Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Latent Gaussian Splatting for 4D Panoptic Occupancy Tracking

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Sparse Gaussians beat dense voxels in 4D occupancy tracking

desk verdict LaGS offers a genuinely new latent Gaussian encoder for 4D-POT, but the headline SOTA margins are provisional until verified under the original TrackOcc metric and against non-reimplemented baselines. read the letter →

arxiv 2602.23172 v2 pith:U7FI6JM3 submitted 2026-02-26 cs.CV cs.AIcs.RO

Maximilian Luz , Rohit Mohan , Thomas Nürnberg , Yakov Miron , Daniele Cattaneo , Abhinav Valada This is my paper
classification cs.CVcs.AIcs.RO
keywords 4DpanopticoccupancytrackinglatentGaussiansplatting3Dpredictionmulti-viewcamerastracking-by-attentionsparsepointrepresentationmask-basedsegmentationautonomousdriving
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

This paper claims that the dense voxel feature grid — the standard intermediate representation for 3D occupancy prediction — is a bottleneck for camera-based 4D panoptic occupancy tracking, and that a sparse set of feature-bearing 3D Gaussians is a better alternative. The authors propose Latent Gaussian Splatting (LaGS): treat Gaussians as dynamic, volume-oriented keypoints, aggregate multi-view image features onto them, let them exchange information over long ranges through point-transformer attention, then splat their features back into a voxel grid for decoding. On the Occ3D nuScenes and Waymo benchmarks, LaGS is reported to beat all re-implemented baselines, with gains up to +18.9 percentage points in Segmentation and Tracking Quality (STQ) and +19.8 points in Association Quality (AQ) on nuScenes. The paper also argues that the existing TrackOcc evaluation metric is flawed because it ignores false positives in free space, and provides a corrected implementation. If these results hold, sparse point-centric encoders could displace dense voxel encoders for dense 3D prediction tasks.

What carries the argument

Latent Gaussian Splatting — a sparse set of 3D Gaussians (center, covariance, opacity, feature vector) used as feature-bearing keypoints, splatted back into a voxel grid. The key operation is Eq. (2): the feature at voxel x is the opacity-weighted mean of nearby Gaussian features, giving a differentiable way to convert a sparse point representation into a dense volume for decoding. The work also introduces Serialized Multi-Stream Attention (SMSA) to merge coarse and fine Gaussian streams, and a split aggregation rule for thing/stuff masks. The Gaussians act as dynamic, data-dependent neighborhoods, allowing windowed attention over ~1000 points versus fixed local neighborhoods in dense voxel

What would settle it

Run the original TrackOcc metric implementation (as released, ignoring free-space false positives) on LaGS's predictions for the Occ3D-Waymo validation split and compare against the TrackOcc numbers reported in the original paper. If LaGS's STQ/AQ advantage under the original metric is close to zero, the claimed improvement is an artifact of the metric correction rather than of the method.

Watch

Extended reading notes

Core claim

The central discovery is that using 3D Gaussians as an intermediate sparse feature representation — rather than as output primitives — improves both accuracy and efficiency for 4D panoptic occupancy tracking. Each Gaussian carries a center, covariance, opacity, and feature vector; multi-view image features are lifted into these points via spatial cross-attention, and the points are refined with windowed self-attention over a space-filling curve. Hierarchical streams (512 coarse 'super-points' and 8192 fine points) are merged with a Serialized Multi-Stream Attention operation that allows the streams to exchange information. The refined Gaussian features are then splatted onto a voxel grid thr

Load-bearing premise

The state-of-the-art claim rests on the authors' own re-implementations of all baselines and on a corrected evaluation metric; if those re-implementations are weaker than the original published systems, or if the corrected metric systematically disadvantages the baselines, the headline margins are not evidence of superiority over the original systems.

