Pith. sign in

REVIEW 2 major objections 6 minor 81 references

MVTrajecter: Multi-View Pedestrian Tracking with Trajectory Motion Cost and Trajectory Appearance Cost

T0 review · 2 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Multi-view pedestrian tracking becomes more reliable when identity association sums motion and appearance evidence over several past timestamps instead of one; MVTrajecter's 7-frame window reports new state-of-the-art results on three bench

desk verdict A credible multi-timestamp association mechanism for end-to-end MVPT, with a real gap in how missing detections are handled and a minor test-set leak. read the letter →

arxiv 2509.01157 v1 pith:WBVAKOFN submitted 2025-09-01 cs.CV

Taiga Yamane , Ryo Masumura , Satoshi Suzuki , Shota Orihashi This is my paper · ORCID
classification cs.CV
keywords multi-viewpedestriantrackingbird'seyeviewtrajectoryassociationmotioncostappearanceattentionmechanismend-to-endoccupancymap
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

Multi-view pedestrian tracking asks which bird's-eye-view detections of people at the current instant continue the tracks built from earlier instants. Previous end-to-end methods compare the current frame with only the immediately preceding one, so a single bad match — occlusion, crowding, a wobbly detection — can break an identity. MVTrajecter's claim is that the comparison should collect evidence from many past timestamps and let the whole trajectory vote: it introduces two costs, a trajectory motion cost and a trajectory appearance cost, each a sum over the past K frames of a per-frame motion distance or identity probability. With K = 7 the tracker reports state-of-the-art results on GMVD, Wildtrack, and MultiviewX, and its own ablation shows steady improvement as K grows from 1 to 7, and from adding each cost separately. If the claim is right, window length is a first-class design axis for end-to-end multi-view trackers.

What carries the argument

The two load-bearing objects are trajectory motion cost (TMC) and trajectory appearance cost (TAC). TMC(i,j) = Σ_k ||p^t_i − (p^{t−k}_j + m^{t−k}_j)|| sums, over all past timestamps in the window, the distance between the current detected position and the position reached by propagating each past position with its predicted motion. TAC(i,j) = −Σ_k Pr^{t−k}(j|i) sums the softmax identity probability from appearance-feature dot products. Summing turns the association into a vote over the track: one bad timestamp can be outweighed by the others. A shared attention mechanism, with sinusoidal temporal embeddings added to pedestrian features, produces the motions and appearance features so that ea

What would settle it

Run MVTrajecter on a sequence where a substantial fraction of pedestrians are undetectable for stretches longer than the K=7 window, or where t−1 is synthetically occluded; if IDF1 with K=7 does not exceed K=1 under these conditions — or if the network cannot even produce the required past features — the redundancy argument fails. Also, a direct check of the α=0.98 weighting: setting TAC's contribution to zero (relying on TMC alone) should drop performance measurably if the appearance term is truly load-bearing rather than a numerical rounding error given the reported ranges.

Watch

Extended reading notes

Core claim

The paper proposes that an end-to-end multi-view pedestrian tracker should associate a current detection with a past track by aggregating association evidence across the whole track window. For a candidate pair (current pedestrian i, past trajectory j), it computes the trajectory motion cost, the sum over k = 1..K of the Euclidean distance between i's detected position and the position j had at t−k advanced by a predicted motion; and the trajectory appearance cost, the negative sum over k of the probability that i is j's identity given dot-product similarity of appearance features at t and t−k. The two costs are weighted and solved with Hungarian matching. Training is end-to-end with losses

Load-bearing premise

The association costs assume every pedestrian has an appearance and motion feature at every past timestamp in the window; the paper says a missing location is handled by a Kalman filter, but it never specifies how the network produces the features (Eqs. 5–7) that feed the costs for a pedestrian who was not detected at a past timestamp, which is exactly the common occlusion case.

Editorial extensions

If this is right

  • If the central claim holds, end-to-end multi-view trackers should treat history length K as a tunable axis: the paper's ablation shows MOTA rising from 74.2 (K=1) to 78.1 (K=7) on GMVD.
  • Motion and appearance evidence are complementary: adding TMC to a motion-only baseline raises IDF1 from 67.8 to 71.8, adding TAC to an appearance-only baseline raises it from 64.7 to 72.3, and both together reach 77.2.
  • Attention-based motion prediction beats a Kalman-filter alternative (IDF1 77.2 vs 76.0), supporting the view that multi-timestamp relational context improves motion forecasting.
  • The association mechanism transfers across datasets of different camera counts and density levels (Wildtrack, MultiviewX, GMVD), suggesting the costs generalize rather than overfitting one benchmark.

