Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

CAMELTrack: Context-Aware Multi-cue ExpLoitation for Online Multi-Object Tracking

T0 review · 4 major / 7 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read One trainable cue-agnostic transformer replaces the hand-crafted association rules of tracking-by-detection, with state-of-the-art results on five benchmarks.

desk verdict A clean, well-ablated learned-association paper for TbD tracking; the ACT transfer assumption is undertested and the MOT17 claim is slightly overclaimed, but the central contribution holds up. read the letter →

arxiv 2505.01257 v1 pith:7YWW6FJT submitted 2025-05-02 cs.CV cs.LG

Vladimir Somers , Baptiste Standaert , Victor Joos , Alexandre Alahi , Christophe De Vleeschouwer This is my paper · ORCID
classification cs.CVcs.LG
keywords multi-objecttrackingtracking-by-detectionlearnedassociationtransformerfeaturefusionre-identificationposeestimationassociation-centrictraining
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

Online multi-object tracking is dominated by tracking-by-detection pipelines in which the three steps of association — summarizing a tracklet's history, blending motion and appearance cues, and staging the matches — are engineered by hand. This paper asserts that all three can be replaced by one compact trainable module, CAMEL, composed of two transformer encoders, while keeping the modularity that lets a tracker use off-the-shelf detectors, re-identification models, and pose estimators. It further asserts that its association-centric training, which learns from image-free synthetic scenarios assembled from pre-computed cues rather than from whole videos, transfers to real benchmarks and completes in under an hour on a single GPU. An oracle study supports the premise: with cues and detections held fixed, an optimal association rule improves HOTA by about 15.5 percent on DanceTrack and 8.3 percent on SportsMOT over heuristic baselines, so the headroom the paper targets is real. If these claims hold, learned association becomes a cheap, modular alternative to both heuristic SORT-style trackers and costly end-to-end DETR trackers.

What carries the argument

The load-bearing machinery is a two-stage transformer stack plus an image-free training scheme. First, one Temporal Encoder (TE) per cue type runs intra-object self-attention over the feature bank — the W most recent detections of a tracklet — after adding sinusoidal positional encodings of each detection's age, and returns a learned [CLS] token as the tracklet-level representation; this replaces the mean/EMA/Kalman heuristics for summarizing history. Second, the Group-Aware Feature Fusion Encoder (GAFFE) linearly projects each cue token, sums the projections into one token per active object, then applies inter-object self-attention across all tracklets and detections at once, producing disentangled embeddings in which same-identity pairs are close and different-identity pairs are far; this replaces both the hand-tuned fusion weights and the staged matching cascade. The training scheme makes the design practical: a pre-generated, image-free dataset of detections, labels, and cues is sampled into synthetic association scenarios (a random frame's detections paired with tracklets from earlier frames, sometimes stitched from different videos), augmented with identity swaps, detection dropout, and noise, and optimized with the InfoNCE loss, so the module learns how much to trust each cue in each context from data rather than from rules.

What would settle it

Run CAMELTrack without retraining on a held-out set of sequences whose detections come from a detector of very different quality than the YOLOX ones used in training (for instance, a much weaker detector with frequent missed detections), giving both CAMELTrack and the strongest heuristic-fusion baseline the same motion, appearance, and pose cues; if CAMELTrack fails to match or beat that baseline on association accuracy (AssA) and identity F1 (IDF1), the context-aware strategy learned from synthetic scenarios did not transfer to a real distribution shift.

Watch

Extended reading notes

Core claim

The central claim is that CAMEL is, to the authors' knowledge, the first fully-learned and cue-agnostic association module for online tracking-by-detection: it consumes any number of tracking cues (bounding boxes, re-identification embeddings, pose keypoints) and replaces the three classic heuristics — tracklet representation, feature fusion, and multi-stage matching — with a single trainable forward pass that outputs one disentangled embedding per tracklet and detection, matched by Euclidean distance and the Hungarian algorithm. Two transformer components do the work: a Temporal Encoder per cue aggregates each tracklet's recent detections into a tracklet-level token, and the Group-Aware Feature Fusion Encoder (GAFFE) projects every cue into a shared space, sums the projections per object, and runs inter-object self-attention to separate identities contextually. Training is association-centric: each sample is built from a random frame plus that frame's detections and tracklets from preceding frames, drawn across videos, with augmentations that simulate identity switches, dropped detections, and noisy cues, optimized by the InfoNCE contrastive loss. The resulting CAMELTrack pipeline, a compact 42.6-million-parameter model, reports state-of-the-art HOTA on SportsMOT (80.4), gains of roughly 7.6 HOTA points on PoseTrack21 and 3.7 on BEE24, surpasses all end-to-end methods on MOT17, and, with keypoints added, surpasses all end-to-end methods on DanceTrack (69.3 HOTA) — while training in one hour on one GPU and running at 13 frames per second.

