CONet while using fewer parameters (79.96M) and less memory (2.90 GB). The paper also studies performance on rainy and nighttime subsets, reports efficiency comparisons, and includes ablations on the number of Gaussians, feature channels, and initialization strategy.

Significance. If the headline comparisons are protocol-matched, this is a meaningful contribution: it is the first multi-modal 3D Gaussian occupancy prediction method, the architecture is coherent, and the efficiency numbers are attractive. The paper deserves credit for testing many sensor combinations and for including ablations over initialization and capacity. However, the central state-of-the-art claim currently rests on baseline numbers whose training protocol is not fully matched, on a camera-only row that uses LiDAR initialization, and on a small radar benefit that is not consistently supported by the reported night-table numbers. The contribution is therefore plausible but not yet established at the level claimed.

major comments (5)
  1. [Table 1 and Section 4.3] The headline claim that GaussianFusionOcc surpasses OccFusion and M-CONet depends on Table 1 baseline numbers being measured under the same protocol (same dense occupancy labels, voxel range, class set, and training schedule). The table only footnotes TPVFormer* for protocol differences; it does not state whether OccFusion, M-CONet, L-CONet, or the other baselines were retrained under the SurroundOcc dense labels and the same schedule. Since Tables 2 and 3 explicitly state that all methods were trained with dense occupancy labels from [9], the absence of such a statement for Table 1 is conspicuous. The authors must either clarify that the Table 1 baselines are protocol-matched, or retrain/report matched baselines before claiming SOTA.
  2. [Table 1 footnote ** and Section 4.3] The camera-only row GaussianFusionOcc** uses LiDAR initialization of Gaussian means, as stated by the footnote. Comparing this row with genuine camera-only baselines such as GaussianFormer-2 and SurroundOcc is therefore unfair and inflates the apparent improvement. A true camera-only variant should use random or learnable initialization without any LiDAR-derived means; this result should be reported and used for the comparison.
  3. [Table 4 and Section 4.5] The efficiency comparison is not apples-to-apples. The OccFusion latency values are explicitly marked as taken from the original paper and measured on a different GPU, yet the text states that GaussianFusionOcc 'shows significantly faster inference than compared models' (480 ms vs 588-601 ms). Memory and parameter counts may also come from different configurations. The authors should measure all methods on the same hardware and report the setup, or clearly qualify that the latency comparison is indicative only.
  4. [Section 4.4 and Table 3] The text claims that adding radar improves nighttime performance 'especially for dynamic object categories such as cars, bicycles, and motorcycles,' but Table 3 shows the opposite for the overall night metrics: C+L+R mIoU is 18.45 versus 18.66 for C+L, IoU is 42.51 versus 42.78, and the car category is 39.77 versus 39.82. Only the motorcycle class improves with radar (29.58 vs 27.66). This overclaim should be corrected, and the night-subset analysis should be revised to reflect the actual per-class results.
  5. [Section 4.3 and Tables 1, 5-7] No error bars, multiple seeds, or code/checkpoints are provided. Some of the key comparisons are small (e.g., radar over C+L is +0.16 mIoU in Table 1; learnable versus random initialization in Table 7 differs by 0.30 mIoU), so it is difficult to assess whether these differences are stable. Reporting at least 2-3 seeds with standard deviations, or releasing code and checkpoints, would substantially strengthen the empirical claims.
minor comments (5)
  1. [Section 1] There is a typo in the first paragraph: 'situation3al awareness' should be 'situational awareness'.
  2. [Section 3] The text contains frequent spacing typos in method names, e.g., 'V oxelNet', 'V oxel', and 'PointPillars [31] voxel encoder' should be 'VoxelNet', 'voxel', and 'PointPillars'. These should be corrected throughout.
  3. [Appendix B, Table 7] The table header says 'Initinalization' instead of 'Initialization'.
  4. [Equation (1)] The notation 'Φof f set, qdenote' is missing a space and likely a comma; it should read 'Φ_offset, q denote' for clarity.
  5. [Section 5, Limitations] The limitations paragraph contains typos: 'extremly' should be 'extremely', 'introuducing' should be 'introducing', and 'fuse its features' should be 'fuse them'. The paragraph itself is a useful self-assessment, but the manuscript would benefit from proofreading.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is benchmark performance against external baselines, not a derivation that reduces to its own inputs.

