Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

GRASPTrack claims that voxelized 3D overlap of mask-guided monocular point clouds, not 2D box overlap, is the association cue that resolves multi-object tracking under occlusion, and reports new best HOTA among compared trackers on MOT17, M

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A depth-aware MOT tracker using mask-guided 3D point clouds, voxelized 3D IoU association, adaptive Kalman noise, and 3D motion consistency surpasses prior TBD methods on MOT17, MOT20, and DanceTrack.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Plausible depth+segmentation add-on to OC-SORT with honest ablations, but the geometric core rests on unvalidated relative depth and no code. the 3 major comments →

arxiv 2508.08117 v1 pith:XAHV5BI3 submitted 2025-08-11 cs.CV cs.AI

GRASPTrack: Geometry-Reasoned Association via Segmentation and Projection for Multi-Object Tracking

Xudong Han , Pengcheng Fang , Yueying Tian , Jianhui Yu , Xiaohao Cai , Daniel Roggen , Philip Birch This is my paper
classification cs.CV cs.AI
keywords multi-object trackingtracking-by-detectionmonocular depth estimationinstance segmentationvoxel-based 3D IoUKalman filterocclusion handlingdata association
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper proposes a tracking-by-detection method that builds a 3D shape for each detected object by combining dense monocular depth maps with per-object segmentation masks, projecting only masked pixels into point clouds, and voxelizing those clouds to measure volumetric overlap between detections and tracks. It claims this voxel-based 3D IoU resolves occlusions that defeat 2D IoU, and adds two depth-aware motion mechanisms: adaptive Kalman-filter noise when an object is occluded, and motion-direction consistency measured in 3D. On three standard benchmarks the method reports the best HOTA among the trackers it compares against: 66.1 on MOT17, 64.5 on MOT20, and 65.3 on DanceTrack. If the result holds, it would mean that a monocular camera plus segmentation can supply enough geometric signal to preserve identities in crowded, complex-motion scenes without LiDAR or multi-view setups.

Core claim

The central discovery is that replacing 2D IoU with a depth-aware volumetric overlap metric substantially improves identity-consistent tracking under occlusion. The paper's pipeline is: a monocular depth estimator produces a dense depth map, an instance segmenter isolates each object, the pixels inside each mask are back-projected into camera coordinates to form a point cloud, and each cloud is discretized into a sparse binary voxel grid. Association scores are then computed as Voxel-Based 3D IoU, the ratio of shared occupied voxels to the union of occupied voxels between a detection and a track. The paper additionally extends the Kalman state with a depth term, scales the process noise acco

What carries the argument

Voxel-Based 3D IoU computed from mask-guided monocular point clouds. For each object, pixels inside its instance segmentation mask are projected through the pinhole camera model using an estimated depth value, forming a per-object 3D point cloud; two such clouds are placed in a shared voxel grid with a fixed voxel size, and the similarity metric is the number of occupied voxels in common divided by the number occupied by either cloud. This metric carries the association decision, replacing 2D box overlap and allowing objects that overlap on the image plane to be separated along the depth axis.

Load-bearing premise

The central claim rests on monocular depth estimates, linearly scaled to 0-255 per frame, being metrically consistent across frames and across objects, plus hand-aligned camera intrinsics being accurate; if depth scale or intrinsics drift, the point clouds, voxel overlap, and depth-state updates lose their meaning.

What would settle it

Take a crowded sequence and run GRASPTrack twice: once with the real per-frame depth maps, and once with each frame's depth values replaced by their within-frame rank order or by a constant. If HOTA does not drop, the depth geometry is not what is carrying the association. A sharper check is to multiply every frame's depth map by a random per-frame scale and observe whether Voxel-Based 3D IoU scores and the final tracks change; the claimed mechanism predicts they must.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the claim is correct, occlusion-driven identity switches should drop in crowded scenes because overlapping rectangles at different depths are separated by their volumetric shape rather than their projected area.
  • Objects moving mostly toward or away from the camera, where the 2D center barely changes, become tractable because the motion-direction consistency term, DOCM, uses displacement in 3D rather than only in the image plane.
  • Mask-guided projection should be most valuable in dense scenes, since including background and occluder pixels corrupts the point cloud; the paper reports a 1.1 HOTA gain on DanceTrack from masking versus full-box projection.
  • The design is modular on top of a standard tracking-by-detection baseline, so the same scoring ideas could be grafted onto other detectors and association pipelines without retraining a tracker.
  • Because the voxel grid is used only for scoring, not training, the metric can be adopted by existing trackers as a drop-in replacement for 2D IoU at the association stage.