Load-bearing premise

The module is trained only on image-free synthetic association scenarios — a random frame's detections paired with tracklets from earlier frames, sometimes stitched from different videos — so the reported gains rest on the assumption that these artificial contexts faithfully represent real occlusions, scene re-entries, and detector noise well enough that the learned fusion transfers to genuine online tracking.

Editorial extensions

If this is right

  • Adding a new tracking cue becomes a matter of training one Temporal Encoder and plugging it into CAMEL, with no redesign of the matching logic; the paper demonstrates this with pose keypoints, which add 3.2 HOTA on DanceTrack.
  • Learned association no longer requires the multi-day, multi-GPU budgets of end-to-end DETR trackers: CAMEL trains in under an hour on a single consumer GPU while outperforming those methods on several benchmarks.
  • The oracle study implies that current SORT-based trackers fail to exploit the cues they already have, so replacing association heuristics offers more headroom than improving detectors or cues within the tracking-by-detection paradigm.
  • Long-term identity recovery after occlusions and scene re-entries improves, because the feature bank and long training windows let the tracker re-associate objects that heuristic trackers would re-initialize as new identities.
  • The same architecture and training recipe transfer across domains with different cue reliability: position-only cues on BEE24, pose-augmented cues on PoseTrack21, and the full motion-plus-appearance-plus-pose stack on DanceTrack and SportsMOT.

Reading between the lines

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

  • The paper reports that cue dropout had no measurable effect during training, which suggests the model is not actually learning a fallback behavior for absent cues; a structured dropout that deletes cues exactly when occlusions or detector failures occur might close that robustness gap, but that is a testable extension, not a claim of the paper.
  • Because every cue enters through its own Temporal Encoder before summation, the same trained pipeline could plausibly ingest qualitatively different modalities — audio, radar, text descriptions, or vehicle license plates — but the paper mentions plates only as a possible extension, not a demonstrated one.
  • The gap between the paper's Feature Fusion Oracle and Association Oracle on DanceTrack implies that even a perfect fusion of the current cues is not enough there, so future gains on that benchmark likely come from designing stronger cues rather than from better fusion architectures.
  • The disentangled per-object embeddings CAMEL produces are, in effect, learned tracklet-level descriptors, so they could be reused for downstream tasks such as tracklet re-identification or clustering; the paper does not explore these uses.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes CAMEL, a transformer-based association module for online multi-object tracking that replaces three classic heuristics (tracklet representation, feature fusion, and multi-stage matching) with a single trainable architecture. CAMELTrack, the full tracking-by-detection pipeline, extracts multiple cues (bounding box, ReID appearance, pose keypoints) via off-the-shelf models, encodes each cue temporally, then fuses them in a group-aware self-attention encoder to produce a shared embedding space for Hungarian matching. A novel Association-Centric Training (ACT) scheme trains the module on pre-computed, image-free scenarios that mix detections and tracklets across frames and videos, with identity-swap, dropout, and perturbation augmentations. The paper reports state-of-the-art results on DanceTrack, SportsMOT, PoseTrack21, and BEE24, and competitive results on MOT17, with training under one hour on a single GPU and code/weights released.

Significance. If the results hold, this is a valuable step: it shows that learned association can match or beat hand-crafted association rules while preserving the modularity of tracking-by-detection, and the low training cost and cue-agnostic design make the approach broadly applicable. The paper ships released code, model weights, detections, and detailed configuration files, which is a strong positive for reproducibility. The ablation study in Table 5 is carefully designed and cleanly supports the role of the temporal encoders, group-aware fusion, and data augmentation. The oracle-based analysis provides an intuitive upper-bound framing.