Editorial extensions

If this is right

  • The latent Gaussian encoder is a drop-in replacement for voxel feature encoders in any occupancy or panoptic occupancy pipeline, as the paper states in its conclusion.
  • Camera-based 4D-POT can narrow the gap to stereo/temporal methods: LaGS improves mIoU by +4.9 over the methodologically closest baseline (BEVDet4D+COTR without stereo/long-term), and the paper notes that restoring those would likely close the gap entirely.
  • The corrected 4D-POT metric (counting false positives in free space) is necessary for fair comparison; the paper provides implementations, so future work should adopt it.
  • Multi-layer decoders and spatio-temporal query refinement are both beneficial and compatible with tracking-by-attention, contrary to earlier single-layer designs.
  • Separate aggregation of thing and stuff masks is a simple, effective way to counter mask imbalance in panoptic segmentation.

Reading between the lines

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

  • The reported gains are measured against re-implemented baselines under a corrected metric; the paper does not report LaGS under the original TrackOcc metric, so the margin over the original published TrackOcc is not directly established.
  • If the Gaussian keypoint representation generalizes, it could be applied to other dense 3D prediction tasks that suffer from the same dense-volume bottleneck, such as semantic scene completion or 4D forecasting, where the same splatting-to-volume trick would convert point features to any required grid.
  • The number of Gaussians (512 coarse / 8192 fine) is fixed; a data-dependent allocation of points (e.g., more points in crowded scenes) is a natural extension that could improve the efficiency/accuracy trade-off further.
  • The split thing/stuff aggregation suggests a general design pattern for mask-based panoptic heads beyond occupancy: treat local instance masks and global semantic masks as separate matching problems before merging.
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

4 major / 6 minor

Summary. The paper introduces Latent Gaussian Splatting (LaGS), a camera-based 4D panoptic occupancy tracking method. LaGS replaces dense voxel feature encoders with a sparse set of feature-bearing 3D Gaussians, processed hierarchically in two streams and splatted back to a voxel grid for mask decoding. Tracking is done by propagating decoder queries across frames with a PF-Track-style refinement module. The authors evaluate on Occ3D nuScenes and Waymo, claiming state-of-the-art 4D-POT performance with up to +18.9 p.p. STQ and +19.8 p.p. AQ on nuScenes and +5.1/+7.9 on Waymo. They also revise the TrackOcc evaluation metric, arguing the original implementation ignores false positives in free space, and they provide new nuScenes 4D-POT label generation based on box-to-voxel instance assignment. The manuscript includes ablations of the latent Gaussian encoder, decoder depth, and mask aggregation strategy.

Significance. The core architectural idea is interesting and timely: modeling intermediate 3D features as latent Gaussians allows a sparse, point-centric transformer encoder with data-dependent receptive fields, while retaining a dense voxel output for mask decoding. If the reported results are robust, LaGS advances the state of the art in a relatively new task and also contributes a corrected evaluation protocol. The paper further provides a nuScenes 4D-POT benchmark and public code, which are useful contributions. However, the SOTA claim is currently conditional: it relies on the authors' revised metric and on re-implemented baselines, and LaGS is not evaluated under the original TrackOcc metric. The empirical margin is large, but the benchmark yardstick is partly self-constructed, so the headline conclusions need additional verification before they can be fully trusted.