Reading between the lines

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

  • Windowed voting is a general principle: the same accumulate-costs-over-a-trajectory trick could transfer to monocular trackers or graph-based global association, where the single-frame cost is the usual bottleneck — the paper only demonstrates it in BEV multi-view tracking.
  • The α=0.98 weighting plus the reported value ranges (CTMC on the order of 10^3, CTAC near 10^−25) implies the motion term practically decides the Hungarian assignment; expect that improving motion prediction quality matters more than refining appearance similarity.
  • The paper's K=7 ceiling is a memory limit and its own conclusion admits failure when detection fails for more than 7 consecutive frames; a natural testable extension is per-trajectory adaptive window length, or a recurrent/compressed memory that extends history beyond 7 frames.
  • A stress test for the claim: deliberately occlude the adjacent-past timestamp (t−1) in a fraction of frames; the multi-timestamp method should degrade much less than K=1 baselines if the redundancy argument is right.
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

2 major / 6 minor

Summary. The paper proposes MVTrajecter, an end-to-end multi-view pedestrian tracking method that extends association beyond the adjacent past timestamp by using information from K past timestamps. It introduces a trajectory motion cost (TMC) and a trajectory appearance cost (TAC), each formed by summing per-timestamp motion/appearance evidence over the past K timestamps, and combines them with a weighting parameter before Hungarian matching. The architecture uses attention layers to predict motions and appearance features from BEV pedestrian features. Experiments on GMVD, Wildtrack, and MultiviewX report state-of-the-art results, supported by ablations on trajectory length, cost components, attention versus Kalman filtering, and inference speed.

Significance. The paper addresses a clear gap in end-to-end MVPT, where previous methods restrict association to the current and adjacent past timestamp. The proposed aggregation over multiple timestamps is simple, well-motivated by analogous developments in monocular tracking, and the empirical gains in Tables 5 and 6 are substantial (e.g., +6.2 IDF1 from K=1 to K=7). The authors provide thorough ablations, including HOTA and inference speed, and make the comparison fair by re-implementing previous methods on GMVD and also testing with identical detections. If the missing-detection handling is resolved, this would be a solid contribution that could influence future MVPT design.

major comments (2)
  1. [Sec. 3.3-3.5 (Eqs. 5-7, 9-10)] The paper defines p^{t-k}_j = ∅ for missing detections (Sec. 3.1) and reports detection recall of only 79.4% on GMVD (Table 10), so missing detections are common. However, the procedure is undefined for missing pedestrians. Pedestrian features f^{t-k}_j are extracted only at detected locations (Sec. 3.4: f^n_t = B^t(p^n_t)). Eqs. 5 and 7 require f^{t-k}_j to compute m^{t-k}_j and d^{t-k}_j; the Kalman-filter fallback for TMC in Sec. 3.3 does not provide m^{t-k}_j to the motion branch, nor does it specify how attention handles a missing token. The training losses (Eqs. 9-10) sum over all n,k and are undefined when p^{t-k}_n=∅. This is a load-bearing gap because the central claim is that aggregation over K > 1 timestamps improves association, and missing detections are more likely with larger K. Please specify the actual handling (masking, imputation, dropping) and evaluate its impact; the
  2. [Sec. 4.3 and Tables 3-4] The hyperparameters (K, α, detection and Hungarian thresholds) are tuned on the MultiviewX validation set, which is the validation data for GMVD experiments (Sec. 4.1: 'Following Vora et al., we use MultiviewX as the validation data in the experiments on GMVD'). The same hyperparameters are then used to report MultiviewX test results in Table 4. Consequently, the MultiviewX result is not an independent test of generalization; it may be optimistic because model selection used validation data from the same distribution. I recommend disclosing this explicitly and ideally reporting MultiviewX results with hyperparameters chosen without access to MultiviewX (e.g., from GMVD validation), or at least discussing the optimism. The GMVD claims (Tables 1-2) are not affected.
minor comments (6)
  1. [Sec. 3.4, Eqs. 5/7] The notation {f^{t-k}_1,...,f^{t-k}_N}^K_{k=0} is ambiguous; please define it as the concatenation of all pedestrian features across timestamps t-K,...,t.
  2. [Sec. 4.3] State for each dataset which split was used for hyperparameter tuning; 'validation data' is ambiguous because GMVD's validation is MultiviewX.
  3. [Supp. E] The reported value range for CTAC, from -6.2 to -4.7e-25, is surprising: Eq. 2 gives probabilities in [0,1], so a sum of K=7 terms should lie in [-7,0]; a minimum of -4.7e-25 appears to be a typo or numerical artifact. Please verify.
  4. [Table 2] Please clarify how the 'same detection results as ours' condition was implemented (e.g., feeding MVTrajecter's detections to all trackers).
  5. [Sec. 3.1] The formulation assumes a fixed number N of pedestrians at all timestamps, but in tracking N changes over time (entries/exits). Please explain how the association matrix is formed when the number of past trajectories differs from the number of current detections.
  6. [Author affiliation] Typo: 'Laboratries' should be 'Laboratories'.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the multi-timestamp association gains are empirical and externally benchmarked; the sole self-citation is a non-load-bearing related-work mention.

