pith. sign in

arxiv: 2602.21484 · v2 · submitted 2026-02-25 · 💻 cs.CV

Unified Unsupervised and Sparsely-Supervised 3D Object Detection by Semantic Pseudo-Labeling and Prototype Learning

Yushen He , Lei Zhao , Weidong Chen This is my paper

Pith reviewed 2026-05-15 19:25 UTC · model grok-4.3

classification 💻 cs.CV
keywords 3D object detectionunsupervised learningpseudo-labelingprototype learningsparse supervisionautonomous drivingKITTInuScenes
0
0 comments X p. Extension

The pith

SPL unifies unsupervised and sparsely-supervised 3D object detection by turning fused semantic pseudo-labels into stable prototypes for feature learning.

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

The paper presents SPL as a single training framework that handles both fully unsupervised and sparsely-supervised 3D object detection. It first builds probabilistic pseudo-labels by combining image-based semantics, point-cloud geometry, and temporal consistency across frames to produce both bounding boxes and point-wise labels. These labels are used only as soft priors. A multi-stage prototype learner then maintains memory banks for initialization and applies momentum updates to steadily improve feature representations from labeled and unlabeled data alike. Experiments on KITTI and nuScenes show this yields better detectors than prior methods in both regimes.

Core claim

SPL generates high-quality pseudo-labels by integrating image semantics, point cloud geometry, and temporal cues, producing 3D bounding boxes for dense objects and 3D point labels for sparse ones. These serve as probabilistic priors inside a multi-stage prototype learning strategy that uses memory-based initialization and momentum-based prototype updating to stabilize feature mining from both labeled and unlabeled data.

What carries the argument

Multi-stage prototype learning that treats semantic pseudo-labels as probabilistic priors and updates prototypes through memory initialization and momentum to mine stable features.

If this is right

  • Produces working 3D detectors from sensor data alone without any manual bounding-box labels.
  • Extends naturally to sparse supervision by anchoring the same prototype process on the few available labels.
  • Reduces error propagation by keeping pseudo-labels as soft priors rather than hard targets.
  • Delivers higher detection accuracy than prior unsupervised or semi-supervised methods on KITTI and nuScenes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same fusion-plus-prototype pattern could be tested on 3D semantic segmentation where point-wise labels are even harder to obtain.
  • Performance may degrade in single-frame or static scenes where temporal consistency is unavailable.
  • The memory-bank mechanism suggests a route to continual adaptation when new unlabeled data arrives after initial training.

Load-bearing premise

Fusing image semantics, point cloud geometry, and temporal cues produces sufficiently accurate probabilistic pseudo-label priors that the prototype updates can refine features without spreading errors from the initial noisy labels.

What would settle it

Remove the temporal cue integration or the momentum prototype updates, retrain on KITTI in the unsupervised setting, and check whether average precision falls below existing state-of-the-art unsupervised baselines.

Figures

Figures reproduced from arXiv: 2602.21484 by Lei Zhao, Weidong Chen, Yushen He.

Figure 1
Figure 1. Figure 1: Comparison between representative unsupervised and sparsely [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of contrastive learning strategies under sparse 3D [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed 3D pseudo-label generation pipeline. Starting from synchronized LiDAR frames and RGB images, the pipeline performs [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the prototype-based training strategy in SPL. The framework unifies supervision into GT supervision labels and pseudo labels, mines [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Feature mining process that fuses prototype similarity with pseudo-heatmap priors. Candidate regions are first selected by high prototype similarity and [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of the three-stage training pipeline used to stabilize [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The visualization of ground truth annotations, pseudo labels, and detection results of our SPL method on KITTI dataset. (Part 1) [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The visualization of ground truth annotations, pseudo labels, and detection results of our SPL method on KITTI dataset. (Part 2) [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
read the original abstract