major comments (4)
  1. [Section IV-B, Tables I-II] The paper revises the 4D-POT metric and evaluates all methods, including LaGS, only under this corrected metric; LaGS is never evaluated under the original TrackOcc metric implementation. This matters because the corrected metric changes TrackOcc's Waymo scores from 20.2 STQ / 13.8 AQ to 15.2 / 10.7 (Table II, gray vs normal rows). The headline improvements in Section IV-E (+5.1 STQ / +7.9 AQ on Waymo) are relative to the corrected TrackOcc values, so they do not establish superiority under the original benchmark definition used by previous work. To support the SOTA claim, please report LaGS and all baselines under both metric implementations, and discuss the effect of the correction on each method. If the original implementation is indeed flawed, this should be demonstrated on all methods, not only on the re-implemented baselines.
  2. [Section IV-C/D, Tables I-II] The SOTA comparison relies almost entirely on baselines re-implemented by the authors. For nuScenes, TrackOcc itself is 'official code adapted for nuScenes'; for Waymo, only TrackOcc uses official code and weights, while MinVIS, CTVIS, 4D-LCA, and AB3DMOT are reconstructed. In addition, all baselines are trained on a 'single-frame-adapted version of TrackOcc'. This design may be internally consistent, but it makes the claimed margins sensitive to the fidelity of the re-implementations. Please provide evidence that the re-implementations are faithful: e.g., report original published numbers (where available) alongside your reproductions under the same metric, or run the official TrackOcc code on nuScenes and compare. Without this, an external reader cannot separate method gains from baseline degradation.
  3. [Section IV-A] The nuScenes 4D-POT ground truth is constructed by assigning instance IDs to Occ3D semantic voxels via maximum-overlap box intersection, with ambiguities resolved by 'choosing the closest instance.' This is a heuristic that could systematically mislabel voxels belonging to large or deformed objects or at object boundaries, and it defines a benchmark on which LaGS claims large gains. Please include a quantitative and qualitative validation of these labels (e.g., fraction of ambiguous voxels, distribution of closest-instance assignments, and visual examples), and discuss how label noise might affect the ranking of methods. The paper would also benefit from a clear statement that the nuScenes 4D-POT benchmark is introduced in this work, so results are not directly comparable to prior published numbers.
  4. [Section IV-E and Table I] The stated improvements, 'up to +18.9p.p. STQ and +19.8p.p. AQ for nuScenes', do not match the differences in Table I under the corrected metric: LaGS vs TrackOcc gives +19.0/+19.9, and vs the best baseline AB3DMOT gives +18.1/+19.3. Please clarify which baseline and which metric are used for these headline numbers, and ensure consistency across the abstract, introduction, and results. This is not merely cosmetic: the margin is the basis of the SOTA contribution.
minor comments (6)
  1. [Figure 2] The architecture figure is dense and the component labels are very small. Consider enlarging or splitting into two figures for readability.
  2. [Section III-F] The sequence length T used in multi-frame training is never specified. Please state the exact value used in the experiments.
  3. [Tables I-II] The top and bottom blocks of Table II would be much clearer with explicit column headers indicating the metric version (original TrackOcc vs corrected). Currently this is only conveyed by gray shading and a footnote.
  4. [Eq. (1)] The Mahalanobis distance notation ||x-µ||^2_{\Sigma^{-1}} is used before it is defined. Please move the definition of ||v||^2_M ahead of the first use.
  5. [Section IV-D] The ablation table (Table IV) would benefit from a sentence explaining why COTR is used with one layer as the primary comparison, and why the number of points (512 coarse, 8192 fine) is chosen.
  6. [References] Several references are dated 2026, including [19]. Please make sure all citations are complete and, where appropriate, provide arXiv identifiers or publication venues.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; LaGS's SOTA claim is an empirical benchmark result on external datasets, not a definitional reduction.

full rationale

LaGS is an empirical systems paper. The central claim is benchmarking performance on externally provided Occ3D nuScenes/Waymo validation sets, not a mathematical derivation that reduces to its own inputs. The only candidate concern is Section IV-B, where the authors correct the TrackOcc metric and re-implement all baselines. This is not circular: the corrected AQ is defined over ground-truth and predicted voxel/instance tubes (Eq. 4), LaGS's trained weights are not inputs to that definition, and the same metric is applied to every method. Likewise, the nuScenes 4D panoptic labels are an annotation protocol derived from Occ3D semantic labels and box instances, not a prediction produced by LaGS. Self-citations such as [2], [16], [19], and [31] are contextual, dataset, or related-work attributions, and none is load-bearing for the state-of-the-art claim. The headline SOTA result does carry a reproducibility/evaluation risk because the metric is revised and baselines are self-reimplemented, but that is an evaluation comparability concern, not circularity by construction.

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

The paper introduces no new physical entities. The latent Gaussians are a representational choice, not a new conserved quantity or force. The main assumptions are the validity of the Occ3D labels and camera geometry, the soundness of the authors' heuristic panoptic label generation on nuScenes, and the correctness of their metric revision.

