Pith. sign in

REVIEW 3 major objections 4 minor 47 references

Learning better representations for crowded pedestrians in offboard LiDAR-camera 3D tracking-by-detection

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper reports that an offboard LiDAR-camera 3D tracker, augmented with density-aware and relationship-aware representation learning, more than doubles pedestrian MOTA on a new crowd benchmark (0.172 to 0.353) and also reaches 72.2…

desk verdict A credible but unproven auto-labeling pipeline for crowded pedestrians; the headline MOTA gain rests on an unvalidated new benchmark that needs release and annotation-quality checks. read the letter →

arxiv 2505.16029 v1 pith:H6VGXW7O submitted 2025-05-21 cs.CV

Shichao Li , Peiliang Li , Qing Lian , Peng Yun , Xiaozhi Chen This is my paper · ORCID
classification cs.CV
keywords crowdedpedestriantrackingoffboardauto-labelingLiDAR-camerafusion3Dmultiple-objectdensity-awarelossrelationship-awareoffsettargethigh-resolutionsparsefeaturesPCP-MVbenchmark
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

Perceiving pedestrians in dense urban crowds is a long-tail problem for autonomous driving, and manually labeling 3D trajectories in such scenes is expensive. This paper argues that an offboard LiDAR-camera tracking-by-detection system, which runs offline with abundant compute, can be made accurate enough to pre-label crowded scenes by changing how the detector represents pedestrians. On the new PCP-MV benchmark, the proposed combination of a density-aware focal loss, relationship-aware offset targets, and high-resolution sparse feature fusion raises pedestrian MOTA from 0.172 to 0.353, a relative gain of 105%, and the same system reaches 72.2 AMOTA on nuScenes. If the results hold, they would directly cut the cost of generating training data for crowded-scene perception.

What carries the argument

The central machinery is a set of three representation-learning additions applied to a BEVFusion detector with center-based tracking. Density-aware spatial weights upweight BEV locations by the number of ground-truth pedestrians within a threshold distance, multiplying the per-location focal loss so that crowded regions dominate the gradient. The relationship-aware target appends, for each pedestrian with a neighbor within 3 meters, a 2D offset vector to its nearest neighbor, providing distinct regression targets that force the model to separate adjacent instances even when point clouds overlap. High-resolution sparse feature learning fuses low-level and high-level sparse-convolution features and doubles the output BEV resolution from 0.6 m to 0.3 m, with a multi-scale version that adds cross-resolution fusion blocks. A supporting change replaces the forward camera-to-BEV view transform, which is unreliable for small objects, with a backward transform that samples camera features at dense voxels re-projected into the images.

What would settle it

Re-annotate a random subset of PCP-MV test frames with an independent annotation team using the same 'at least one point' rule, measure per-pedestrian box IoU and identity agreement against the released ground truth, then retrain and re-evaluate the baseline and the full system on the subset where both teams agree; if the 105% MOTA gain shrinks substantially or inter-annotator agreement is poor, the central claim is an annotation artifact.

Watch

Extended reading notes

Core claim

The central claim is that the failure of offboard LiDAR-camera 3D MOT in highly crowded pedestrian scenes is not primarily a tracking or association problem but a representation problem, and that three targeted changes to the detector's representation—density-aware focal weighting, relationship-aware offset targets, and high-resolution sparse feature fusion—remedy it. On PCP-MV, the collected benchmark with 3.8 pedestrians on average within a 2-meter radius (versus 0.7 for nuScenes), the modified BEVFusion baseline improves from 0.172 MOTA to 0.353 MOTA with the full multi-scale system, a 105.2% relative gain; identity switches drop from 2945 to 2774 and the most-lost ratio from 0.307 to 0.268. The same system reaches 72.2 overall AMOTA on the nuScenes validation set, surpassing a strong LiDAR-only center-based tracker and a LiDAR-camera probabilistic tracker. This is offered as evidence that the crowd-specific training objectives and denser BEV features generalize beyond the new benchmark, and that offboard auto-labeling of crowded urban scenes can operate at a level where manual annotators mainly verify rather than create annotations.