full rationale

GaussianFusionOcc is an empirical architecture paper. Its central claim is measured 3D occupancy prediction accuracy on the nuScenes validation set compared with external methods such as OccFusion, M-CONet, and GaussianFormer-2. The equations in Section 3 define modular components (deformable attention feature extraction, concatenation and MLP fusion, Gaussian refinement, Gaussian-to-voxel splatting, and loss terms), but no equation defines the reported IoU or mIoU in terms of fitted values of the same quantities. The Gaussian-to-voxel splatting module is explicitly cited from GaussianFormer-2, an external group's work, and is not used to derive the paper's own performance advantage. There is no self-citation chain that carries the state-of-the-art claim, and the ablations in Appendix B are internally consistent but are not presented as independent predictions. Concerns about comparability of baselines, such as the Table 1 footnote that TPVFormer* was retrained on dense occupancy annotations and the Table 4 note that OccFusion latency was measured on a different GPU, are validity and fairness risks for the benchmark comparison, not circularity. Accordingly, no step in the paper reduces by construction to its own inputs, and the appropriate finding is no significant circularity.

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

The paper's central claim depends on domain assumptions about label quality and baseline comparability, not on mathematical axioms. The only hand-set parameters are architectural hyperparameters; there are no fitted physical constants and no invented entities.

free parameters (4)
  • Number of Gaussians = 6400 in main tables; 25600 in ablation
    Hand-set hyperparameter controlling the accuracy-memory trade-off; Table 5 shows 25600 Gaussians improve mIoU from 30.21 to 30.83 with modest memory change.
  • Feature channels = 128; 192 in ablation
    Hand-set hyperparameter for per-Gaussian features; Table 6 shows 192 channels improve mIoU from 30.37 to 30.85 with higher memory use.
  • Number of Gaussian fusion blocks = 4
    Architectural depth chosen without a dedicated ablation in the paper.
  • Gaussian initialization strategy = learnable
    Default strategy; Table 7 shows random and probabilistic variants change mIoU by less than 0.3 points, so the choice is not strongly constrained by data.
assumptions (3)
  • domain assumption SurroundOcc-generated dense occupancy labels are treated as ground truth for training and evaluation.
    Section 4.1 states nuScenes lacks native dense occupancy ground truth, so SurroundOcc annotations are used. If these labels are noisy or biased, all IoU and mIoU numbers shift.
  • domain assumption Baseline numbers in Table 1 from prior papers are directly comparable to GaussianFusionOcc, despite mixed training-label protocols (e.g., TPVFormer* footnote).
    Section 4.3 and Table 1 compare methods whose supervision differs; the paper only footnotes one model. Comparability is load-bearing for the state-of-the-art claim.
  • domain assumption The Gaussian-to-voxel splatting from GaussianFormer-2 is an unbiased way to convert sparse Gaussians into dense occupancy and does not introduce artifacts that favor the proposed model.
    Eq. (5) uses the splatting module from reference [13] without modification; the accuracy claim assumes this splatting is adequate for all classes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GaussianFusionOcc: A Seamless Sensor Fusion Approach for 3D Occupancy Prediction Using 3D Gaussians." pith.science (2026). https://pith.science/paper/M35FU5WP

@misc{pith2026250718522,
  author       = {Pith},
  title        = {Pith review of: GaussianFusionOcc: A Seamless Sensor Fusion Approach for 3D Occupancy Prediction Using 3D Gaussians},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M35FU5WP}},
  note         = {Machine review of arXiv:2507.18522}
}
read the original abstract

3D semantic occupancy prediction is one of the crucial tasks of autonomous driving. It enables precise and safe interpretation and navigation in complex environments. Reliable predictions rely on effective sensor fusion, as different modalities can contain complementary information. Unlike conventional methods that depend on dense grid representations, our approach, GaussianFusionOcc, uses semantic 3D Gaussians alongside an innovative sensor fusion mechanism. Seamless integration of data from camera, LiDAR, and radar sensors enables more precise and scalable occupancy prediction, while 3D Gaussian representation significantly improves memory efficiency and inference speed. GaussianFusionOcc employs modality-agnostic deformable attention to extract essential features from each sensor type, which are then used to refine Gaussian properties, resulting in a more accurate representation of the environment. Extensive testing with various sensor combinations demonstrates the versatility of our approach. By leveraging the robustness of multi-modal fusion and the efficiency of Gaussian representation, GaussianFusionOcc outperforms current state-of-the-art models.

Figures

Figures reproduced from arXiv: 2507.18522 by the authors.