free parameters (4)
  • number of fine stream points = 8192
    Selected via ablation on the validation set (Table IV); larger sets improved performance up to a point.
  • number of coarse stream points = 512
    Selected via ablation; the coarse stream is intended to make full attention feasible.
  • attention window size = 1024
    Chosen to balance memory and receptive field; not derived from first principles.
  • encoder/decoder transformer layers = 4
    Ablation (Table V) shows 4 layers outperform 1; chosen configuration for the final model.
assumptions (3)
  • domain assumption Ground-truth occupancy labels from Occ3D are correct and camera intrinsic/extrinsic plus ego-motion are known.
    Section III-A states that camera parameters and ego-motion are assumed known; the entire pipeline depends on accurate sensor calibration and ground-truth labels.
  • ad hoc to paper The nuScenes 4D-POT labels generated by assigning instance IDs via maximum-overlap box intersection are a valid proxy for true instance identities.
    Section IV-A describes a heuristic assignment to boxes; voxels not intersecting any box are assigned to the closest instance. This is a paper-specific annotation procedure with no external validation.
  • ad hoc to paper The revised AQ metric (Eq. 4) correctly measures 4D panoptic tracking quality and the TrackOcc implementation is flawed.
    Section IV-B argues that the original TrackOcc metric ignores free-space false positives. The re-evaluation and all subsequent comparisons depend on this metric choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent Gaussian Splatting for 4D Panoptic Occupancy Tracking." pith.science (2026). https://pith.science/paper/U7FI6JM3

@misc{pith2026260223172,
  author       = {Pith},
  title        = {Pith review of: Latent Gaussian Splatting for 4D Panoptic Occupancy Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U7FI6JM3}},
  note         = {Machine review of arXiv:2602.23172}
}
read the original abstract

Capturing 4D spatiotemporal scene structure is crucial for the safe and reliable operation of robots in dynamic environments. However, existing approaches typically address only part of the problem: they either provide coarse geometric tracking via bounding boxes or detailed 3D occupancy estimates that lack explicit temporal association and instance-level reasoning. In this work, we present Latent Gaussian Splatting (LaGS) for 4D Panoptic Occupancy Tracking (4D-POT). We revisit the underlying representation and model 3D features as a sparse set of feature-bearing Gaussians. These act as dynamic, volume-oriented keypoints that enable spatially continuous, distance-weighted aggregation of multi-view features before being splatted into a voxel grid for decoding. This point-centric formulation enables flexible, data-dependent receptive fields and long-range spatial interactions that are difficult to capture with local and dense voxel-based operators. A hierarchical Gaussian representation further enables multi-scale reasoning by combining global context from coarse super-points with fine-grained detail from higher-resolution streams. Extensive experiments on Occ3D nuScenes and Waymo demonstrate state-of-the-art performance for 4D-POT. We provide code and models at https://lags.cs.uni-freiburg.de/.

Figures

Figures reproduced from arXiv: 2602.23172 by the authors.

Figure 1
Figure 1. Illustration of our latent Gaussian representation. Bottom left: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture overview (left to right). An image encoder ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results on the Occ3D-nuScenes validation split. Our [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative results on the Occ3D-Waymo validation split. Our [PITH_FULL_IMAGE:figures/full_fig_p006_4.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. Streaming Gaussian Encoding for 4D Panoptic Occupancy Tracking

    cs.CV 2026-06 unverdicted novelty 7.0 of 10

    Introduces a streaming Gaussian encoder maintaining persistent volumetric representations via ego-motion compensation and confidence-guided updates for improved 4D panoptic occupancy tracking from cameras.

  2. OccTrack360: 4D Panoptic Occupancy Tracking from Surround-View Fisheye Cameras

    cs.CV 2026-03 conditional novelty 6.0 of 10

    A new benchmark and baseline method for 4D panoptic occupancy tracking with surround-view fisheye cameras, built from KITTI-360.