3D object detection is essential for autonomous driving and robotic perception, yet its reliance on large-scale manually annotated data limits scalability and adaptability. To reduce annotation dependency, unsupervised and sparsely-supervised paradigms have emerged. However, they face intertwined challenges: low-quality pseudo-labels, unstable feature mining, and a lack of a unified training framework. This paper proposes SPL, a unified training framework for both unsupervised and sparsely-supervised 3D object detection via \underline{S}emantic \underline{P}seudo-labeling and prototype \underline{L}earning. SPL first generates high-quality pseudo-labels by integrating image semantics, point cloud geometry, and temporal cues, producing both 3D bounding boxes for dense objects and 3D point labels for sparse ones. These pseudo-labels are not used directly but as probabilistic priors within a novel, multi-stage prototype learning strategy. This strategy stabilizes feature representation learning through memory-based initialization and momentum-based prototype updating, effectively mining features from both labeled and unlabeled data. Extensive experiments on KITTI and nuScenes datasets demonstrate that SPL significantly outperforms state-of-the-art methods in both settings. Our work provides a robust and generalizable solution for learning 3D object detectors with minimal or no manual annotations. Our code is available at https://github.com/TossherO/SPL.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims to introduce SPL, a unified training framework for unsupervised and sparsely-supervised 3D object detection. It generates probabilistic pseudo-labels by integrating image semantics, point cloud geometry, and temporal cues for both dense 3D bounding boxes and sparse 3D point labels. These priors are then used in a multi-stage prototype learning strategy involving memory-based initialization and momentum-based prototype updating to mine features from labeled and unlabeled data. Experiments on KITTI and nuScenes datasets show that SPL significantly outperforms state-of-the-art methods in both settings.

Significance. If the results hold, this provides a robust solution for learning 3D detectors with minimal annotations, which is significant for scalable autonomous driving perception. Strengths include the unified framework, the use of probabilistic priors rather than hard labels, the prototype mechanism for stable feature mining, and the public release of code for reproducibility on public benchmarks.

major comments (2)
  1. [§4.1] §4.1 (Prototype Updating): the momentum coefficient is listed among free parameters; the manuscript must clarify whether it is held fixed across all experiments or tuned per dataset, as this directly affects the stability claim for the multi-stage learning.
  2. [Table 2] Table 2 (nuScenes unsupervised row): the reported gains over prior methods are presented without standard deviations or number of runs; this undermines the strength of the outperformance claim given the known variability in pseudo-label quality.
minor comments (2)
  1. [Abstract] Abstract: the notation for SPL is underlined for emphasis; this is unnecessary in the final version and should be removed for standard formatting.
  2. [§5] §5 (Experiments): ensure all ablation studies explicitly state the contribution of each cue (semantic, geometric, temporal) with quantitative deltas rather than qualitative description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment of our work and the recommendation for minor revision. We address the major comments point by point below.

read point-by-point responses

  1. Referee: [§4.1] §4.1 (Prototype Updating): the momentum coefficient is listed among free parameters; the manuscript must clarify whether it is held fixed across all experiments or tuned per dataset, as this directly affects the stability claim for the multi-stage learning.

    Authors: We thank the referee for this clarification request. The momentum coefficient is held fixed at the same value across all experiments on both KITTI and nuScenes to ensure consistent and stable prototype updating in the multi-stage learning process. We will revise §4.1 to explicitly state the fixed value used and confirm that it is not tuned per dataset. revision: yes

  2. Referee: [Table 2] Table 2 (nuScenes unsupervised row): the reported gains over prior methods are presented without standard deviations or number of runs; this undermines the strength of the outperformance claim given the known variability in pseudo-label quality.