Where Pith is reading between the lines

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

  • An extension the paper leaves implicit is that Voxel-Based 3D IoU is a generic volumetric shape similarity, so it could also serve re-identification or exemplar matching problems where appearance cues are ambiguous, not just multi-object tracking.
  • The method's assumption of per-frame depth consistency invites a testable upgrade: align depth scale and shift across frames using an estimated ground plane; if the geometric mechanism is real, association on depth-dominant motion should improve beyond the reported HOTA numbers.
  • The interactive per-sequence camera intrinsic calibration is a practical bottleneck; automating it would determine whether the tracked gains survive outside carefully hand-aligned benchmark sequences.
  • The ablation design suggests the depth-enhanced motion term and the adaptive noise scaling contribute differently across scene types, implying that future depth-aware trackers could tune these modules separately for crowded-static scenes versus uniform-appearance, complex-motion scenes such as DanceTrack.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. GRASPTrack is a tracking-by-detection MOT method that augments an OC-SORT-style pipeline with monocular depth estimation (Depth Anything v2) and instance segmentation (EfficientTAM). It projects pixels inside each instance mask to 3D via Eq. (3), voxelizes the per-object point clouds, and replaces 2D IoU with Voxel-Based 3D IoU for association. It also introduces Depth-Aware Adaptive Noise Compensation (DANC) to the Kalman filter and Depth-enhanced Observation-Centric Momentum (DOCM) in 3D. On the MOT17, MOT20, and DanceTrack test sets it reports HOTA scores of 66.1, 64.5, and 65.3, respectively, claiming superiority over the compared TBD methods; ablations on the DanceTrack validation set show monotonic improvements as components are added.

Significance. If the results hold, the paper demonstrates a practical way to inject geometric reasoning into monocular TBD tracking and shows consistent gains on a challenging benchmark. The ablation design is useful: it separates the contributions of the 3D IoU, DANC, and DOCM, and it tests the voxel-size and occlusion-sensitivity factors. The use of standard benchmarks and metrics aids comparability. The strongest limitations are reproducibility (no code or calibration release) and the unvalidated metric interpretation of monocular depth, which directly underpins the geometric components.

major comments (3)
  1. [Section III-A1, Eq. (3); Section IV-A3] Depth Anything v2 is a relative/disparity model; linearly scaling each frame's depth map to [0,255] does not produce metric Z. Since Eq. (3) treats Z as camera-coordinate depth, the point clouds, Eq. (7) voxelization, Eq. (9) state depth, Eq. (10) occlusion depth comparisons, and Eq. (14) DOCM all inherit unknown per-frame scale/shift. No cross-frame or cross-object consistency validation is provided. This is load-bearing for the claimed 'high-fidelity 3D point clouds' and Voxel-Based 3D IoU. The authors should either use a metric depth source, calibrate scale/shift (e.g., via the estimated ground plane or known object sizes), or provide experiments that validate depth consistency and show the results are not artifacts of normalization.
  2. [Section IV-A3] Camera intrinsics (fx, fy, cx, cy) are estimated interactively per sequence following UCMCTrack, and no calibration files, code, or automated procedure are released. All geometric computations in Eq. (3) depend on these values. This is a reproducibility gap and a potential source of overfitting to the test sequences; the paper should release the estimated intrinsics or a script, and analyze sensitivity to calibration errors.
  3. [Section IV-C, Tables IV-V] The hyperparameters δv = 0.4 and α = 3 are selected on the DanceTrack validation set, and τIoU = 0.6 is fixed without an ablation. Because these parameters interact with the depth scale and scene geometry, tuning them on one benchmark does not establish that the SOTA results generalize. A sensitivity analysis on MOT17/MOT20 or a principled choice (e.g., tied to object scale) is needed for the cross-dataset claim.
minor comments (6)
  1. [Eq. (3)] The notation '(cx, cy)' is described as 'the center point of the box corresponding to the object', but in the pinhole camera model it should be the principal point. Please correct.
  2. [Eq. (10)] The text calls τIoU a 'spatial overlap threshold' and the surrounding text mentions 3D IoU, but Eq. (10) computes standard 2D bounding-box IoU. Clarify which IoU is used.
  3. [Table III] The check-mark alignment in the table is ambiguous: it is unclear which configuration corresponds to each row, making it hard to verify the claim that DOCM gives the largest gain. Please reformat.
  4. [Section I] Several reference citations appear as '??' in the text (e.g., [4], [6], [9]), indicating broken LaTeX cross-references. Please fix these before final submission.
  5. [Section IV-C, Table IV] The FPS column likely measures only part of the pipeline; specify exactly which modules are included in the reported FPS.
  6. [General] The claim 'superior performance compared to other methods' should be qualified as 'compared to the TBD methods listed here', since the selected baseline set is not exhaustive and some recent methods are omitted.