Load-bearing premise

The load-bearing premise is that PCP-MV's manual annotations are accurate and unbiased, since no inter-annotator agreement or independent cross-check is reported; if heavily occluded pedestrians are systematically underlabeled in dense crowds, both the training supervision and the MOTA evaluation would be biased and the claimed doubling could partly reflect annotation artifacts.

Editorial extensions

If this is right

  • The proposed offboard system can pre-label highly crowded pedestrian scenes with MOTA above 0.35, meaning annotators only need to correct a minority of boxes and identities rather than draw them from scratch.
  • Because density weighting and relationship targets operate only during training, they add zero inference cost and can be dropped into any existing center-based detector.
  • The system transfers to less crowded scenes: on nuScenes it improves overall AMOTA to 72.2, beating strong LiDAR-only and LiDAR-camera baselines, suggesting the crowd-specific training signals do not hurt general performance.
  • High-resolution sparse feature fusion is the largest single contributor (MOTA jump from 0.172 to 0.325), indicating that BEV resolution is a primary bottleneck for small-object detection in crowds.
  • The PCP-MV benchmark, with pedestrian density 3.8 within 2 meters versus 0.7 for nuScenes, provides a public testbed for future crowd-specific 3D MOT research.

Reading between the lines

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

  • A testable extension: because the density-aware weighting upweights loss by local count, the same per-location reweighting could be applied to other dense-scene perception tasks such as vehicle tracking in congestion or robot navigation in crowds, where the same bias against closely spaced objects exists.
  • The relationship-offset target is a cheap auxiliary supervision that forces the network to attend to inter-instance differences; the same idea could be transferred to 3D instance segmentation, where adjacent objects with few LiDAR points are often merged into one mask.
  • Given that DAW and RAT incur no inference cost, the paper's design predicts that they will keep providing additive gains when combined with any future stronger detector or backbone; a quick check is to append them to a state-of-the-art online tracker and measure the MOTA delta on PCP-MV.
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

3 major / 4 minor

Summary. The paper studies offboard LiDAR-camera 3D multi-object tracking (MOT) of crowded pedestrians. It introduces PCP-MV, a new multi-view LiDAR-camera benchmark with dense pedestrian scenes, and a tracking-by-detection system built on BEVFusion. Three contributions are proposed: density-aware spatial weighting (DAW) in the detection loss, relationship-aware offset targets (RAT) that regress a vector to the nearest neighbor for close pedestrians, and high-resolution sparse feature learning (HRL) with two variants. On PCP-MV, the reported MOTA improves from 0.172 (baseline) to 0.353 (Full-MS), a 105.2% relative gain; on the nuScenes validation set the system achieves 72.2 AMOTA, comparing favorably to several prior methods.

Significance. If the PCP-MV annotations are reliable and the results reproducible, this is a useful contribution to offboard auto-labeling: the paper identifies a practical long-tail scenario, proposes a dedicated benchmark, and provides a clean ablation isolating each component. The nuScenes result indicates the design transfers beyond the crowded setting. The planned code release is a strength, though the benchmark itself is not yet released.