Figure 1
Figure 1. GaussianFusionOcc pipeline: Our method utilizes modality-specific encoders to extract feature maps from the input data. Extracted features are then fed to GaussianFusion blocks along with Gaussians and queries from the previous block. The GaussianFusion block, visualized in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. GaussianFusionBlock architecture. Modality-agnostic Gaussian encoders employ a deformable attention mechanism to extract relevant features for each initial Gaussian. These per-Gaussian features are then fed to the fusion module along with the resulting features of the sparse convolution layer applied to voxelized Gaussians. Fusion module concatenates Gaussian encoder features, applies an MLP fuser, and adds sparse c… view at source ↗
Figure 3
Figure 3. Occupancy visualizations on nuScenes. Our model is able to predict both comprehensive and realistic semantic 3D occupancy. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Occupancy visualizations on nuScenes night scenes. Visualization shows the importance of additional sensors in low-light conditions. significant efficiency degradation. The increase in memory and parameter number can be attributed to the increased size of sensor featur…
Figure 5
Figure 5. Figure 5: Occupancy and Gaussian representation visualizations on nuScenes dataset. Visualizations demonstrate the importance of additional sensors for the allocation of the Gaussians and final occupancy prediction. B.1 Evaluation metrics To evaluate our method and compare the r…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 25 canonical work pages

  1. [9]

    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 Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729–21740, 2023

    2023

  2. [1]

    Occfusion: Multi-sensor fusion framework for 3d semantic occupancy prediction

    Zhenxing Ming, Julie Stephany Berrio, Mao Shan, and Stewart Worrall. Occfusion: Multi-sensor fusion framework for 3d semantic occupancy prediction. IEEE Transactions on Intelligent Vehicles, 2024

    work page 2024

  3. [2]

    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. Advances in Neural Information Processing Systems, 37:79618–79641, 2024

    work page 2024

  4. [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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020

    2020

  5. [4]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850–17859, 2023

    work page 2023

  6. [5]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087–9098, 2023

    work page 2023

  7. [6]

    Monoscene: Monocular 3d semantic scene completion

    Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022

    work page 2022

  8. [7]

    Symphonize 3d semantic scene completion with contextual instance queries

    Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Tianwei Lin, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20258–20267, 2024

    work page 2024

Show all 36 references
  1. [8]

    Occdepth: A depth-aware method for 3d semantic scene completion

    Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 9

    2023 arXiv

  2. [10]

    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 Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433–9443, 2023

    2023

  3. [11]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9223–9232, 2023

    2023

  4. [12]

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

    Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In European Conference on Computer Vision, pages 376–393. Springer, 2024

    2024

  5. [13]

    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

  6. [14]

    Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation

    Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774–2781. IEEE, 2023

    2023

  7. [15]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), July 2023

    2023

  8. [16]

    Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20331–20341, 2024

    2024

  9. [17]

    Bevformer: learning bird’s-eye-view representation from lidar-camera 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 lidar-camera via spatiotemporal transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  10. [18]

    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

  11. [19]

    Gaussianocc: Fully self-supervised and efficient 3d occupancy estimation with gaussian splatting, 2024

    Wanshui Gan, Fang Liu, Hongbin Xu, Ningkai Mo, and Naoto Yokoya. Gaussianocc: Fully self-supervised and efficient 3d occupancy estimation with gaussian splatting, 2024

    2024

  12. [20]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,...

    2024

  13. [21]

    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

  14. [22]

    Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection

    Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conferenc...

    2023

  15. [23]

    Futr3d: A unified sensor fusion framework for 3d detection

    Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 172–181, 2023

    2023

  16. [24]

    Craft: Camera-radar 3d object detection with spatio- contextual fusion transformer

    Youngseok Kim, Sanmin Kim, Jun Won Choi, and Dongsuk Kum. Craft: Camera-radar 3d object detection with spatio- contextual fusion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 1160–1168, 2023

    2023

  17. [25]

    Centerfusion: Center-based radar and camera fusion for 3d object detection

    Ramin Nabati and Hairong Qi. Centerfusion: Center-based radar and camera fusion for 3d object detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1527–1536, 2021

    2021

  18. [26]

    Deformable detr: Deformable transformers for end-to-end object detection

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020

    2010 arXiv

  19. [27]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  20. [28]

    Deformable convolutional networks

    Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764–773, 2017

    2017

  21. [29]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117–2125, 2017. 10

    2017

  22. [30]

    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. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4490–4499, 2018

    2018

  23. [31]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697–12705, 2019

    2019

  24. [32]

    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 Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4413–...

    2018

  25. [33]

    Fcos3d: Fully convolutional one-stage monocular 3d object detection

    Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 913–922, 2021

    2021

  26. [34]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  27. [35]

    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 European conference on computer vision, pages 414–431. Springer, 2020. A Additional visualizations The visuali...

    2020

  28. [36]

    Probabilistic initialization, proposed by GaussianFormer-2 [13], degrades the performance of the model while significantly slowing down the inference

    Learnable initialization shows the highest mIoU with a slightly higher number of parameters, memory usage, and latency, compared to random initialization. Probabilistic initialization, proposed by GaussianFormer-2 [13], degrades the performance of the model while significantly...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.