Circularity Check

0 steps flagged

No significant circularity: the paper's benchmark results are empirical and its components are defined from external inputs, not from the target metrics.

full rationale

GRASPTrack's central claim is a set of benchmark scores (HOTA 66.1/64.5/65.3 on MOT17/MOT20/DanceTrack) obtained by running a tracking pipeline on test data. The method's components—mask-guided projection (Eq. 3), voxelization (Eq. 7), voxel IoU (Eq. 8), DANC (Eqs. 10–13), and DOCM (Eqs. 14–15)—are explicit functions of depth estimates, segmentation masks, detections, and Kalman states. None of these equations defines a benchmark metric in terms of itself, and none of the reported scores is a fitted parameter renamed as a prediction. The hyperparameters δv, α, and τIoU are tuned on the DanceTrack validation set (Tables IV and V) and then frozen for test evaluation; this is standard model selection rather than circular reasoning. The only self-citation is reference [15] (ETrack), which appears in a list of related TBD methods and is not used to justify the method's correctness, uniqueness, or performance. There is no imported uniqueness theorem and no ansatz smuggled in via citation. The paper's reliance on unvalidated monocular depth scale consistency and interactively estimated camera intrinsics is a legitimate empirical robustness concern, but it is not circularity: treating Depth Anything v2 output as metric depth may be unjustified, but it is not equivalent to assuming the paper's conclusion. The derivation chain does not reduce to its inputs, so the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on empirical engineering choices: three hand-set hyperparameters, interactively estimated intrinsics, and an unvalidated cross-frame depth scale consistency. No new physical or model entities are introduced.

free parameters (5)
  • Voxel size delta_v = 0.4
    Tuned on the DanceTrack validation set (Table IV). Controls the granularity of 3D IoU and the FPS trade-off.
  • Occlusion sensitivity factor alpha = 3
    Tuned on the DanceTrack validation set (Table V). Scales Kalman process noise under occlusion.
  • Occlusion overlap threshold tau_IoU = 0.6
    Set in Section IV-A3 without an ablation study. Determines when a detection is considered occluded.
  • Depth scaling = linear scale to [0,255]
    Section IV-A3: depth maps from Depth Anything v2 are linearly scaled to [0,255], imposing an arbitrary per-frame scale that affects all 3D geometry.
  • Camera intrinsics (fx, fy, cx, cy) = estimated per sequence by interactive alignment
    Section IV-A3: estimated by interactively aligning a ground-plane grid with each image. This is not automated, not released, and affects every 3D projection.
axioms (4)
  • domain assumption Monocular depth maps are metrically consistent across frames and across objects after linear scaling to [0,255].
    Invoked in Eq. (3) projection and Eq. (9) state. Depth Anything v2 returns affine-invariant depth; no cross-frame alignment is described.
  • domain assumption EfficientTAM segmentation masks for detection boxes are accurate enough that mask-guided projection removes background and occluder pixels.
    Section III-A1: the claimed advantage over full-bounding-box projection depends on mask quality.
  • domain assumption The shared pmin/pmax bounding volume in voxelization defines a meaningful comparison for IoU.
    Equations (5)-(7): for pairs at very different depths, the common voxel grid can distort overlap granularity.
  • standard math Pinhole camera projection with estimated intrinsics applies to benchmark videos.
    Eq. (3) uses standard pinhole projection; assumes a calibrated camera model.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GRASPTrack: Geometry-Reasoned Association via Segmentation and Projection for Multi-Object Tracking." pith.science (2026). https://pith.science/paper/XAHV5BI3