    Authors: We agree that reporting run information strengthens the claims. We will update Table 2 with a footnote indicating that results are from single runs (due to the high computational cost of nuScenes training) and add a short discussion in the experimental section on the consistency of gains across datasets despite pseudo-label variability. revision: partial

Circularity Check

0 steps flagged

No significant circularity; method is empirically grounded

full rationale

The SPL framework is presented as an empirical pipeline that generates probabilistic pseudo-label priors from image semantics, point-cloud geometry and temporal cues, then feeds them into a multi-stage prototype-learning procedure that uses memory initialization and momentum updates to mine features from labeled and unlabeled data. No equations or first-principles derivations appear that reduce by construction to fitted parameters or to self-citations; the central claim of outperformance is supported solely by direct experimental comparisons against prior methods on the public KITTI and nuScenes benchmarks. The pseudo-labels are explicitly described as non-hard priors, and the prototype mechanism is designed to operate across both labeled and unlabeled splits, so the argument does not collapse into a self-definitional loop or a fitted-input-called-prediction. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract provides insufficient technical detail to enumerate free parameters or invented entities; the central claim rests on the unverified assumption that multi-cue pseudo-labels act as reliable priors.

free parameters (1)
  • momentum coefficient for prototype updating
    Described as momentum-based updating; value must be chosen or tuned but not specified in abstract.
axioms (1)
  • domain assumption Integration of image semantics, point cloud geometry, and temporal cues yields high-quality probabilistic pseudo-labels suitable as priors
    Invoked as the first step that enables the subsequent prototype learning.

pith-pipeline@v0.9.0 · 5541 in / 1322 out tokens · 62533 ms · 2026-05-15T19:25:38.321343+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages · 1 internal anchor

  1. [1]

    V oxel r- cnn: Towards high performance voxel-based 3d object detection,

    J. Deng, S. Shi, P. Li, W. Zhou, Y . Zhang, and H. Li, “V oxel r- cnn: Towards high performance voxel-based 3d object detection,” in Proceedings of the AAAI conference on artificial intelligence, vol. 35, no. 2, 2021, pp. 1201–1209

    work page 2021

  2. [2]

    Pv-rcnn: Point-voxel feature set abstraction for 3d object detection,

    S. Shi, C. Guo, L. Jiang, Z. Wang, J. Shi, X. Wang, and H. Li, “Pv-rcnn: Point-voxel feature set abstraction for 3d object detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 10 529–10 538

    work page 2020

  3. [3]

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

    Z. Liu, H. Tang, A. Amini, X. Yang, H. Mao, D. Rus, and S. Han, “Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,”arXiv preprint arXiv:2205.13542, 2022

    work page arXiv 2022

  4. [4]

    Cross modal transformer: Towards fast and robust 3d object detection,

    J. Yan, Y . Liu, J. Sun, F. Jia, S. Li, T. Wang, and X. Zhang, “Cross modal transformer: Towards fast and robust 3d object detection,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 18 268–18 278

    work page 2023

  5. [5]

    Are we ready for autonomous driving? the kitti vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in2012 IEEE conference on computer vision and pattern recognition. IEEE, 2012, pp. 3354–3361

    work page 2012

  6. [6]

    nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 621–11 631. 10

    work page 2020

  7. [7]

    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. Caineet al., “Scalability in perception for autonomous driving: Waymo open dataset,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2446–2454

    work page 2020

  8. [8]

    Towards unsupervised object detection from lidar point clouds,

    L. Zhang, A. J. Yang, Y . Xiong, S. Casas, B. Yang, M. Ren, and R. Urtasun, “Towards unsupervised object detection from lidar point clouds,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9317–9328

    work page 2023

  9. [9]

    Commonsense prototype for outdoor unsupervised 3d object detection,

    H. Wu, S. Zhao, X. Huang, C. Wen, X. Li, and C. Wang, “Commonsense prototype for outdoor unsupervised 3d object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 14 968–14 977

    work page 2024

  10. [10]

    Learning to detect mobile objects from lidar scans without labels,

    Y . You, K. Luo, C. P. Phoo, W.-L. Chao, W. Sun, B. Hariharan, M. Campbell, and K. Q. Weinberger, “Learning to detect mobile objects from lidar scans without labels,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1130–1140

    work page 2022

  11. [11]