major comments (3)
  1. [Section III and Table II] The central claim that the proposed system doubles pedestrian MOTA on PCP-MV rests on ground-truth annotations that are never independently validated. The annotation protocol described in Section III -- 'annotators first exhaustively label objects that have at least one point' and then verify correctness and identity coherence -- is precisely where systematic bias is likely in dense crowds: heavily occluded pedestrians may be underlabeled, and adjacent point clouds may be merged. No inter-annotator agreement, re-labeling cross-check, error analysis, or benchmark release is provided. Because the same labels supervise DAW, RAT, and HRL (Section IV) and define the MOTA denominator and FP/FN/IDS counts in Eq. (6), annotation bias would be propagated into both training and evaluation and could account for a substantial part of the reported 0.172-to-0.353 gain. The paper should add inter-annotator agreement on a subset, an analysis of label quality versus occlusion/distance, or at least a public release of the annotations to allow independent checking.
  2. [Table II] The final incremental claim that Full-MS is the best system is not supported by the accompanying metrics: compared to Full, Full-MS improves MOTA (0.353 vs 0.333) but worsens MTR (0.315 vs 0.371), MLR (0.268 vs 0.229), and IDS (2774 vs 2460). Since all results are single runs on a single hand-collected test set with no error bars or significance tests, the paper should either report variance across multiple training runs or provide a clear argument for why the MOTA gain dominates the observed degradation in these other metrics. Without this, the incremental 6% relative gain over Full is not robust.
  3. [Sections IV-C and IV-D] The two key hyperparameters in the proposed methods -- the density radius th in Eq. (3) and the 3 m neighbor distance used to define relationship offsets in Section IV-D -- are hand-chosen and no sensitivity analysis is provided. Given that DAW and RAT contribute 14.5% and 7.5% relative MOTA gains respectively in Table II, the paper should demonstrate that these gains are not artifacts of threshold tuning by varying th and the neighbor radius over a reasonable range (e.g., 1-5 m) and reporting MOTA.
minor comments (4)
  1. [Eq. (3)] The indicator expression has a misplaced parenthesis; it should read 1(dist(j,k,x_i,y_i) < th).
  2. [Section IV-C] In the DAW loss definition, the condition c_{j,k}=1 is ambiguous because c_{j,k} in Eq. (2) is a Gaussian heatmap rather than a binary mask; please clarify that the positive-loss branch applies at object-center cells.
  3. [Section III] The paper states the code will be publicly available but does not state whether the PCP-MV benchmark annotations will be released; please add an availability statement for the benchmark, as the reproducibility of the main claim depends on it.
  4. [Table III] VoxelNeXt per-class AMOTA values are omitted (shown as '-'), although the text compares overall performance; reporting the pedestrian category for VoxelNeXt would make the comparison more informative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed gains are empirical ablations on a held-out test split plus an external nuScenes benchmark, with no load-bearing self-citation or fitted quantity renamed as prediction.

full rationale

The paper's central claims are empirical rather than derived. Table II reports MOTA on PCP-MV for ablations, and Table III reports AMOTA on the external nuScenes validation set. The proposed components are training-time modifications: DAW reweights the focal loss using a count of nearby GT objects (Eq. 3), RAT adds an auxiliary regression target to the nearest neighbor within 3 m (Eq. 5), and HRL changes the sparse feature encoder resolution (Fig. 6). None of these quantities enters the MOTA computation (Eq. 6), and none is equivalent to the final metric by construction. The PCP-MV training and test splits come from different sequences, so the reported MOTA gains are not fits to the evaluation labels. The nuScenes comparison is an independent external benchmark and grounds the generality claim. Remaining concerns—no inter-annotator agreement statistics for the new benchmark, the undisclosed DAW threshold, and absent sensitivity or variance analysis—are validation and reporting limitations, not circular derivations. No load-bearing step reduces to its own input or to a self-citation chain.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The paper's contributions are empirical: a new dataset and three training or architecture modifications. The ledger therefore contains no fitted model parameters in the physics sense; it lists the hand-set thresholds that the core losses depend on, plus the dataset and metric assumptions that the central comparison rests on. The main unstated burden is that PCP-MV is simultaneously the training signal, the evaluation set, and the evidence for a labeling-efficiency claim.