@misc{pith2026250808117,
  author       = {Pith},
  title        = {Pith review of: GRASPTrack: Geometry-Reasoned Association via Segmentation and Projection for Multi-Object Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XAHV5BI3}},
  note         = {Machine review of arXiv:2508.08117}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multi-object tracking (MOT) in monocular videos is fundamentally challenged by occlusions and depth ambiguity, issues that conventional tracking-by-detection (TBD) methods struggle to resolve owing to a lack of geometric awareness. To address these limitations, we introduce GRASPTrack, a novel depth-aware MOT framework that integrates monocular depth estimation and instance segmentation into a standard TBD pipeline to generate high-fidelity 3D point clouds from 2D detections, thereby enabling explicit 3D geometric reasoning. These 3D point clouds are then voxelized to enable a precise and robust Voxel-Based 3D Intersection-over-Union (IoU) for spatial association. To further enhance tracking robustness, our approach incorporates Depth-aware Adaptive Noise Compensation, which dynamically adjusts the Kalman filter process noise based on occlusion severity for more reliable state estimation. Additionally, we propose a Depth-enhanced Observation-Centric Momentum, which extends the motion direction consistency from the image plane into 3D space to improve motion-based association cues, particularly for objects with complex trajectories. Extensive experiments on the MOT17, MOT20, and DanceTrack benchmarks demonstrate that our method achieves competitive performance, significantly improving tracking robustness in complex scenes with frequent occlusions and intricate motion patterns.

Figures

Figures reproduced from arXiv: 2508.08117 by Daniel Roggen, Jianhui Yu, Pengcheng Fang, Philip Birch, Xiaohao Cai, Xudong Han, Yueying Tian.

Figure 1
Figure 1. Figure 1: An illustration of associating occluded detections in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The pipeline of the proposed GRASPTrack. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. CylindTrack: Depth-Aware Cylindrical Motion Modeling for Panoramic Multi-Object Tracking

    cs.CV 2026-06 unverdicted novelty 6.0

    CylindTrack improves identity preservation in panoramic multi-object tracking by combining depth-temporal trajectory modeling, spherical spatio-temporal consistency learning, and topology-aware cylindrical motion prediction.

  2. Handling fabrication defects in hex-grid surface codes

    quant-ph 2025-08 unverdicted novelty 4.0

    Extensions to the LUCI framework keep hex-grid surface codes resilient to isolated broken qubits and couplers, with circuit distance dropping by at most one.

Reference graph

Works this paper leans on

35 extracted references · 29 canonical work pages · cited by 2 Pith papers

  1. [1]

    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, pp. 2446–2454. I

    work page 2020

  2. [2]

    Glamr: Global occlusion-aware human mesh recovery with dynamic cameras,

    Y . Yuan, U. Iqbal, P. Molchanov, K. Kitani, and J. Kautz, “Glamr: Global occlusion-aware human mesh recovery with dynamic cameras,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11 038–11 049. I

    work page 2022

  3. [3]

    Tracking systems in team sports: a narrative review of applications of the data and sport specific analysis,

    L. Torres-Ronda, E. Beanland, S. Whitehead, A. Sweeting, and J. Clubb, “Tracking systems in team sports: a narrative review of applications of the data and sport specific analysis,” Sports Medicine-Open, vol. 8, no. 1, p. 15, 2022. I

    work page 2022

  4. [4]

    Sparsetrack: Multi- object tracking by performing scene decomposition based on pseudo- depth,

    Z. Liu, X. Wang, C. Wang, W. Liu, and X. Bai, “Sparsetrack: Multi- object tracking by performing scene decomposition based on pseudo- depth,” IEEE Transactions on Circuits and Systems for Video Technol- ogy, 2025. I, II-B, ??

    work page 2025

  5. [5]

    Camot: Camera angle-aware multi- object tracking,

    F. Limanta, K. Uto, and K. Shinoda, “Camot: Camera angle-aware multi- object tracking,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2024, pp. 6479–6488. I, II-B

    work page 2024

  6. [6]

    Quo vadis: Is trajectory forecasting the key towards long-term multi-object track- ing?