full rationale

The paper's central claim is that aggregating motion and appearance association evidence over multiple past timestamps (TMC and TAC) improves multi-view pedestrian tracking. This claim is supported by controlled experiments on held-out test splits of GMVD, Wildtrack, and MultiviewX, with ablations over K, TMC/TAC, attention vs. Kalman filter, and EMA aggregation. The training losses LTMC and LTAC minimize the same per-timestamp motion and appearance objectives used in the inference costs, but that is standard supervised consistency rather than circular derivation: the reported MOTA/IDF1 numbers are measured against ground-truth tracks on test data, not by the training loss. Hyperparameters such as α and detection thresholds are tuned on validation data, not fitted to the test metrics. The only author self-citation is [61] (Suzuki, Tora, Masumura) in a related-work list of data augmentations; it is not used to justify the proposed architecture or to establish the SOTA claim, so it is not load-bearing. The paper also states a genuine limitation: past information is limited to 7 timestamps due to GPU memory, and it cannot track someone whose detection continuously fails for more than 7 timestamps. A related robustness gap is that the handling of p^{t-k}_j = ∅ is underspecified for the attention branches (Eqs. 5 and 7 require f^{t-k}_j, but features are only extracted at detected locations). These are correctness/robustness concerns, not circularity, and do not raise the circularity score beyond the minor self-citation tier.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities. It relies on standard network components and dataset assumptions. The key free parameters are the cost weighting and trajectory length, which are dataset-sensitive and warrant careful re-tuning on each new deployment.

free parameters (4)
  • alpha = 0.98
    Weighting between TMC and TAC in the overall cost C(i,j) = (1-alpha)CTMC + alpha*CTAC. Tuned on validation data and shown to be sensitive in Table 13 (MOTA varies by 2.5 points between 0.97 and 0.98).
  • K (trajectory length) = 7
    Number of past timestamps used in TMC and TAC. Chosen as the maximum feasible under GPU memory, but performance improves monotonically with K (Table 5), so it functions as a capacity parameter rather than a fixed design choice.
  • detection threshold = 0.4
    Threshold on the occupancy map to extract pedestrian locations; tuned on validation data.
  • Hungarian matching threshold = 0.1
    Threshold for accepting an association from the Hungarian algorithm; tuned on validation data.
assumptions (5)
  • domain assumption Calibrated cameras and ground-plane homographies are available for all datasets.
    The perspective transformation in Supp. A requires intrinsic and extrinsic calibration, which is standard for Wildtrack, MultiviewX, and GMVD.
  • domain assumption BEV features at detected pedestrian locations contain sufficient appearance and motion information for identity association over K timestamps.
    The entire association branch operates on B^t(p^t_n), the BEV feature at the detected location. This is a design premise not independently verified.
  • domain assumption The attention mechanism can capture temporal changes and identity matching from pedestrian features.
    The paper justifies this by citing prior work [44,70,74,26,80,81], but it is an architectural assumption that the learned features are sufficient for the attention to recover identity and motion.
  • domain assumption The softmax normalization in Eq. 2 over all N pedestrians at each timestamp yields valid identity probabilities.
    This assumes that the set of pedestrians at each timestamp is exactly the set that should be compared, and that the dot-product similarity is a proper score for identity.
  • domain assumption Ground-truth occupancy maps and trajectories are available for training and evaluation on all three datasets.
    The losses (Eqs. 8, 9, 10) rely on ground-truth annotations, which are provided by the datasets but are external to the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MVTrajecter: Multi-View Pedestrian Tracking with Trajectory Motion Cost and Trajectory Appearance Cost." pith.science (2026). https://pith.science/paper/WBVAKOFN

@misc{pith2026250901157,
  author       = {Pith},
  title        = {Pith review of: MVTrajecter: Multi-View Pedestrian Tracking with Trajectory Motion Cost and Trajectory Appearance Cost},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBVAKOFN}},
  note         = {Machine review of arXiv:2509.01157}
}
read the original abstract