Reference graph

Works this paper leans on

42 extracted references · 1 linked inside Pith · cited by 2 Pith papers

  1. [1]

    TrackOcc: Camera-based 4d panoptic occupancy tracking,

    Z. Chen, K. Li, X. Yang, T. Jiang, Y . Li, and H. Zhao, “TrackOcc: Camera-based 4d panoptic occupancy tracking,” inICRA, 2025

    2025

  2. [2]

    Amodal optical flow,

    M. Luz, R. Mohan, A. R. Sekkat, O. Sawade, E. Matthes, T. Brox, and A. Valada, “Amodal optical flow,” inICRA, 2024

    2024

  3. [3]

    3d multi-object tracking using graph neural networks with cross-edge modality attention,

    M. B ¨uchner and A. Valada, “3d multi-object tracking using graph neural networks with cross-edge modality attention,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 9707–9714, 2022

    2022

  4. [4]

    Self-supervised multi-object tracking for autonomous driving from consistency across timescales,

    C. Lang, A. Braun, L. Schillingmann, and A. Valada, “Self-supervised multi-object tracking for autonomous driving from consistency across timescales,”IEEE Robotics and Automation Letters, vol. 8, no. 11, pp. 7711–7718, 2023

    2023

  5. [5]

    Bridging perspectives: Foundation model guided bev maps for 3d object detection and tracking,

    M. K¨appeler, ¨O. C ¸ic ¸ek, D. Cattaneo, C. Gl¨aser, Y . Miron, and A. Valada, “Bridging perspectives: Foundation model guided bev maps for 3d object detection and tracking,”arXiv preprint arXiv:2510.10287, 2025

    arXiv 2025

  6. [6]

    OccFormer: Dual-path transformer for vision-based 3D semantic occupancy prediction,

    Y . Zhang, Z. Zhu, and D. Du, “OccFormer: Dual-path transformer for vision-based 3D semantic occupancy prediction,” inICCV, 2023

    2023

  7. [7]

    SurroundOcc: Multi-camera 3D occupancy prediction for autonomous driving,

    Y . Wei, L. Zhao, W. Zheng, Z. Zhu, J. Zhou, and J. Lu, “SurroundOcc: Multi-camera 3D occupancy prediction for autonomous driving,” in ICCV, 2023, pp. 21 729–21 740

    2023

  8. [8]

    COTR: Compact occupancy transformer for vision-based 3D occupancy prediction,

    Q. Ma, X. Tan, Y . Qu, L. Ma, Z. Zhang, and Y . Xie, “COTR: Compact occupancy transformer for vision-based 3D occupancy prediction,” in CVPR, 2024, pp. 19 936–19 945

    2024

Show all 42 references
  1. [9]

    GaussianFormer: Scene as gaussians for vision-based 3D semantic occupancy prediction,

    Y . Huang, W. Zheng, Y . Zhang, J. Zhou, and J. Lu, “GaussianFormer: Scene as gaussians for vision-based 3D semantic occupancy prediction,” inECCV, 2025, pp. 376–393

    2025

  2. [10]

    GaussianFormer-2: Probabilistic gaussian superposition for efficient 3D occupancy prediction,

    Y . Huang, A. Thammatadatrakoon, W. Zheng, Y . Zhang, D. Du, and J. Lu, “GaussianFormer-2: Probabilistic gaussian superposition for efficient 3D occupancy prediction,” inCVPR, 2025, pp. 27 477–27 486

    2025

  3. [11]

    GaussianOcc: Fully self-supervised and efficient 3D occupancy estimation with gaussian splatting,

    W. Gan, F. Liu, H. Xu, N. Mo, and N. Yokoya, “GaussianOcc: Fully self-supervised and efficient 3D occupancy estimation with gaussian splatting,” inICCV, 2025, pp. 28 980–28 990

    2025

  4. [12]

    TrackFormer: Multi-object tracking with transformers,

    T. Meinhardt, A. Kirillov, L. Leal-Taix ´e, and C. Feichtenhofer, “TrackFormer: Multi-object tracking with transformers,” inCVPR, 2022, pp. 8844–8854

    2022

  5. [13]

    MOTR: End-to-end multiple-object tracking with transformer,

    F. Zeng, B. Dong, Y . Zhang, T. Wang, X. Zhang, and Y . Wei, “MOTR: End-to-end multiple-object tracking with transformer,” inECCV, 2022

    2022

  6. [14]

    MUTR3D: A multi-camera tracking framework via 3D-to-2D queries,

    T. Zhang, X. Chen, Y . Wang, Y . Wang, and H. Zhao, “MUTR3D: A multi-camera tracking framework via 3D-to-2D queries,” inCVPRW, 2022

    2022

  7. [15]