    Liso: Lidar-only self- supervised 3d object detection,

    S. A. Baur, F. Moosmann, and A. Geiger, “Liso: Lidar-only self- supervised 3d object detection,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 253–270

    work page 2024

  12. [12]

    Motal: Unsupervised 3d object detection by modality and task-specific knowledge transfer,

    H. Wu, H. Lin, X. Guo, X. Li, M. Wang, C. Wang, and C. Wen, “Motal: Unsupervised 3d object detection by modality and task-specific knowledge transfer,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 6284–6293

    work page 2025

  13. [13]

    Fgr: Frustum-aware geometric reasoning for weakly supervised 3d vehicle detection,

    Y . Wei, S. Su, J. Lu, and J. Zhou, “Fgr: Frustum-aware geometric reasoning for weakly supervised 3d vehicle detection,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 4348–4354

    work page 2021

  14. [14]

    Annofreeod: Detecting all classes at low frame rates without human annotations,

    B. Sun, Y . Liu, H. He, Y . Tian, and F.-Y . Wang, “Annofreeod: Detecting all classes at low frame rates without human annotations,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 5315–5325

    work page 2025

  15. [15]

    Enhancing pseudo-boxes via data- level lidar-camera fusion for unsupervised 3d object detection,

    M. Ji, J. Yang, and S. Zhang, “Enhancing pseudo-boxes via data- level lidar-camera fusion for unsupervised 3d object detection,” in Proceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 896–904

    work page 2025

  16. [16]

    Approaching outside: scaling unsupervised 3d object detection from 2d scene,

    R. Zhang, H. Zhang, H. Yu, and Z. Zheng, “Approaching outside: scaling unsupervised 3d object detection from 2d scene,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 249–266

    work page 2024

  17. [17]

    Union: Unsupervised 3d ob- ject detection using object appearance-based pseudo-classes,

    T. Lentsch, H. Caesar, and D. M. Gavrila, “Union: Unsupervised 3d ob- ject detection using object appearance-based pseudo-classes,”Advances in Neural Information Processing Systems, vol. 37, pp. 22 028–22 046, 2024

    work page 2024

  18. [18]

    Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations,

    Q. Xia, J. Deng, C. Wen, H. Wu, S. Shi, X. Li, and C. Wang, “Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 6254–6263

    work page 2023

  19. [19]

    Learning class prototypes for unified sparse-supervised 3d object detection,

    Y . Zhu, L. Hui, H. Yang, J. Qian, J. Xie, and J. Yang, “Learning class prototypes for unified sparse-supervised 3d object detection,” inPro- ceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 9911–9920

    work page 2025

  20. [20]

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

    Y . Zhou and O. 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, 2018, pp. 4490–4499

    work page 2018

  21. [21]

    Second: Sparsely embedded convolutional detection,

    Y . Yan, Y . Mao, and B. Li, “Second: Sparsely embedded convolutional detection,”Sensors, vol. 18, no. 10, p. 3337, 2018

    work page 2018

  22. [22]

    Center-based 3d object detection and tracking,

    T. Yin, X. Zhou, and P. Krahenbuhl, “Center-based 3d object detection and tracking,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 11 784–11 793

    work page 2021

  23. [23]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 12 697–12 705

    work page 2019

  24. [24]

    Smoke: Single-stage monocular 3d object detection via keypoint estimation,

    Z. Liu, Z. Wu, and R. T ´oth, “Smoke: Single-stage monocular 3d object detection via keypoint estimation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 996–997

    work page 2020

  25. [25]

    Categorical depth distribution network for monocular 3d object detection,

    C. Reading, A. Harakeh, J. Chae, and S. L. Waslander, “Categorical depth distribution network for monocular 3d object detection,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 8555–8564

    work page 2021

  26. [26]