major comments (4)
  1. [Section 1, contribution bullet 3, Table 4] The claim that CAMELTrack 'outperforms both E2E and SORT-based methods across five challenging benchmarks' is not supported by Table 4: on MOT17, the SORT-based ByteTrack and GHOST report HOTA 62.8, above CAMELTrack's 62.4. The abstract's 'state-of-the-art performance on multiple tracking benchmarks' should be qualified, e.g., by naming the benchmarks where SOTA is achieved, and the bullet should be corrected to avoid a false claim.
  2. [Section 3.3, Appendix E.2] The central transfer assumption of ACT is that synthetic scenarios built from random frames and cross-video mixing, with identity-swap and dropout augmentations, are representative of real online contexts. This is not tested directly. Please add a controlled experiment, such as training on one dataset (e.g., DanceTrack) and evaluating on another (e.g., SportsMOT or BEE24) without fine-tuning, or compare ACT-trained CAMEL against the same architecture trained with a recursive real-frame protocol. Without such a test, the generalization claim of ACT beyond the training distribution remains an assumption.
  3. [Section 4.4, Table 5] All reported results are single-run. The main gains (+3.2 HOTA on DanceTrack, +3.2 HOTA on SportsMOT) should be accompanied by mean and standard deviation over at least three random seeds, so the reader can judge whether the improvements are within noise.
  4. [Section 4.4, 'Complete Architecture and Training' paragraph] The statement that 'Exp. 10-11 achieve close performances' is not accurate: the gap is 4.7 HOTA on DanceTrack and 2.2 HOTA on SportsMOT. Please rephrase to 'narrow the gap' without overstating closeness.
minor comments (7)
  1. [Section 4.3] '+7.6% HOTA on PoseTrack21' should be '+7.7%' per Table 6 (66.0 vs 58.3).
  2. [Title and abstract] 'ExpLoitation' is likely a spelling error for 'Exploitation'; please confirm and correct throughout the manuscript.
  3. [Figure 2] Figure 2 is dense; consider splitting it into two figures (pipeline overview and model detail) to improve readability.
  4. [Appendix F.3] 'eEach' should be 'Each'.
  5. [Section 3.2] 'CAMEL build upon' should be 'CAMEL builds upon'.
  6. [Table 5] The delta rows in Table 5 are ambiguous; specify the baseline for each delta (e.g., relative to Exp. 1).
  7. [Appendix C.2] The statement about DiffMOT's interpolation is supported by a repository link, but the version/date of the release is not given; please add it for reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central claim is an empirical association module evaluated on held-out benchmarks; same-group components (BPBReID, TrackLab) serve as fixed off-the-shelf extractors and are not the claim being tested.

full rationale

The paper's central claim is that a trainable association module (CAMEL) can replace heuristic tracklet representation, feature fusion, and multi-stage matching, and that the resulting CAMELTrack pipeline reaches state-of-the-art HOTA on several benchmarks. This is an empirical claim evaluated on held-out test or validation sets (DanceTrack, SportsMOT, MOT17, PoseTrack21, BEE24), not a derivation from a fitted parameter. The main same-group components are BPBReID and TrackLab, but they are used as fixed off-the-shelf cue extractors and as an implementation framework; the headline numbers do not reduce to their design. The oracle experiments in Sec. 4.4 and Appendix F are explicitly ground-truth upper bounds: the Feature Fusion Oracle chooses its fusion weight with privileged access to annotations, and the Association Oracle matches detections to IoU-closest ground-truth tracks. These are labeled as oracles, so they are not presented as predictions derived from the model, and they cannot constitute circularity. The Association-Centric Training scheme builds synthetic association scenarios by mixing frames and videos, but whether these scenarios transfer to real tracking is a generalization or correctness risk, not a circularity: the model is still evaluated on real test data with no test-time fitting. Appendix D even describes a deliberate split to avoid using ReID embeddings that were overfit to the training data, which mitigates rather than creates circularity. The only self-referential elements are the citations to the authors' BPBReID and TrackLab; these are load-bearing only as infrastructure and cue extractors, not as the argument for CAMEL's learned association being effective. Consequently, no step in the paper's derivation chain reduces to its own inputs by construction, and the appropriate finding is minor self-citation without circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central contribution is an empirical method, not a derivation, so the ledger lists the hand-chosen architectural and life-cycle parameters and the domain assumptions the benchmark gains depend on. No new physical entities are introduced.

free parameters (7)
  • Feature bank window W = 50
    Length of the per-tracklet detection history used by the Temporal Encoders; set in Sec. 4.2 and affects how much temporal context the model can exploit.
  • Training sample pair count P = 32
    Number of tracklet-detection pairs per synthetic training sample, set in Sec. 4.2.
  • Detection confidence threshold per dataset = 0.4 (DanceTrack), 0.1 (SportsMOT), 0.3 (PoseTrack21), 0.5 (MOT17)
    Life-cycle filtering thresholds in Appendix D that control which detections enter association.
  • Tracklet initialization confidence threshold per dataset = 0.9, 0.4, 0.4, 0.55
    Minimum confidence for creating new tracklets, tuned per dataset in Appendix D.
  • CAMEL matching threshold per dataset = 0.1, 0.1, 0.45, 0.5
    Euclidean distance threshold in the learned embedding space below which a tracklet-detection pair is matched, tuned per dataset in Appendix D.
  • Minimum hits for tracklet confirmation = 0 except MOT17=1
    Number of hits required before confirming a tracklet, set per dataset in Appendix D to filter detector noise.
  • Data augmentation parameters = not reported (grid-searched)
    Optimal augmentation parameters are selected by grid search on each dataset's validation set, including dropout probability p_drop and noise scales, as stated in Appendix E.3.
assumptions (5)
  • standard math Transformer self-attention and InfoNCE contrastive learning produce discriminative embeddings for the association task.
    The method assumes the standard machinery of transformer encoders and contrastive loss works as intended; no proof is given.
  • domain assumption The IoU-closest ground-truth assignment used to label detections is a sufficiently accurate supervision signal.
    Appendix E.1 assigns each detection the identity of its IoU-closest ground truth; if detections frequently straddle multiple identities, the learned embeddings inherit label noise.
  • domain assumption Off-the-shelf cue extractors such as YOLOX, BPBReID and RTMPose provide informative, complementary cues in the target domains.
    The pipeline depends on external models for appearance, motion and pose cues; on DanceTrack the reported ReID mAP is only 34.4, so the method must lean on other cues without a guarantee that they exist in new domains.
  • ad hoc to paper Synthetic association scenarios built from random frames across videos are representative of real online tracking contexts.
    Sec. 3.3 and Appendix E.2 construct training samples by mixing detections and tracklets, sometimes from distinct videos; the paper does not quantify the domain gap between these artificial contexts and real test sequences.
  • domain assumption The validation splits and grid search over augmentation and life-cycle hyperparameters do not overfit the benchmark choices.
    Augmentation parameters and life-cycle thresholds are grid-searched on each dataset's validation set (Appendix E.3); on small validation sets this selection can be an optimistic bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CAMELTrack: Context-Aware Multi-cue ExpLoitation for Online Multi-Object Tracking." pith.science (2026). https://pith.science/paper/7YWW6FJT