free parameters (2)
  • DAW density radius th
    Hand-set radius in Eq. (3) that decides which BEV locations receive density-aware focal weighting; its value is not reported in the main text.
  • Relationship neighbor radius = 3 m
    Section IV-D restricts the relationship-offset target to pedestrians with a neighbor within 3 meters; hand-chosen and not swept in an ablation.
assumptions (5)
  • domain assumption PCP-MV ground-truth annotations are correct, complete, and temporally coherent.
    Section III says annotators label objects with at least one point and verify identity coherence, but no inter-annotator agreement, error rate, or independent cross-check is reported.
  • domain assumption A single center-based Gaussian heatmap per object is a sufficient detection supervision target, and the extra relationship offsets fix the residual ambiguity.
    Eq. (2) inherits the BEVFusion heatmap; Section IV-D argues this target is identical for different humans and thus motivates the added nearest-neighbor target, but the sufficiency of the combined target is not independently tested.
  • domain assumption MOTA computed with BEV IoU threshold 0.5 is an adequate proxy for auto-labeling efficiency.
    Section V uses MOTA and AMOTA as the only success metrics and the conclusion extends to labeling efficiency, but no labeling-cost, time-saved, or correction-rate measurement is reported.
  • domain assumption Input ego poses, calibrations, and rectified pinhole images are accurate enough for cross-sensor feature sampling.
    Section IV-A takes calibration and ego poses as given inputs; the backward view transform in Section IV-B re-projects voxels through these transforms, so calibration error would directly corrupt fused features.
  • domain assumption Sparse convolution features in BEV preserve enough information for adjacent-instance discrimination when higher resolution is used.
    Section IV-E asserts HR sparse features help cluttered small objects; this is empirically validated on PCP-MV but not on an independent crowded set.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning better representations for crowded pedestrians in offboard LiDAR-camera 3D tracking-by-detection." pith.science (2026). https://pith.science/paper/H6VGXW7O

@misc{pith2026250516029,
  author       = {Pith},
  title        = {Pith review of: Learning better representations for crowded pedestrians in offboard LiDAR-camera 3D tracking-by-detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H6VGXW7O}},
  note         = {Machine review of arXiv:2505.16029}
}
read the original abstract

Perceiving pedestrians in highly crowded urban environments is a difficult long-tail problem for learning-based autonomous perception. Speeding up 3D ground truth generation for such challenging scenes is performance-critical yet very challenging. The difficulties include the sparsity of the captured pedestrian point cloud and a lack of suitable benchmarks for a specific system design study. To tackle the challenges, we first collect a new multi-view LiDAR-camera 3D multiple-object-tracking benchmark of highly crowded pedestrians for in-depth analysis. We then build an offboard auto-labeling system that reconstructs pedestrian trajectories from LiDAR point cloud and multi-view images. To improve the generalization power for crowded scenes and the performance for small objects, we propose to learn high-resolution representations that are density-aware and relationship-aware. Extensive experiments validate that our approach significantly improves the 3D pedestrian tracking performance towards higher auto-labeling efficiency. The code will be publicly available at this HTTP URL.

Figures

Figures reproduced from arXiv: 2505.16029 by the authors.

Figure 1
Figure 1. A comparison of the pedestrian density with vary￾ing circle radii. MV datasets Density-2 nuScenes [11] 0.7 H3D [12] 1.5 Waymo Open [10] 1.0 PCP-MV (Ours) 3.8 TABLE I: Comparison of the pedestrian density using a cir￾cle radius of 2 meters. The statistics of other datasets are taken from [13]. labeling costs due to a large number of objects and requires a systematic study for efficiency improvement. However, several … view at source ↗
Figure 2
Figure 2. An example frame of PCP-MV. Left: Bird’s Eye View of the captured point cloud and box annotations. Right: six [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Diagram of our offboard system for 3D MOT of crowded pedestrians. Our proposed representation learning approaches [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Ground truth relationship offset targets shown as [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 4
Figure 4. Figure 4: (a) A ground truth heatmap and (b) the density-aware [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: We enhance system performance with a high [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results of 3D multiple-pedestrian tracking in crowded environments. Detected objects in the video sequence [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Predicted relationship vectors on PCP-MV. An arrow [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 32 canonical work pages

  1. [1]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in European conference on computer vision . Springer, 2020, pp. 213– 229

    2020

  2. [2]