Multi-View Pedestrian Tracking (MVPT) aims to track pedestrians in the form of a bird's eye view occupancy map from multi-view videos. End-to-end methods that detect and associate pedestrians within one model have shown great progress in MVPT. The motion and appearance information of pedestrians is important for the association, but previous end-to-end MVPT methods rely only on the current and its single adjacent past timestamp, discarding the past trajectories before that. This paper proposes a novel end-to-end MVPT method called Multi-View Trajectory Tracker (MVTrajecter) that utilizes information from multiple timestamps in past trajectories for robust association. MVTrajecter introduces trajectory motion cost and trajectory appearance cost to effectively incorporate motion and appearance information, respectively. These costs calculate which pedestrians at the current and each past timestamp are likely identical based on the information between those timestamps. Even if a current pedestrian could be associated with a false pedestrian at some past timestamp, these costs enable the model to associate that current pedestrian with the correct past trajectory based on other past timestamps. In addition, MVTrajecter effectively captures the relationships between multiple timestamps leveraging the attention mechanism. Extensive experiments demonstrate the effectiveness of each component in MVTrajecter and show that it outperforms the previous state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2509.01157 by the authors.

Figure 1
Figure 1. Comparison of (a) previous end-to-end MVPT meth [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) Overview of MVTrajecter. It consists of a (b) detection network, (c) motion branch, (d) appearance branch, TMC, and TAC [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. MODA on the validation data of models trained with our [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Sample frames of Wildtrack [9], MultiviewX [32], and GMVD [65]. Top to bottom rows represent Wildtrack, MultiviewX, and GMVD, respectively. Left and right columns show the training split and testing split of each dataset, respectively. While GMVD contains 7 scenes, we …
Figure 5
Figure 5. Figure 5: Qualitative comparison of detection results between ground truth (GT), TrackTacular [ [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of tracking results between ground truth (GT), TrackTacular [ [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison of tracking results between ground truth (GT), TrackTacular [ [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of tracking results between ground truth (GT), TrackTacular [ [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

81 extracted references · 72 canonical work pages

  1. [1]

    BoT-SORT: Robust associations multi-pedestrian tracking

    Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. BoT-SORT: Robust associations multi-pedestrian tracking. arXiv:2206.14651, 2022. 3, 4

    arXiv 2022

  2. [2]

    Enhancing multi-view pedestrian detection through generalized 3D feature pulling

    Sithu Aung, Haesol Park, Hyungjoo Jung, and Junghyun Cho. Enhancing multi-view pedestrian detection through generalized 3D feature pulling. In WACV, 2024. 3, 2

    2024

  3. [3]

    Deep oc- clusion reasoning for multi-camera multi-target detection

    Pierre Baqu ´e, Franc ¸ois Fleuret, and Pascal Fua. Deep oc- clusion reasoning for multi-camera multi-target detection. In ICCV, 2017. 3

    work page 2017

  4. [4]

    Berclaz, F

    J. Berclaz, F. Fleuret, and P. Fua. Robust people tracking with global trajectory optimization. In CVPR, 2006. 1, 3

    work page 2006

  5. [5]

    Learning a neural solver for multiple object tracking

    Guillem Bras ´o and Laura Leal-Taix ´e. Learning a neural solver for multiple object tracking. In CVPR, 2020. 3

    work page 2020

  6. [6]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In ECCV, 2020. 3

    work page 2020

  7. [7]

    Uni- fying short and long-term tracking with graph hierarchies

    Orcun Cetintas, Guillem Bras ´o, and Laura Leal-Taix´e. Uni- fying short and long-term tracking with graph hierarchies. In CVPR, 2023. 3

    work page 2023

  8. [8]

    Deep multi- camera people detection

    Tatjana Chavdarova and Franc ¸ois Fleuret. Deep multi- camera people detection. In ICMLA, 2017. 3

    work page 2017

Show all 81 references
  1. [9]

    Wildtrack: A multi-camera hd dataset for dense unscripted pedestrian detection

    Tatjana Chavdarova, Pierre Baqu ´e, St ´ephane Bouquet, An- drii Maksai, Cijo Jose, Timur Bagautdinov, Louis Lettry, Pascal Fua, Luc Van Gool, and Franc ¸ois Fleuret. Wildtrack: A multi-camera hd dataset for dense unscripted pedestrian detection. In CVPR, 2018. 1, 2, 6, 3

    2018

  2. [10]

    ReST: A reconfigurable spatial-temporal graph model for multi-camera multi-object tracking

    Cheng-Che Cheng, Min-Xuan Qiu, Chen-Kuo Chiang, and Shang-Hong Lai. ReST: A reconfigurable spatial-temporal graph model for multi-camera multi-object tracking. In ICCV, 2023. 3, 6, 7

    2023

  3. [11]

    Online multi- camera people tracking with spatial-temporal mechanism and anchor-feature hierarchical clustering

    Riu Cherdchusakulchai, Sasin Phimsiri, Visarut Trairat- tanapa, Suchat Tungjitnob, Wasu Kudisthalert, Pornprom Ki- awjak, Ek Thamwiwatthana, Phawat Borisuitsawat, Teep- akorn Tosawadi, Pakcheera Choppradi, et al. Online multi- camera people tracking with spatial-temporal mecha...