@misc{pith2026250501257,
  author       = {Pith},
  title        = {Pith review of: CAMELTrack: Context-Aware Multi-cue ExpLoitation for Online Multi-Object Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7YWW6FJT}},
  note         = {Machine review of arXiv:2505.01257}
}
read the original abstract

Online multi-object tracking has been recently dominated by tracking-by-detection (TbD) methods, where recent advances rely on increasingly sophisticated heuristics for tracklet representation, feature fusion, and multi-stage matching. The key strength of TbD lies in its modular design, enabling the integration of specialized off-the-shelf models like motion predictors and re-identification. However, the extensive usage of human-crafted rules for temporal associations makes these methods inherently limited in their ability to capture the complex interplay between various tracking cues. In this work, we introduce CAMEL, a novel association module for Context-Aware Multi-Cue ExpLoitation, that learns resilient association strategies directly from data, breaking free from hand-crafted heuristics while maintaining TbD's valuable modularity. At its core, CAMEL employs two transformer-based modules and relies on a novel association-centric training scheme to effectively model the complex interactions between tracked targets and their various association cues. Unlike end-to-end detection-by-tracking approaches, our method remains lightweight and fast to train while being able to leverage external off-the-shelf models. Our proposed online tracking pipeline, CAMELTrack, achieves state-of-the-art performance on multiple tracking benchmarks. Our code is available at https://github.com/TrackingLaboratory/CAMELTrack.

Figures

Figures reproduced from arXiv: 2505.01257 by the authors.

Figure 1
Figure 1. Our proposed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture overview of CAMELTrack, our online tracking-by-detection pipeline that operates in three steps: (i) object detec [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of tracking results on v 00HRwkvvjtQ c007 from SportsMOT. Ground truth tracks are depicted with horizontal lines, while colors indicates predicted identities. Blue zones highlight scene re-entries and red zones show occlusions. Frames where the ground truth identity has left the scene are represented with a black line, and missing predictions are left blank. Bottom gray plots show cumulative tracked id… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Similarity distributions between positive and negative tracklet-detection pairs. (a) IoU between KF’s predictions and de￾tections. Cosine distances between : (b) EMA tracklet and detec￾tion ReID embeddings (c) pairs of CAMEL’s output embeddings. (a) Motion (b) Appearan…
Figure 5
Figure 5. Figure 5: t-SNE on the first 150 frames of dancetrack0019. Each identity is assigned a unique color, with light/dark shades indicat￾ing detections/tracklets respectively. Analysis of Similarity Distributions [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Visualization of tracking results on additional videos from the SportsMOT and DanceTrack validation sets. (a-c) video [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Learning Association via Track-Detection Matching for Multi-Object Tracking

    cs.CV 2025-12 conditional novelty 6.0 of 10

    TDLP uses a link-prediction head to match tracks to detections, beating heuristic and metric-learning trackers on several MOT benchmarks while underperforming on MOT17.

  2. Zero-Shot Multi-Animal Tracking in the Wild

    cs.CV 2025-11 conditional novelty 5.0 of 10

    A zero-shot multi-animal tracker combining Grounding DINO + SAM 2 with three hand-designed heuristics beats prior methods on four animal-tracking benchmarks with fixed hyperparameters.