    Petr: Position embedding trans- formation for multi-view 3d object detection,

    Y . Liu, T. Wang, X. Zhang, and J. Sun, “Petr: Position embedding trans- formation for multi-view 3d object detection,” inEuropean conference on computer vision. Springer, 2022, pp. 531–548

    work page 2022

  27. [27]

    Mvx-net: Multimodal voxelnet for 3d object detection,

    V . A. Sindagi, Y . Zhou, and O. Tuzel, “Mvx-net: Multimodal voxelnet for 3d object detection,” in2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019, pp. 7276–7282

    work page 2019

  28. [28]

    Synet: A synergistic network for 3d object detection through geometric-semantic-based multi- interaction fusion,

    X. Zhang, K. Bi, S. Chan, S. Lu, and X. Zhou, “Synet: A synergistic network for 3d object detection through geometric-semantic-based multi- interaction fusion,”IEEE Transactions on Multimedia, 2025

    work page 2025

  29. [29]

    Detection, classification and tracking of moving objects in a 3d environment,

    A. Azim and O. Aycard, “Detection, classification and tracking of moving objects in a 3d environment,” in2012 IEEE Intelligent Vehicles Symposium. IEEE, 2012, pp. 802–807

    work page 2012

  30. [30]

    Robust moving objects detection in lidar data exploiting visual cues,

    G. Postica, A. Romanoni, and M. Matteucci, “Robust moving objects detection in lidar data exploiting visual cues,” in2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2016, pp. 1093–1098

    work page 2016

  31. [31]

    Dynamic multi-lidar based multiple object detection and tracking,

    M. Sualeh and G.-W. Kim, “Dynamic multi-lidar based multiple object detection and tracking,”Sensors, vol. 19, no. 6, p. 1474, 2019

    work page 2019

  32. [32]

    Ss3d: Sparsely- supervised 3d object detection from point cloud,

    C. Liu, C. Gao, F. Liu, J. Liu, D. Meng, and X. Gao, “Ss3d: Sparsely- supervised 3d object detection from point cloud,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8428–8437

    work page 2022

  33. [33]

    Hinted: Hard instance enhanced detector with mixed-density feature fusion for sparsely-supervised 3d object detection,

    Q. Xia, W. Ye, H. Wu, S. Zhao, L. Xing, X. Huang, J. Deng, X. Li, C. Wen, and C. Wang, “Hinted: Hard instance enhanced detector with mixed-density feature fusion for sparsely-supervised 3d object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 321–15 330

    work page 2024

  34. [34]

    Sp3d: Boosting sparsely-supervised 3d object detection via accurate cross-modal semantic prompts,

    S. Zhao, Q. Xia, X. Guo, P. Zou, M. Zheng, H. Wu, C. Wen, and C. Wang, “Sp3d: Boosting sparsely-supervised 3d object detection via accurate cross-modal semantic prompts,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 29 374– 29 384

    work page 2025

  35. [35]

    Hardness-aware scene syn- thesis for semi-supervised 3d object detection,

    S. Zeng, W. Zheng, J. Lu, and H. Yan, “Hardness-aware scene syn- thesis for semi-supervised 3d object detection,”IEEE Transactions on Multimedia, vol. 26, pp. 9644–9656, 2024

    work page 2024

  36. [36]

    Weakly supervised object detection with class prototypical network,

    H. Li, Y . Li, Y . Cao, Y . Han, Y . Jin, and Y . Wei, “Weakly supervised object detection with class prototypical network,”IEEE Transactions on Multimedia, vol. 25, pp. 1868–1878, 2022

    work page 2022

  37. [37]

    Break- ing immutable: Information-coupled prototype elaboration for few-shot object detection,

    X. Lu, W. Diao, Y . Mao, J. Li, P. Wang, X. Sun, and K. Fu, “Break- ing immutable: Information-coupled prototype elaboration for few-shot object detection,” inProceedings of the AAAI conference on artificial intelligence, vol. 37, no. 2, 2023, pp. 1844–1852

    work page 2023

  38. [38]