    2024

  4. [12]

    SportsMOT: A large multi- object tracking dataset in multiple sports scenes

    Yutao Cui, Chenkai Zeng, Xiaoyu Zhao, Yichun Yang, Gangshan Wu, and Limin Wang. SportsMOT: A large multi- object tracking dataset in multiple sports scenes. In ICCV,

  5. [13]

    Learning a proposal classifier for multiple object tracking

    Peng Dai, Renliang Weng, Wongun Choi, Changshui Zhang, Zhangping He, and Wei Ding. Learning a proposal classifier for multiple object tracking. In CVPR, 2021. 3

    2021

  6. [14]

    Resource-efficient multiview percep- tion: Integrating semantic masking with masked autoen- coders

    Kosta Dakic, Kanchana Thilakarathna, Rodrigo N Calheiros, and Teng Joon Lim. Resource-efficient multiview percep- tion: Integrating semantic masking with masked autoen- coders. arXiv:2410.04817, 2024. 2, 3, 7, 8

    2024

  7. [15]

    MOT20: A benchmark for multi object tracking in crowded scenes

    Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taix´e. MOT20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003,

    2003 arXiv

  8. [16]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 6

    2009

  9. [17]

    GIAOTracker: A compre- hensive framework for mcmot with global information and optimizing strategies in visdrone 2021

    Yunhao Du, Junfeng Wan, Yanyun Zhao, Binyu Zhang, Zhi- hang Tong, and Junhao Dong. GIAOTracker: A compre- hensive framework for mcmot with global information and optimizing strategies in visdrone 2021. In ICCV, 2021. 2, 3, 4

    2021

  10. [18]

    StrongSORT: Make deepsort great again

    Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. StrongSORT: Make deepsort great again. IEEE Transactions on Multimedia ,

  11. [19]

    Multi-object detection and track- ing (MODT) machine learning model for real-time video surveillance systems

    Mohamed Elhoseny. Multi-object detection and track- ing (MODT) machine learning model for real-time video surveillance systems. CSSP, 2020. 1

    2020

  12. [20]

    Multi-view tracking using weakly supervised human motion prediction

    Martin Engilberge, Weizhe Liu, and Pascal Fua. Multi-view tracking using weakly supervised human motion prediction. In WACV, 2023. 1, 2, 3, 7, 4, 5

    2023

  13. [21]

    Two-level data augmentation for calibrated multi-view de- tection

    Martin Engilberge, Haixin Shi, Zhiye Wang, and Pascal Fua. Two-level data augmentation for calibrated multi-view de- tection. In WACV, 2023. 3

    2023

  14. [22]

    Homography based multiple camera detection and tracking of people in a dense crowd

    Ran Eshel and Yael Moses. Homography based multiple camera detection and tracking of people in a dense crowd. In CVPR, 2008. 1

    2008

  15. [23]

    Unveiling the power of self-supervision for multi-view multi-human association and tracking

    Wei Feng, Feifan Wang, Ruize Han, Yiyang Gan, Zekun Qian, Junhui Hou, and Song Wang. Unveiling the power of self-supervision for multi-view multi-human association and tracking. TPAMI, 2024. 1, 3

    2024

  16. [24]

    PETS2009: Dataset and challenge

    James Ferryman and Ali Shahrokni. PETS2009: Dataset and challenge. In PETS, 2009. 1

    2009

  17. [25]

    Multicamera people tracking with a probabilistic occupancy map

    Francois Fleuret, Jerome Berclaz, Richard Lengagne, and Pascal Fua. Multicamera people tracking with a probabilistic occupancy map. TPAMI, 2007. 1, 3

    2007

  18. [26]

    Multi-object tracking as attention mechanism

    Hiroshi Fukui, Taiki Miyagawa, and Yusuke Morishita. Multi-object tracking as attention mechanism. InICIP, 2023. 5

    2023

  19. [27]

    Multiple human association and tracking from ego- centric and complementary top views