Reference graph

Works this paper leans on

98 extracted references · 63 canonical work pages · cited by 2 Pith papers

  1. [1]

    Engineering an efficient object tracker for non-linear motion

    Momir Ad ˇzemovi´c, Predrag Tadi ´c, Andrija Petrovi ´c, and Mladen Nikoli´c. Engineering an efficient object tracker for non-linear motion. arXiv, abs/2407.00738, 2024. 2, 6

    arXiv 2024

  2. [2]

    Beyond Kalman Filters: Deep Learning-Based Filters for Improved Object Tracking

    Momir Ad ˇzemovi´c, Predrag Tadi ´c, Andrija Petrovi ´c, and Mladen Nikoli ´c. Beyond kalman filters: Deep learning- based filters for improved object tracking. arXiv, abs/2402.09865, 2024. 2, 6

    work page Pith review arXiv 2024

  3. [3]

    Bot- SORT: Robust associations multi-pedestrian tracking.arXiv, abs/2206.14651, 2022

    Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Bot- SORT: Robust associations multi-pedestrian tracking.arXiv, abs/2206.14651, 2022. 1, 2, 3, 4, 6

    arXiv 2022

  4. [4]

    Tracking without bells and whistles

    Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pages 941–951. Inst. Electr. Electron. Eng. (IEEE), 2019. 2, 3, 6

    2019

  5. [5]

    Simple online and realtime tracking

    Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In IEEE Int. Conf. Image Process. (ICIP) , pages 3464–3468, Phoenix, AZ, USA, 2016. Inst. Electr. Electron. Eng. (IEEE). 1, 2, 6

    2016

  6. [6]

    Learning a neural solver for multiple object tracking

    Guillem Braso and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. InIEEE/CVF Conf. Com- put. Vis. Pattern Recognit. (CVPR), pages 6246–6256, Seat- tle, W A, USA, 2020. Inst. Electr. Electron. Eng. (IEEE). 3, 6

    2020

  7. [7]

    MeMOT: Multi-object tracking with memory

    Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. MeMOT: Multi-object tracking with memory. arXiv, abs/2203.16761, 2022. 3, 8

    arXiv 2022

  8. [8]

    Observation-centric SORT: Rethinking SORT for robust multi-object tracking

    Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirod- kar, and Kris Kitani. Observation-centric SORT: Rethinking SORT for robust multi-object tracking. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , pages 9686–9696, Vancouver, Can., 2023. Inst. Electr. Electron. Eng. (IEEE). 1, 2, 6, 3, 4

    2023

Show all 98 references
  1. [9]

    TOPIC: A parallel association paradigm for multi-object tracking under complex motions and diverse scenes

    Xiaoyan Cao, Yiyao Zheng, Yao Yao, Hua-Peng Qin, Xi- aoyu Cao, and Shihui Guo. TOPIC: A parallel association paradigm for multi-object tracking under complex motions and diverse scenes. IEEE Trans. Image Process. , 34:743– 758, 2025. 5, 2

    2025

  2. [10]

    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 Eur. Conf. Comput. Vis. (ECCV), pages 213–229. Springer Nat. Switz.,

  3. [11]

    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 IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 22877–22887, Vancouver, Can., 2023. Inst. Electr. Electron. Eng. (IEEE). 3, 6

    2023

  4. [12]

    MMDetection: Open MMLab detection toolbox and benchmark

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, and et al. MMDetection: Open MMLab detection toolbox and benchmark. arXiv, abs/1906.07155, 2019. 4

    1906 arXiv

  5. [13]

    Delv- ing into the trajectory long-tail distribution for muti-object tracking

    Sijia Chen, En Yu, Jinyang Li, and Wenbing Tao. Delv- ing into the trajectory long-tail distribution for muti-object tracking. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 19341–19351, Seattle, W A, USA, 2024. Inst. Electr. Electron. Eng. (IEEE). 3, 6

    2024

  6. [14]

    Beyond appearance: A semantic controllable self-supervised learning framework for human-centric visual tasks

    Weihua Chen, Xianzhe Xu, Jian Jia, Hao Luo, Yaohua Wang, Fan Wang, Rong Jin, and Xiuyu Sun. Beyond appearance: A semantic controllable self-supervised learning framework for human-centric visual tasks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 202...

    2023

  7. [15]

    TransMOT: Spatial-temporal graph trans- former for multiple object tracking

    Peng Chu, Jiang Wang, Quanzeng You, Haibin Ling, and Zicheng Liu. TransMOT: Spatial-temporal graph trans- former for multiple object tracking. In IEEE/CVF Win- ter Conf. Appl. Comput. Vis. (WACV) , pages 4859–4869, Waikoloa, HI, USA, 2023. Inst. Electr. Electron. Eng. (IEEE). ...

    2023

  8. [16]

    SoccerNet-tracking: Multiple object tracking dataset and benchmark in soccer videos

    Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. SoccerNet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Work. (CVPRW), pag...