    An end-to-end transformer model for 3d object detection,

    I. Misra, R. Girdhar, and A. Joulin, “An end-to-end transformer model for 3d object detection,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 2906–2917

    work page 2021

  3. [3]

    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,” in European Conference on Computer Vision . Springer, 2022, pp. 659–675

    work page 2022

  4. [4]

    Planning-oriented autonomous driving,

    Y . Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wang et al. , “Planning-oriented autonomous driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 17 853–17 862

    2023

  5. [5]

    Monodetr: Depth-guided transformer for monocular 3d object detection,

    R. Zhang, H. Qiu, T. Wang, Z. Guo, Z. Cui, Y . Qiao, H. Li, and P. Gao, “Monodetr: Depth-guided transformer for monocular 3d object detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 9155–9166

    2023

  6. [6]

    3dmot- former: Graph transformer for online 3d multi-object tracking,

    S. Ding, E. Rehder, L. Schneider, M. Cordts, and J. Gall, “3dmot- former: Graph transformer for online 3d multi-object tracking,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 9784–9794

    work page 2023

  7. [7]

    Auto4D: Learning to Label 4D Objects from Sequential Point Clouds

    B. Yang, M. Bai, M. Liang, W. Zeng, and R. Urtasun, “Auto4d: Learning to label 4d objects from sequential point clouds,” arXiv preprint arXiv:2101.06586, 2021

    work page Pith review arXiv 2021

  8. [8]

    Once detected, never lost: Surpassing human performance in offline lidar based 3d object detection,

    L. Fan, Y . Yang, Y . Mao, F. Wang, Y . Chen, N. Wang, and Z. Zhang, “Once detected, never lost: Surpassing human performance in offline lidar based 3d object detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 19 820– 19 829

    work page 2023

Show all 47 references
  1. [9]

    Detzero: Rethinking offboard 3d object detection with long-term sequential point clouds,

    T. Ma, X. Yang, H. Zhou, X. Li, B. Shi, J. Liu, Y . Yang, Z. Liu, L. He, Y . Qiaoet al., “Detzero: Rethinking offboard 3d object detection with long-term sequential point clouds,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 6736–6747

    2023

  2. [10]

    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,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , ...

    2020

  3. [11]

    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,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 11 621–11 631

    2020

  4. [12]

    The h3d dataset for full-surround 3d multi-object detection and tracking in crowded urban scenes,

    A. Patil, S. Malla, H. Gang, and Y .-T. Chen, “The h3d dataset for full-surround 3d multi-object detection and tracking in crowded urban scenes,” in 2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019, pp. 9552–9557

    2019

  5. [13]

    Stcrowd: A multimodal dataset for pedes- trian perception in crowded scenes,

    P. Cong, X. Zhu, F. Qiao, Y . Ren, X. Peng, Y . Hou, L. Xu, R. Yang, D. Manocha, and Y . Ma, “Stcrowd: A multimodal dataset for pedes- trian perception in crowded scenes,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 19 608–19 617

    2022

  6. [14]

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

    Z. Liu, H. Tang, A. Amini, X. Yang, H. Mao, D. L. Rus, and S. Han, “Bevfusion: Multi-task multi-sensor fusion with unified bird’s- eye view representation,” in 2023 IEEE international conference on robotics and automation (ICRA) . IEEE, 2023, pp. 2774–2781

    2023

  7. [15]

    Center-based 3d object detec- tion and tracking,

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

    2021

  8. [16]

    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,” in 2020 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS) . IEEE, 2020, pp. 10 359–10 366

    2020

  9. [17]

    Gnn3dmot: Graph neural network for 3d multi-object tracking with 2d-3d multi-feature learning,

    X. Weng, Y . Wang, Y . Man, and K. M. Kitani, “Gnn3dmot: Graph neural network for 3d multi-object tracking with 2d-3d multi-feature learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 6499–6508

    2020

  10. [18]