    Ruize Han, Wei Feng, Yujun Zhang, Jiewen Zhao, and Song Wang. Multiple human association and tracking from ego- centric and complementary top views. TPAMI, 2021. 1

    2021

  20. [28]

    Benchmarking the complementary- view multi-human association and tracking

    Ruize Han, Wei Feng, Feifan Wang, Zekun Qian, Haomin Yan, and Song Wang. Benchmarking the complementary- view multi-human association and tracking. IJCV, 2024. 1

    2024

  21. [29]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR,

  22. [30]

    Hyper- graphs for joint multi-view reconstruction and multi-object tracking

    Martin Hofmann, Daniel Wolf, and Gerhard Rigoll. Hyper- graphs for joint multi-view reconstruction and multi-object tracking. In CVPR, 2013. 1

    2013

  23. [31]

    Multiview detection with shadow transformer (and view-coherent data augmentation)

    Yunzhong Hou and Liang Zheng. Multiview detection with shadow transformer (and view-coherent data augmentation). In ACMMM, 2021. 3, 2

    2021

  24. [32]

    Multiview detection with feature perspective transformation

    Yunzhong Hou, Liang Zheng, and Stephen Gould. Multiview detection with feature perspective transformation. In ECCV,

  25. [33]

    Principal axis-based correspon- dence between multiple cameras for people tracking.TPAMI,

    Weiming Hu, Min Hu, Xue Zhou, Tieniu Tan, Jianguang Lou, and Steve Maybank. Principal axis-based correspon- dence between multiple cameras for people tracking.TPAMI,

  26. [34]

    Multiple target tracking by learning-based hierarchical association of detection responses

    Chang Huang, Yuan Li, and Ramakant Nevatia. Multiple target tracking by learning-based hierarchical association of detection responses. TPAMI, 2012. 3

    2012

  27. [35]

    Hao Jiang, Sidney Fels, and James J. Little. A linear pro- gramming approach for multiple object tracking. In CVPR,

  28. [36]

    A new approach to linear filtering and prediction problems

    Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. J Fluids Eng, 1960. 2, 5, 8

    1960

  29. [37]

    Framework for performance evaluation of face, text, and vehicle detection and tracking in video: Data, metrics, and protocol

    Rangachar Kasturi, Dmitry Goldgof, Padmanabhan Soundararajan, Vasant Manohar, John Garofolo, Rachel Bowers, Matthew Boonstra, Valentina Korzhova, and Jing Zhang. Framework for performance evaluation of face, text, and vehicle detection and tracking in video: Data, metrics, and...

    2008

  30. [38]

    Multi-task learning using uncertainty to weigh losses for scene geome- try and semantics

    Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geome- try and semantics. In CVPR, 2018. 6

    2018

  31. [39]

    The hungarian method for the assignment problem

    Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 4

    1955

  32. [40]

    Multi- view target transformation for pedestrian detection

    Wei-Yu Lee, Ljubomir Jovanov, and Wilfried Philips. Multi- view target transformation for pedestrian detection. In WACVW, 2023. 3

    2023

  33. [41]

    Cou- pled detection and trajectory estimation for multi-object tracking

    Bastian Leibe, Konrad Schindler, and Luc Van Gool. Cou- pled detection and trajectory estimation for multi-object tracking. In ICCV, 2007. 3

    2007

  34. [42]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regu- larization. In ICLR, 2018. 7, 2

    2018

  35. [43]

    HOTA: A higher order metric for evaluating multi-object tracking

    Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taix´e, and Bastian Leibe. HOTA: A higher order metric for evaluating multi-object tracking. IJCV, 2021. 4

    2021

  36. [44]

    TrackFormer: Multi-object track- ing with transformers

    Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. TrackFormer: Multi-object track- ing with transformers. In CVPR, 2022. 5

    2022

  37. [45]

    MOT16: A benchmark for multi-object tracking

    Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. MOT16: A benchmark for multi-object tracking. arXiv:1603.00831, 2016. 1, 3

    2016 arXiv

  38. [46]

    LMGP: Lifted multicut meets geometry projections for multi-camera multi-object tracking

    Duy MH Nguyen, Roberto Henschel, Bodo Rosenhahn, Daniel Sonntag, and Paul Swoboda. LMGP: Lifted multicut meets geometry projections for multi-camera multi-object tracking. In CVPR, 2022. 1

    2022

  39. [47]

    MCTR: Multi camera tracking transformer

    Alexandru Niculescu-Mizil, Deep Patel, and Iain Melvin. MCTR: Multi camera tracking transformer. arXiv:2408.13243, 2024. 3

    2024

  40. [48]