    Standing between past and future: Spatio-temporal modeling for multi- camera 3D multi-object tracking,

    Z. Pang, J. Li, P. Tokmakov, D. Chen, S. Zagoruyko, and Y .-X. Wang, “Standing between past and future: Spatio-temporal modeling for multi- camera 3D multi-object tracking,” inCVPR, 2023, pp. 17 928–17 938

    2023

  8. [16]

    Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking,

    W. K. Fong, R. Mohan, J. V . Hurtado, L. Zhou, H. Caesar, O. Beijbom, and A. Valada, “Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 3795–3802, 2022

    2022

  9. [17]

    Scalability in perception for autonomous driving: Waymo open dataset,

    P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V . Patnaik, P. Tsui, J. Guo, Y . Zhou, Y . Chai, B. Caine,et al., “Scalability in perception for autonomous driving: Waymo open dataset,” inCVPR, 2020, pp. 2446–2454

    2020

  10. [18]

    Progressive multi- modal fusion for robust 3d object detection,

    R. Mohan, D. Cattaneo, F. Drews, and A. Valada, “Progressive multi- modal fusion for robust 3d object detection,” inCoRL, 2024

    2024

  11. [19]

    Forecas- tOcc: Vision-based semantic occupancy forecasting,

    R. Mohan, J. V . Hurtado, R. Mohan, and A. Valada, “Forecas- tOcc: Vision-based semantic occupancy forecasting,”arXiv preprint arXiv:2602.08006, 2026

    2026

  12. [20]

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

    X. Tian, T. Jiang, L. Yun, Y . Mao, H. Yang, Y . Wang, Y . Wang, and H. Zhao, “Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving,” inNeurIPS, 2023

    2023

  13. [21]

    Per-pixel classification is not all you need for semantic segmentation,

    B. Cheng, A. Schwing, and A. Kirillov, “Per-pixel classification is not all you need for semantic segmentation,” inNeurIPS, vol. 34, 2021, pp. 17 864–17 875

    2021

  14. [22]

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

    P. Tang, Z. Wang, G. Wang, J. Zheng, X. Ren, B. Feng, and C. Ma, “SparseOcc: Rethinking sparse latent representation for vision-based semantic occupancy prediction,” inCVPR, 2024, pp. 15 035–15 044

    2024

  15. [23]

    PaSCo: Urban 3D panoptic scene completion with uncertainty awareness,

    A.-Q. Cao, A. Dai, and R. de Charette, “PaSCo: Urban 3D panoptic scene completion with uncertainty awareness,” inCVPR, 2024, pp. 14 554–14 564

    2024

  16. [24]

    Tri-perspective view for vision-based 3D semantic occupancy prediction,

    Y . Huang, W. Zheng, Y . Zhang, J. Zhou, and J. Lu, “Tri-perspective view for vision-based 3D semantic occupancy prediction,” inCVPR, 2023, pp. 9223–9232

    2023

  17. [25]

    OPUS: Occupancy prediction using a sparse set,

    J. Wang, Z. Liu, Q. Meng, L. Yan, K. Wang, J. Yang, W. Liu, Q. Hou, and M.-M. Cheng, “OPUS: Occupancy prediction using a sparse set,” inNeurIPS, 2024, pp. 119 861–119 885

    2024

  18. [26]

    LION: Linear group RNN for 3D object detection in point clouds,

    Z. Liu, J. Hou, X. Wang, X. Ye, J. Wang, H. Zhao, and X. Bai, “LION: Linear group RNN for 3D object detection in point clouds,” inNeurIPS, vol. 37, 2024, pp. 13 601–13 626

    2024

  19. [27]

    Point transformer v3: Simpler faster stronger,

    X. Wu, L. Jiang, P.-S. Wang, Z. Liu, X. Liu, Y . Qiao, W. Ouyang, T. He, and H. Zhao, “Point transformer v3: Simpler faster stronger,” inCVPR, 2024, pp. 4840–4851

    2024

  20. [28]

    V oxelNeXt: Fully sparse voxelnet for 3D object detection and tracking,

    Y . Chen, J. Liu, X. Zhang, X. Qi, and J. Jia, “V oxelNeXt: Fully sparse voxelnet for 3D object detection and tracking,” inCVPR, 2023, pp. 21 674–21 683