    2022

  9. [17]

    Openmmlab pose estimation tool- box and benchmark

    MMPose Contributors. Openmmlab pose estimation tool- box and benchmark. https://github.com/open- mmlab/mmpose, 2020. 4

    2020

  10. [18]

    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 IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , pages 9887– 9897, Paris, Fr., 2023. Inst. Electr. Electron. Eng. (IEEE). 1, 5, 6, 4

    2023

  11. [19]

    Joint spatial-temporal and appearance model- ing with transformer for multiple object tracking

    Peng Dai, Yiqiang Feng, Renliang Weng, and Chang- shui Zhang. Joint spatial-temporal and appearance model- ing with transformer for multiple object tracking. CoRR, abs/2205.15495, 2022. 1

    2022 arXiv

  12. [20]

    BERT: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional trans- formers for language understanding. In Proceedings of the 9 2019 Conference of the North American Chapter of the As- sociation for Computational Linguistics: Human...

    2019

  13. [21]

    ADA-track: End-to-end multi-camera 3D multi- object tracking with alternating detection and association

    Shuxiao Ding, Lukas Schneider, Marius Cordts, and Juer- gen Gall. ADA-track: End-to-end multi-camera 3D multi- object tracking with alternating detection and association. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 15184–15194, Seattle, W A, USA, 2024. Inst....

    2024

  14. [22]

    A gated attention trans- former for multi-person pose tracking

    Andreas Doering and Juergen Gall. A gated attention trans- former for multi-person pose tracking. In IEEE/CVF Int. Conf. Comput. Vis. Work. (ICCV Work.), pages 3181–3190, Paris, Fr., 2023. Inst. Electr. Electron. Eng. (IEEE). 2, 4

    2023

  15. [23]

    PoseTrack21: A dataset for person search, multi-object tracking and multi-person pose tracking

    Andreas Doering, Di Chen, Shanshan Zhang, Bernt Schiele, and Juergen Gall. PoseTrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , pages 20931–20940, New Orleans, LA, USA, 2022. I...

    2022

  16. [24]

    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 Trans. Multimedia, 25:8725– 8737, 2023. 1, 2, 3, 6

    2023

  17. [25]

    SoccerNet 2022 challenges results

    Giancola et al. SoccerNet 2022 challenges results. In Pro- ceedings of the 5th International ACM Workshop on Multi- media Content Analysis in Sports, pages 75–86. ACM, 2022. 1

    2022

  18. [26]

    MeMOTR: Long-term memory-augmented transformer for multi-object tracking

    Ruopeng Gao and Limin Wang. MeMOTR: Long-term memory-augmented transformer for multi-object tracking. In IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , pages 9867– 9876, Paris, Fr., 2023. Inst. Electr. Electron. Eng. (IEEE). 3, 5, 6, 2, 4

    2023

  19. [27]

    Multiple ob- ject tracking as ID prediction

    Ruopeng Gao, Yijun Zhang, and Limin Wang. Multiple ob- ject tracking as ID prediction. arXiv, abs/2403.16848, 2024. 2, 3, 5, 6, 7, 4, 8

    2024 arXiv

  20. [28]

    YOLOX: Exceeding YOLO series in 2021

    Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: Exceeding YOLO series in 2021. arXiv, abs/2107.08430, 2021. 1, 5, 4

    2021 arXiv

  21. [29]

    ETTrack: Enhanced temporal motion predictor for multi-object track- ing

    Xudong Han, Nobuyuki Oishi, Yueying Tian, Elif Ucurum, Rupert Young, Chris Chatwin, and Philip Birch. ETTrack: Enhanced temporal motion predictor for multi-object track- ing. arXiv, abs/2405.15755, 2024. 2, 6

    2024 arXiv

  22. [30]

    FeatureSORT: Essential features for effective track- ing

    Hamidreza Hashempoor, Rosemary Koikara, and Yu Dong Hwang. FeatureSORT: Essential features for effective track- ing. arXiv, abs/2407.04249, 2024. 2, 6

    2024

  23. [31]

    Fastreid: A pytorch toolbox for general instance re-identification

    Lingxiao He, Xingyu Liao, Wu Liu, Xinchen Liu, Peng Cheng, and Tao Mei. Fastreid: A pytorch toolbox for general instance re-identification. arXiv preprint arXiv:2006.02631,

    2006 arXiv

  24. [32]

    Exploring learning- based motion models in multi-object tracking

    Hsiang-Wei Huang, Cheng-Yen Yang, Wenhao Chai, Zhongyu Jiang, and Jenq-Neng Hwang. Exploring learning- based motion models in multi-object tracking. arXiv, abs/2403.10826, 2024. 2, 6

    2024 arXiv

  25. [33]

    Iterative scale-up ExpansionIoU and deep features association for multi-object tracking in sports

    Hsiang-Wei Huang, Cheng-Yen Yang, Jiacheng Sun, Pyong- Kun Kim, Kwang-Ju Kim, Kyoungoh Lee, Chung-I Huang, and Jenq-Neng Hwang. Iterative scale-up ExpansionIoU and deep features association for multi-object tracking in sports. In IEEE/CVF Winter Conf. Appl. Comput. Vis. Work. ...