    A bayesian filter for multi-view 3D multi- object tracking with occlusion handling

    Jonah Ong, Ba-Tuong V o, Ba-Ngu V o, Du Yong Kim, and Sven Nordholm. A bayesian filter for multi-view 3D multi- object tracking with occlusion handling. TPAMI, 2020. 1

    2020

  41. [49]

    Multi-object tracking through simultaneous long occlusions and split-merge conditions

    AG Amitha Perera, Chukka Srinivas, Anthony Hoogs, Glen Brooksby, and Wensheng Hu. Multi-object tracking through simultaneous long occlusions and split-merge conditions. In CVPR, 2006. 3

    2006

  42. [50]

    From a bird’s eye view to see: Joint camera and subject registration without the camera calibration

    Zekun Qian, Ruize Han, Wei Feng, and Song Wang. From a bird’s eye view to see: Joint camera and subject registration without the camera calibration. In CVPR, 2024. 1

    2024

  43. [51]

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

    Zheng Qin, Sanping Zhou, Le Wang, Jinghai Duan, Gang Hua, and Wei Tang. MotionTrack: Learning robust short- term and long-term motions for multi-object tracking. In CVPR, 2023. 2, 3

    2023

  44. [52]

    3D random occlusion and multi-layer projection for deep multi-camera pedestrian localization

    Rui Qiu, Ming Xu, Yuyao Yan, Jeremy S Smith, and Xi Yang. 3D random occlusion and multi-layer projection for deep multi-camera pedestrian localization. In ECCV, 2022. 3

    2022

  45. [53]

    DyGLIP: A dynamic graph model with link prediction for accurate multi-camera multiple object tracking

    Kha Gia Quach, Pha Nguyen, Huu Le, Thanh-Dat Truong, Chi Nhan Duong, Minh-Triet Tran, and Khoa Luu. DyGLIP: A dynamic graph model with link prediction for accurate multi-camera multiple object tracking. In CVPR, 2021. 1, 3

    2021

  46. [54]

    Faster R-CNN: Towards real-time object detection with re- gion proposal networks

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with re- gion proposal networks. In NIPS, 2015. 1, 3

    2015

  47. [55]

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

    Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV, 2016. 6

    2016

  48. [56]

    Conditional random fields for multi-camera object de- tection

    Gemma Roig, Xavier Boix, Horesh Ben Shitrit, and Pascal Fua. Conditional random fields for multi-camera object de- tection. In ICCV, 2011. 3

    2011

  49. [57]

    U- Net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- Net: Convolutional networks for biomedical image segmen- tation. In MICCAI, 2015. 6

    2015

  50. [58]

    Stacked homography transformations for multi-view pedestrian detection

    Liangchen Song, Jialian Wu, Ming Yang, Qian Zhang, Yuan Li, and Junsong Yuan. Stacked homography transformations for multi-view pedestrian detection. In ICCV, 2021. 3

    2021

  51. [59]

    OCMCTrack: Online multi-target multi- camera tracking with corrective matching cascade

    Andreas Specker. OCMCTrack: Online multi-target multi- camera tracking with corrective matching cascade. In CVPRW, 2024. 1

    2024

  52. [60]

    DanceTrack: Multi-object track- ing in uniform appearance and diverse motion

    Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. DanceTrack: Multi-object track- ing in uniform appearance and diverse motion. In CVPR,

  53. [61]

    Scene generalized multi-view pedestrian detection with rotation- based augmentation and regularization

    Satoshi Suzuki, Shotaro Tora, and Ryo Masumura. Scene generalized multi-view pedestrian detection with rotation- based augmentation and regularization. In ICIP, 2024. 3

    2024

  54. [62]

    EarlyBird: Early-fusion for multi- view tracking in the bird’s eye view

    Torben Teepe, Philipp Wolters, Johannes Gilg, Fabian Her- zog, and Gerhard Rigoll. EarlyBird: Early-fusion for multi- view tracking in the bird’s eye view. In WACVW, 2024. 2, 3, 5, 6, 7, 8, 4

    2024

  55. [63]

    Lifting multi-view detection and tracking to the bird’s eye view

    Torben Teepe, Philipp Wolters, Johannes Gilg, Fabian Her- zog, and Gerhard Rigoll. Lifting multi-view detection and tracking to the bird’s eye view. In CVPRW, 2024. 2, 3, 6, 7, 4, 5, 8, 9

    2024

  56. [64]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need. In NIPS, 2017. 2, 3, 5, 7

    2017

  57. [65]