    2023

  21. [29]

    Video Panoptic Segmentation,

    D. Kim, S. Woo, J.-Y . Lee, and I. S. Kweon, “Video Panoptic Segmentation,” inCVPR, 2020, pp. 9859–9868

    2020

  22. [30]

    Masked-attention mask transformer for universal image segmentation,

    B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, “Masked-attention mask transformer for universal image segmentation,” inCVPR, 2022, pp. 1290–1299

    2022

  23. [31]

    CoDEPS: Online continual learning for depth estimation and panoptic segmentation,

    N. V ¨odisch, K. Petek, W. Burgard, and A. Valada, “CoDEPS: Online continual learning for depth estimation and panoptic segmentation,” RSS, 2023

    2023

  24. [32]

    BEVFormer: Learning bird’s-eye-view representation from LiDAR- camera via spatiotemporal transformers,

    Z. Li, W. Wang, H. Li, E. Xie, C. Sima, T. Lu, Q. Yu, and J. Dai, “BEVFormer: Learning bird’s-eye-view representation from LiDAR- camera via spatiotemporal transformers,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 3, pp. 2020–2036, 2025

    2020

  25. [33]

    PETR: Position embedding transformation for multi-view 3D object detection,

    Y . Liu, T. Wang, X. Zhang, and J. Sun, “PETR: Position embedding transformation for multi-view 3D object detection,” inECCV, 2022, pp. 531–548

    2022

  26. [34]

    STEP: Segmenting and tracking every pixel,

    M. Weber, J. Xie, M. D. Collins, Y . Zhu, P. V oigtlaender, H. Adam, B. Green, A. Geiger, B. Leibe, D. Cremers,et al., “STEP: Segmenting and tracking every pixel,” inNeurIPS, 2021

    2021

  27. [35]

    SemanticKITTI: A dataset for semantic scene understand- ing of lidar sequences,

    J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall, “SemanticKITTI: A dataset for semantic scene understand- ing of lidar sequences,” inICCV, 2019, pp. 9297–9307

    2019

  28. [36]

    MinVIS: A minimal video instance segmentation framework without video-based training,

    D.-A. Huang, Z. Yu, and A. Anandkumar, “MinVIS: A minimal video instance segmentation framework without video-based training,” in NeurIPS, vol. 35, 2022, pp. 31 265–31 277

    2022

  29. [37]

    CTVIS: Consistent training for online video instance segmentation,

    K. Ying, Q. Zhong, W. Mao, Z. Wang, H. Chen, L. Y . Wu, Y . Liu, C. Fan, Y . Zhuge, and C. Shen, “CTVIS: Consistent training for online video instance segmentation,” inCVPR, 2023, pp. 899–908

    2023

  30. [38]

    3D multi-object tracking: A baseline and new evaluation metrics,

    X. Weng, J. Wang, D. Held, and K. Kitani, “3D multi-object tracking: A baseline and new evaluation metrics,” inIROS, 2020

    2020

  31. [39]

    4d panoptic lidar segmentation,

    M. Ayg¨un, A. Osep, M. Weber, M. Maximov, C. Stachniss, J. Behley, and L. Leal-Taix ´e, “4d panoptic lidar segmentation,” inCVPR, 2021, pp. 5527–5537

    2021

  32. [40]

    An energy and GPU- computation efficient backbone network for real-time object detection,

    Y . Lee, J.-w. Hwang, S. Lee, Y . Bae, and J. Park, “An energy and GPU- computation efficient backbone network for real-time object detection,” inCVPRW, 2019, pp. 752–760

    2019

  33. [41]

    BEVDet4D: Exploit temporal cues in multi- camera 3D object detection,

    J. Huang and G. Huang, “BEVDet4D: Exploit temporal cues in multi- camera 3D object detection,” arXiv preprint, arXiv:2203.17054, 2022

    2022 arXiv

  34. [42]

    Vision transformer with deformable attention,

    Z. Xia, X. Pan, S. Song, L. E. Li, and G. Huang, “Vision transformer with deformable attention,” inCVPR, 2022, pp. 4794–4803

    2022

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.