    Exploring simple 3d multi-object tracking for autonomous driving,

    C. Luo, X. Yang, and A. Yuille, “Exploring simple 3d multi-object tracking for autonomous driving,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 488–10 497

    2021

  11. [19]

    Spot: Spatiotemporal modeling for 3d object tracking,

    C. Stearns, D. Rempe, J. Li, R. Ambrus ¸, S. Zakharov, V . Guizilini, Y . Yang, and L. J. Guibas, “Spot: Spatiotemporal modeling for 3d object tracking,” in European Conference on Computer Vision . Springer, 2022, pp. 639–656

    2022

  12. [20]

    Tracking objects as points,

    X. Zhou, V . Koltun, and P. Kr ¨ahenb¨uhl, “Tracking objects as points,” in European conference on computer vision . Springer, 2020, pp. 474–490

    2020

  13. [21]

    Probabilistic 3d multi- modal, multi-object tracking for autonomous driving,

    H.-k. Chiu, J. Li, R. Ambrus ¸, and J. Bohg, “Probabilistic 3d multi- modal, multi-object tracking for autonomous driving,” in 2021 IEEE international conference on robotics and automation (ICRA) . IEEE, 2021, pp. 14 227–14 233

    2021

  14. [22]

    Pedestrian detection in crowded scenes,

    B. Leibe, E. Seemann, and B. Schiele, “Pedestrian detection in crowded scenes,” in 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR’05) , vol. 1. IEEE, 2005, pp. 878–885

    2005

  15. [23]

    How far are we from solving pedestrian detection?

    S. Zhang, R. Benenson, M. Omran, J. Hosang, and B. Schiele, “How far are we from solving pedestrian detection?” in Proceedings of the iEEE conference on computer vision and pattern recognition , 2016, pp. 1259–1267

    2016

  16. [24]

    Multi-view 3d human tracking in crowded scenes,

    X. Liu, “Multi-view 3d human tracking in crowded scenes,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 30, no. 1, 2016

    2016

  17. [25]

    On multi-modal people tracking from mobile platforms in very crowded and dynamic environments,

    T. Linder, S. Breuers, B. Leibe, and K. O. Arras, “On multi-modal people tracking from mobile platforms in very crowded and dynamic environments,” in 2016 IEEE international conference on robotics and automation (ICRA). IEEE, 2016, pp. 5512–5519

    2016

  18. [26]

    Where, what, whether: Multi-modal learning meets pedestrian detection,

    Y . Luo, C. Zhang, M. Zhao, H. Zhou, and J. Sun, “Where, what, whether: Multi-modal learning meets pedestrian detection,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 14 065–14 073

    2020

  19. [27]

    Detection in crowded scenes: One proposal, multiple predictions,

    X. Chu, A. Zheng, X. Zhang, and J. Sun, “Detection in crowded scenes: One proposal, multiple predictions,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 12 214–12 223

    2020

  20. [28]

    Temporal- context enhanced detection of heavily occluded pedestrians,

    J. Wu, C. Zhou, M. Yang, Q. Zhang, Y . Li, and J. Yuan, “Temporal- context enhanced detection of heavily occluded pedestrians,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 13 430–13 439

    2020

  21. [29]

    Graininess-aware deep feature learning for robust pedestrian detection,

    C. Lin, J. Lu, G. Wang, and J. Zhou, “Graininess-aware deep feature learning for robust pedestrian detection,” IEEE Transactions on Image Processing, vol. 29, pp. 3820–3834, 2020

    2020

  22. [30]

    Occlusion handling and multi-scale pedestrian detection based on deep learning: A review,

    F. Li, X. Li, Q. Liu, and Z. Li, “Occlusion handling and multi-scale pedestrian detection based on deep learning: A review,” IEEE Access, vol. 10, pp. 19 937–19 957, 2022

    2022

  23. [31]

    Tracking pedestrian heads in dense crowd,

    R. Sundararaman, C. De Almeida Braga, E. Marchand, and J. Pettre, “Tracking pedestrian heads in dense crowd,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 3865–3875