    2024

  26. [34]

    RTMPose: Real- time multi-person pose estimation based on MMPose.arXiv, abs/2303.07399, 2023

    Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. RTMPose: Real- time multi-person pose estimation based on MMPose.arXiv, abs/2303.07399, 2023. 5, 4

    2023 arXiv

  27. [35]

    TrackLab

    Victor Joos, Vladimir Somers, and Baptiste Stan- daert. TrackLab. https : / / github . com / TrackingLaboratory/tracklab, 2024. 5

    2024

  28. [36]

    Kasturi, D

    R. Kasturi, D. Goldgof, P. Soundararajan, V . Manohar, J. Garofolo, R. Bowers, M. Boonstra, V . Korzhova, and Jing Zhang. Framework for performance evaluation of face, text, and vehicle detection and tracking in video: Data, metrics, and protocol. IEEE Trans. Pattern Anal. Mac...

    2009

  29. [37]

    Deep- reid: Deep filter pairing neural network for person re- identification

    Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deep- reid: Deep filter pairing neural network for person re- identification. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 152–159. IEEE Computer Society,

    2014

  30. [38]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vi- sion, ICCV 2021, Montreal, QC, Canada, October 10-17, ...

    2021

  31. [39]

    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. Int. J. Comput. Vis., 129(2):548–578, 2020. 5

    2020

  32. [40]

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

    Weiyi Lv, Yuhang Huang, Ning Zhang, Ruei-Sung Lin, Mei Han, and Dan Zeng. DiffMOT: A real-time diffusion- based multiple object tracker with non-linear prediction. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , pages 19321–19330, Seattle, W A, USA, 2024. Inst. Elect...

    2024

  33. [41]

    Deep OC-sort: Multi-pedestrian tracking by adap- tive re-identification

    Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep OC-sort: Multi-pedestrian tracking by adap- tive re-identification. In IEEE Int. Conf. Image Process. (ICIP), pages 3025–3029, Kuala Lumpur, Malaysia, 2023. Inst. Electr. Electron. Eng. (IEEE). 1, 2, 6, 3, 4

    2023

  34. [42]

    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 IEEE/CVF Conf. Comput. Vis. Pat- tern Recognit. (CVPR), pages 8834–8844, New Orleans, LA, USA, 2022. Inst. Electr. Electron. Eng. (IEEE)...

    2022

  35. [43]

    Reid, Stefan Roth, and Konrad Schindler

    Anton Milan, Laura Leal-Taix ´e, Ian D. Reid, Stefan Roth, and Konrad Schindler. MOT16: A benchmark for multi- object tracking. CoRR, abs/1603.00831, 2016. 6, 4

    2016 arXiv

  36. [44]

    Repre- sentation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre- sentation learning with contrastive predictive coding. arXiv, abs/1807.03748, 2018. 5

    2018 arXiv

  37. [45]

    Transformer-based assignment decision net- 10 work for multiple object tracking

    Athena Psalta, Vasileios Tsironis, and Konstantinos Karantzalos. Transformer-based assignment decision net- 10 work for multiple object tracking. Comput. Vis. Image Un- derst., 241:103957, 2024. 3, 6, 7

    2024

  38. [46]

    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 IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , pages 17939–17948, Vancouver, Can., 2023. Inst. Electr. E...

    2023

  39. [47]

    Focus on details: Online multi-object tracking with diverse fine-grained representa- tion

    Hao Ren, Shoudong Han, Huilin Ding, Ziwen Zhang, Hong- wei Wang, and Faquan Wang. Focus on details: Online multi-object tracking with diverse fine-grained representa- tion. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 11289–11298, Vancouver, Can., 2023. Inst....

    2023

  40. [48]

    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 Computer Vision - ECCV Workshops, pages 17–35. Springer Int. Publ., 2016. 5

    2016

  41. [49]

    Simple cues lead to a strong multi-object tracker

    Jenny Seidenschwarz, Guillem Bras ´o, Victor Castro Serrano, Ismail Elezi, and Laura Leal-Taix ´e. Simple cues lead to a strong multi-object tracker. InIEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 13813–13823, Vancouver, Can., 2023. Inst. Electr. Electron. Eng. ...

    2023

  42. [50]

    Crowdhuman: A benchmark for detecting human in a crowd

    Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xi- angyu Zhang, and Jian Sun. Crowdhuman: A benchmark for detecting human in a crowd. CoRR, abs/1805.00123, 2018. 7, 3

    2018 arXiv

  43. [51]

    Body part-based representation learn- ing for occluded person re-identification

    Vladimir Somers, Christophe De Vleeschouwer, and Alexandre Alahi. Body part-based representation learn- ing for occluded person re-identification. In IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV) , pages 1613– 1623, Waikoloa, HI, USA, 2023. Inst. Electr. Electron. Eng. (IEEE...