    Query- guided prototype evolution network for few-shot segmentation,

    R. Cong, H. Xiong, J. Chen, W. Zhang, Q. Huang, and Y . Zhao, “Query- guided prototype evolution network for few-shot segmentation,”IEEE Transactions on Multimedia, vol. 26, pp. 6501–6512, 2024

    work page 2024

  39. [39]

    Prototypical votenet for few-shot 3d point cloud object detection,

    S. Zhao and X. Qi, “Prototypical votenet for few-shot 3d point cloud object detection,”Advances in neural information processing systems, vol. 35, pp. 13 838–13 851, 2022

    work page 2022

  40. [40]

    Gpa-3d: Geometry- aware prototype alignment for unsupervised domain adaptive 3d object detection from point clouds,

    Z. Li, J. Guo, T. Cao, L. Bingbing, and W. Yang, “Gpa-3d: Geometry- aware prototype alignment for unsupervised domain adaptive 3d object detection from point clouds,” inProceedings of the IEEE/CVF interna- tional conference on computer vision, 2023, pp. 6394–6403

    work page 2023

  41. [41]

    Cl3d: Unsupervised domain adaptation for cross-lidar 3d detection,

    X. Peng, X. Zhu, and Y . Ma, “Cl3d: Unsupervised domain adaptation for cross-lidar 3d detection,” inProceedings of the AAAI conference on artificial intelligence, vol. 37, no. 2, 2023, pp. 2047–2055

    work page 2023

  42. [42]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9729–9738

    work page 2020

  43. [43]

    Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud,

    S. Lee, H. Lim, and H. Myung, “Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 13 276–13 283

    work page 2022

  44. [44]

    YOLOv12: Attention-Centric Real-Time Object Detectors

    Y . Tian, Q. Ye, and D. Doermann, “Yolov12: Attention-centric real-time object detectors,”arXiv preprint arXiv:2502.12524, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  45. [45]

    Available from: https://arxiv.org/ abs/2206.14651

    N. Aharon, R. Orfaig, and B.-Z. Bobrovsky, “Bot-sort: Robust associa- tions multi-pedestrian tracking,”arXiv preprint arXiv:2206.14651, 2022

    work page arXiv 2022

  46. [46]

    A density-based algorithm for discovering clusters in large spatial databases with noise,

    M. Ester, H.-P. Kriegel, J. Sander, X. Xuet al., “A density-based algorithm for discovering clusters in large spatial databases with noise,” inkdd, vol. 96, no. 34, 1996, pp. 226–231

    work page 1996

  47. [47]

    K-nearest neighbor,

    L. E. Peterson, “K-nearest neighbor,”Scholarpedia, vol. 4, no. 2, p. 1883, 2009

    work page 2009

  48. [48]

    Efficient l-shape fitting for vehicle detection using laser scanners,

    X. Zhang, W. Xu, C. Dong, and J. M. Dolan, “Efficient l-shape fitting for vehicle detection using laser scanners,” in2017 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2017, pp. 54–59

    work page 2017

  49. [49]

    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,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 21 674–21 683. 11 APPENDIX A. Implementation Details In this section, we provide comprehensive implementation details of SPL, including pseud...

    work page 2023

  50. [50]

    These settings are selected according to the characteristics of KITTI and nuScenes to balance pseudo-label precision and recall

    Pseudo-Label Generation:The detailed parameters of pseudo-label generation are summarized in Table VIII. These settings are selected according to the characteristics of KITTI and nuScenes to balance pseudo-label precision and recall. Concretely, the mask-dilation and depth-range constraints are used to improve point association quality in crowded scenes, ...

    work page

  51. [51]

    Network and Training:Following the baseline detector settings, we report only the additional parameters introduced by SPL and its multi-stage prototype learning strategy. The re- ported settings are directly tied to the core method components in sections III-B and III-C: prototype construction and update, pseudo-label-guided feature mining, contrastive ob...

    work page