    Bringing generalization to deep multi-view pedestrian detection

    Jeet V ora, Swetanjal Dutta, Kanishk Jain, Shyamgopal Karthik, and Vineet Gandhi. Bringing generalization to deep multi-view pedestrian detection. In WACV, 2023. 2, 5, 6, 1, 3

    2023

  58. [66]

    BEV- SUSHI: Multi-target multi-camera 3d detection and tracking in bird’s-eye view

    Yizhou Wang, Tim Meinhardt, Orcun Cetintas, Cheng-Yen Yang, Sameer Satish Pusegaonkar, Benjamin Missaoui, Su- jit Biswas, Zheng Tang, and Laura Leal-Taix ´e. BEV- SUSHI: Multi-target multi-camera 3d detection and tracking in bird’s-eye view. arXiv:2412.00692, 2024. 1, 3, 7

    2024 arXiv

  59. [67]

    Towards real-time multi-object tracking

    Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In ECCV, 2020. 2, 3, 4

    2020

  60. [68]

    A multi modal people tracker for real time human robot interaction

    Tim Wengefeld, Steffen M ¨uller, Benjamin Lewandowski, and Horst-Michael Gross. A multi modal people tracker for real time human robot interaction. In RO-MAN, 2019. 1

    2019

  61. [69]

    Multi-view people tracking via hierarchical trajectory com- position

    Yuanlu Xu, Xiaobai Liu, Yang Liu, and Song-Chun Zhu. Multi-view people tracking via hierarchical trajectory com- position. In CVPR, 2016. 1

    2016

  62. [70]

    TransCenter: Transformers with dense representations for multiple-object tracking

    Yihong Xu, Yutong Ban, Guillaume Delorme, Chuang Gan, Daniela Rus, and Xavier Alameda-Pineda. TransCenter: Transformers with dense representations for multiple-object tracking. TPAMI, 2022. 5

    2022

  63. [71]

    An end-to-end tracking framework via multi- view and temporal feature aggregation

    Yihan Yang, Ming Xu, Jason F Ralph, Yuchen Ling, and Xi- aonan Pan. An end-to-end tracking framework via multi- view and temporal feature aggregation. CVIU, 2024. 2, 3, 6, 7, 4, 5

    2024

  64. [72]

    Real-time 3D deep multi- camera tracking

    Quanzeng You and Hao Jiang. Real-time 3D deep multi- camera tracking. arxiv:2003.11753, 2020. 1

    2003 arXiv

  65. [73]

    Towards dis- criminative representation: Multi-view trajectory contrastive learning for online multi-object tracking

    En Yu, Zhuoling Li, and Shoudong Han. Towards dis- criminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In CVPR, 2022. 2, 3, 4

    2022

  66. [74]

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

    Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi- angyu Zhang, and Yichen Wei. MOTR: End-to-end multiple- object tracking with transformer. In ECCV, 2022. 5

    2022

  67. [75]

    Global data as- sociation for multi-object tracking using network flows

    Li Zhang, Yuan Li, and Ramakant Nevatia. Global data as- sociation for multi-object tracking using network flows. In CVPR, 2008. 3

    2008

  68. [76]

    Cross-view cross- scene multi-view crowd counting

    Qi Zhang, Wei Lin, and Antoni B Chan. Cross-view cross- scene multi-view crowd counting. In CVPR, 2021. 2

    2021

  69. [77]

    Mahalanobis distance-based multi-view optimal transport for multi-view crowd localization

    Qi Zhang, Kaiyi Zhang, Antoni B Chan, and Hui Huang. Mahalanobis distance-based multi-view optimal transport for multi-view crowd localization. In ECCV, 2024. 3

    2024

  70. [78]

    FairMOT: On the fairness of detection and re-identification in multiple object tracking

    Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. FairMOT: On the fairness of detection and re-identification in multiple object tracking. IJCV, 2021. 6

    2021

  71. [79]

    Ob- jects as points

    Xingyi Zhou, Dequan Wang, and Philipp Kr ¨ahenb¨uhl. Ob- jects as points. arXiv:1904.07850, 2019. 1, 3, 5

    1904 arXiv

  72. [80]

    Global tracking transformers

    Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Kr¨ahenb¨uhl. Global tracking transformers. In CVPR, 2022. 2, 3, 5

    2022

  73. [81]

    Online multi-object tracking with dual matching attention networks

    Ji Zhu, Hua Yang, Nian Liu, Minyoung Kim, Wenjun Zhang, and Ming-Hsuan Yang. Online multi-object tracking with dual matching attention networks. In ECCV, 2018. 5 MVTrajecter: Multi-View Pedestrian Tracking with Trajectory Motion Cost and Trajectory Appearance Cost Supplementar...

    2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.