    2021

  24. [32]

    Mask-guided attention network for occluded pedestrian detection,

    Y . Pang, J. Xie, M. H. Khan, R. M. Anwer, F. S. Khan, and L. Shao, “Mask-guided attention network for occluded pedestrian detection,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 4967–4975

    2019

  25. [33]

    Occlusion-aware r- cnn: Detecting pedestrians in a crowd,

    S. Zhang, L. Wen, X. Bian, Z. Lei, and S. Z. Li, “Occlusion-aware r- cnn: Detecting pedestrians in a crowd,” inProceedings of the European conference on computer vision (ECCV) , 2018, pp. 637–653

    2018

  26. [34]

    Nms by representative region: Towards crowded pedestrian detection by proposal pairing,

    X. Huang, Z. Ge, Z. Jie, and O. Yoshie, “Nms by representative region: Towards crowded pedestrian detection by proposal pairing,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 10 750–10 759

    2020

  27. [35]

    Detr for crowd pedestrian detection,

    M. Lin, C. Li, X. Bu, M. Sun, C. Lin, J. Yan, W. Ouyang, and Z. Deng, “Detr for crowd pedestrian detection,” arXiv preprint arXiv:2012.06785, 2020

    2012 arXiv

  28. [36]

    Crowdhuman: A benchmark for detecting human in a crowd,

    S. Shao, Z. Zhao, B. Li, T. Xiao, G. Yu, X. Zhang, and J. Sun, “Crowdhuman: A benchmark for detecting human in a crowd,” arXiv preprint arXiv:1805.00123, 2018

    2018 arXiv

  29. [37]

    Crowdpose: Efficient crowded scenes pose estimation and a new benchmark,

    J. Li, C. Wang, H. Zhu, Y . Mao, H.-S. Fang, and C. Lu, “Crowdpose: Efficient crowded scenes pose estimation and a new benchmark,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 10 863–10 872

    2019

  30. [38]

    Mot20: A bench- mark for multi object tracking in crowded scenes,

    P. Dendorfer, H. Rezatofighi, A. Milan, J. Shi, D. Cremers, I. Reid, S. Roth, K. Schindler, and L. Leal-Taix ´e, “Mot20: A bench- mark for multi object tracking in crowded scenes,” arXiv preprint arXiv:2003.09003, 2020

    2003 arXiv

  31. [39]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 10 012–10 022

    2021

  32. [40]

    Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,

    J. Philion and S. Fidler, “Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16. Springer, 2020, pp. 194–210

    2020

  33. [41]

    Bevdepth: Acquisition of reliable depth for multi-view 3d object detection,

    Y . Li, Z. Ge, G. Yu, J. Yang, Z. Wang, Y . Shi, J. Sun, and Z. Li, “Bevdepth: Acquisition of reliable depth for multi-view 3d object detection,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 2, 2023, pp. 1477–1485

    2023

  34. [42]

    Repair: Removing representation bias by dataset resampling,

    Y . Li and N. Vasconcelos, “Repair: Removing representation bias by dataset resampling,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 9572–9581

    2019

  35. [43]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2980–2988

    2017

  36. [44]

    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

    2018

  37. [45]

    Deep high-resolution representation learning for visual recognition,

    J. Wang, K. Sun, T. Cheng, B. Jiang, C. Deng, Y . Zhao, D. Liu, Y . Mu, M. Tan, X. Wang et al., “Deep high-resolution representation learning for visual recognition,” IEEE transactions on pattern analysis and machine intelligence, vol. 43, no. 10, pp. 3349–3364, 2020

    2020

  38. [46]

    Mot16: A benchmark for multi-object tracking,

    A. Milan, L. Leal-Taix ´e, I. Reid, S. Roth, and K. Schindler, “Mot16: A benchmark for multi-object tracking,” arXiv preprint arXiv:1603.00831, 2016

    2016 arXiv

  39. [47]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 21 674–21 683

    2023

Pith tools

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