    P. Dendorfer, V . Yugay, A. Osep, and L. Leal-Taix ´e, “Quo vadis: Is trajectory forecasting the key towards long-term multi-object track- ing?” Advances in Neural Information Processing Systems , vol. 35, pp. 15 657–15 671, 2022. I, II-B, ??

    work page 2022

  7. [7]

    Detecting invisible people,

    T. Khurana, A. Dave, and D. Ramanan, “Detecting invisible people,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 3174–3184. I

    work page 2021

  8. [8]

    Contributions to the theory of optimal control,

    R. E. Kalman et al. , “Contributions to the theory of optimal control,” Bol. soc. mat. mexicana , vol. 5, no. 2, pp. 102–119, 1960. I

    work page 1960

  9. [9]

    Observation- centric sort: Rethinking sort for robust multi-object tracking,

    J. Cao, J. Pang, X. Weng, R. Khirodkar, and K. Kitani, “Observation- centric sort: Rethinking sort for robust multi-object tracking,” in Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 9686–9696. I, II-A, II-A, III-B, ??

    work page 2023

  10. [10]

    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,

    Pith/arXiv arXiv

  11. [11]

    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. I, IV-A1

    Pith/arXiv arXiv 2003

  12. [12]

    Dancetrack: Multi-object tracking in uniform appearance and diverse motion,

    P. Sun, J. Cao, Y . Jiang, Z. Yuan, S. Bai, K. Kitani, and P. Luo, “Dancetrack: Multi-object tracking in uniform appearance and diverse motion,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 20 993–21 002. I, IV-A1

    work page 2022

  13. [13]

    Simple online and realtime tracking,

    A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft, “Simple online and realtime tracking,” in 2016 IEEE international conference on image processing (ICIP). Ieee, 2016, pp. 3464–3468. II-A, II-A, III-B

    work page 2016

  14. [14]

    Bytetrack: Multi-object tracking by associating every detection box,

    Y . Zhang, P. Sun, Y . Jiang, D. Yu, F. Weng, Z. Yuan, P. Luo, W. Liu, and X. Wang, “Bytetrack: Multi-object tracking by associating every detection box,” in European conference on computer vision . Springer, 2022, pp. 1–21. II-A, II-A, III-B, ??, IV-A3

    work page 2022

  15. [15]

    Ettrack: enhanced temporal motion predictor for multi-object tracking,

    X. Han, N. Oishi, Y . Tian, E. Ucurum, R. Young, C. Chatwin, and P. Birch, “Ettrack: enhanced temporal motion predictor for multi-object tracking,” Applied Intelligence, vol. 55, no. 1, p. 33, 2025. II-A, III-B JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 8

    work page 2025

  16. [16]

    Simple online and realtime tracking with a deep association metric,

    N. Wojke, A. Bewley, and D. Paulus, “Simple online and realtime tracking with a deep association metric,” in 2017 IEEE international conference on image processing (ICIP) . IEEE, 2017, pp. 3645–3649. II-A

    work page 2017

  17. [17]

    Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification,

    G. Maggiolino, A. Ahmad, J. Cao, and K. Kitani, “Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification,” in 2023 IEEE International conference on image processing (ICIP) . IEEE, 2023, pp. 3025–3029. II-A, ??

    work page 2023

  18. [18]

    Ucmc- track: Multi-object tracking with uniform camera motion compensation,

    K. Yi, K. Luo, X. Luo, J. Huang, H. Wu, R. Hu, and W. Hao, “Ucmc- track: Multi-object tracking with uniform camera motion compensation,” in Proceedings of the AAAI conference on artificial intelligence , vol. 38, no. 7, 2024, pp. 6702–6710. II-A, ??, IV-A3

    work page 2024

  19. [19]

    Ab3dmot: A baseline for 3d multi-object tracking and new evaluation metrics,

    X. Weng, J. Wang, D. Held, and K. Kitani, “Ab3dmot: A baseline for 3d multi-object tracking and new evaluation metrics,” arXiv preprint arXiv:2008.08063, 2020. II-B

    Pith/arXiv arXiv 2008

  20. [20]

    Center-based 3d object detection and tracking,

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

    work page 2021

  21. [21]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,” Advances in Neural Information Processing Sys- tems, vol. 37, pp. 21 875–21 911, 2024. III-A, IV-A3

    work page 2024

  22. [22]

    Efficient track anything,

    Y . Xiong, C. Zhou, X. Xiang, L. Wu, C. Zhu, Z. Liu, S. Suri, B. Varadarajan, R. Akula, F. Iandola et al. , “Efficient track anything,” arXiv preprint arXiv:2411.18933 , 2024. III-A, IV-A3

    Pith/arXiv arXiv 2024

  23. [23]

    Spherical mask: Coarse-to-fine 3d point cloud instance segmentation with spher- ical representation,

    S. Shin, K. Zhou, M. Vankadari, A. Markham, and N. Trigoni, “Spherical mask: Coarse-to-fine 3d point cloud instance segmentation with spher- ical representation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 4060–4069. III-A2

    work page 2024

  24. [24]

    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,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 4490–4499. III-A2

    work page 2018

  25. [25]

    Hard to track objects with irregular motions and similar appearances? make it easier by buffering the matching space,

    F. Yang, S. Odashima, S. Masui, and S. Jiang, “Hard to track objects with irregular motions and similar appearances? make it easier by buffering the matching space,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision , 2023, pp. 4799–4808. ??

    work page 2023

  26. [26]

    Motiontrack: Learning robust short-term and long-term motions for multi-object tracking,

    Z. Qin, S. Zhou, L. Wang, J. Duan, G. Hua, and W. Tang, “Motiontrack: Learning robust short-term and long-term motions for multi-object tracking,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 17 939–17 948. ??

    work page 2023

  27. [27]

    Bot-sort: Robust associa- tions multi-pedestrian tracking,

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

    Pith/arXiv arXiv 2022

  28. [28]

    Simple cues lead to a strong multi-object tracker,

    J. Seidenschwarz, G. Bras ´o, V . C. Serrano, I. Elezi, and L. Leal-Taix ´e, “Simple cues lead to a strong multi-object tracker,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 13 813–13 823. ??

    work page 2023

  29. [29]

    Strong- sort: Make deepsort great again,

    Y . Du, Z. Zhao, Y . Song, Y . Zhao, F. Su, T. Gong, and H. Meng, “Strong- sort: Make deepsort great again,” IEEE Transactions on Multimedia , vol. 25, pp. 8725–8737, 2023. ??

    work page 2023

  30. [30]

    Diffmot: A real-time diffusion-based multiple object tracker with non-linear prediction,

    W. Lv, Y . Huang, N. Zhang, R.-S. Lin, M. Han, and D. Zeng, “Diffmot: A real-time diffusion-based multiple object tracker with non-linear prediction,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2024, pp. 19 321–19 330. ??, IV-A3

    work page 2024

  31. [31]

    Temporal coherent object flow for multi-object tracking,

    Z. Song, R. Luo, L. Ma, Y . Tang, Y .-P. P. Chen, J. Yu, and W. Yang, “Temporal coherent object flow for multi-object tracking,” in Proceed- ings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 7, 2025, pp. 6978–6986. ??

    work page 2025

  32. [32]

    Evaluating multiple object tracking performance: the clear mot metrics,

    K. Bernardin and R. Stiefelhagen, “Evaluating multiple object tracking performance: the clear mot metrics,” EURASIP Journal on Image and Video Processing, vol. 2008, no. 1, p. 246309, 2008. IV-A2

    work page 2008

  33. [33]

    Performance measures and a data set for multi-target, multi-camera tracking,

    E. Ristani, F. Solera, R. Zou, R. Cucchiara, and C. Tomasi, “Performance measures and a data set for multi-target, multi-camera tracking,” in European conference on computer vision . Springer, 2016, pp. 17–35. IV-A2

    work page 2016

  34. [34]

    Hota: A higher order metric for evaluating multi-object tracking,

    J. Luiten, A. Osep, P. Dendorfer, P. Torr, A. Geiger, L. Leal-Taix ´e, and B. Leibe, “Hota: A higher order metric for evaluating multi-object tracking,” International journal of computer vision , vol. 129, no. 2, pp. 548–578, 2021. IV-A2, IV-A2

    work page 2021

  35. [35]

    Yolox: Exceeding yolo series in 2021,

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “Yolox: Exceeding yolo series in 2021,” arXiv preprint arXiv:2107.08430 , 2021. IV-A3

    Pith/arXiv arXiv 2021

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.