    2023

  44. [52]

    Keypoint promptable re-identification

    Vladimir Somers, Alexandre Alahi, and Christophe De Vleeschouwer. Keypoint promptable re-identification. In Eur. Conf. Comput. Vis. (ECCV) , pages 216–233. Springer Nat. Switz., 2024. 1

    2024

  45. [53]

    Mansourian, Xin Zhou, Shohreh Kasaei, Bernard Ghanem, Alexandre Alahi, Marc Van Droogenbroeck, and Christophe De Vleeschouwer

    Vladimir Somers, Victor Joos, Anthony Cioppa, Silvio Gi- ancola, Seyed Abolfazl Ghasemzadeh, Floriane Magera, Baptiste Standaert, Amir M. Mansourian, Xin Zhou, Shohreh Kasaei, Bernard Ghanem, Alexandre Alahi, Marc Van Droogenbroeck, and Christophe De Vleeschouwer. Soc- cerNet ...

    2024

  46. [54]

    Deep high-resolution representation learning for human pose esti- mation

    Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose esti- mation. In CVPR, 2019. 4

    2019

  47. [55]

    TransTrack: Multiple object tracking with transformer

    Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. TransTrack: Multiple object tracking with transformer. arXiv, abs/2012.15460, 2020. 3, 6

    2012 arXiv

  48. [56]

    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. InIEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 20961– 20970, New Orleans, LA, USA, 2022. Inst. Electr...

    2022

  49. [57]

    Visualizing data using t-sne

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579–2605, 2008. 8

    2008

  50. [58]

    Brea, and Manuel Mucientes

    Lorenzo Vaquero, Yihong Xu, Xavier Alameda-Pineda, Vic- tor M. Brea, and Manuel Mucientes. Lost and found: Overcoming detector failures in online multi-object tracking. arXiv, abs/2407.10151, 2024. 3, 6

    2024 arXiv

  51. [59]

    SMILEtrack: SiM- Ilarity LEarning for occlusion-aware multiple object track- ing

    Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen, Ming- Ching Chang, Hung-Hin So, and Xin Li. SMILEtrack: SiM- Ilarity LEarning for occlusion-aware multiple object track- ing. In AAAI Conf. Artif. Intell., pages 5740–5748. Associa- tion for the Advancement of Artificial Intellige...

  52. [60]

    Towards real-time multi-object tracking

    Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. Computer Vision - ECCV 2020 , pages 107–122, 2020. 2, 6

    2020

  53. [61]

    Simple online and realtime tracking with a deep association metric

    Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In IEEE Int. Conf. Image Process. (ICIP), pages 3645–3649, Beijing, China, 2017. Inst. Electr. Electron. Eng. (IEEE). 1, 2, 6

    2017

  54. [62]

    MambaTrack: A simple baseline for multiple object tracking with state space model

    Changcheng Xiao, Qiong Cao, Zhigang Luo, and Long Lan. MambaTrack: A simple baseline for multiple object tracking with state space model. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4082–4091. ACM, 2024. 2, 6

    2024

  55. [63]

    Spatial- temporal relation networks for multi-object tracking

    Jiarui Xu, Yue Cao, Zheng Zhang, and Han Hu. Spatial- temporal relation networks for multi-object tracking. In IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , pages 3987–

  56. [64]

    Bridging the gap between end-to-end and non-end-to- end multi-object tracking

    Feng Yan, Weixin Luo, Yujie Zhong, Yiyang Gan, and Lin Ma. Bridging the gap between end-to-end and non-end-to- end multi-object tracking. arXiv, abs/2305.12724, 2023. 3, 6

    2023 arXiv

  57. [65]

    Hybrid-SORT: Weak cues matter for online multi-object tracking

    Mingzhan Yang, Guangxin Han, Bin Yan, Wenhua Zhang, Jinqing Qi, Huchuan Lu, and Dong Wang. Hybrid-SORT: Weak cues matter for online multi-object tracking. In AAAI Conf. Artif. Intell., pages 6504–6512. Association for the Ad- vancement of Artificial Intelligence (AAAI), 2024. ...

    2024

  58. [66]

    UTM: A unified multiple object tracking model with identity-aware feature enhancement

    Sisi You, Hantao Yao, Bing-kun Bao, and Changsheng Xu. UTM: A unified multiple object tracking model with identity-aware feature enhancement. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 21876– 21886, Vancouver, Can., 2023. Inst. Electr. Electron. Eng. (IEEE). 3, 6

    2023

  59. [67]

    BDD100K: A diverse driving dataset for heteroge- neous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingy- ing Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heteroge- neous multitask learning. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , pages 2633–2642, Seattle, W...

    2020

  60. [68]

    Hrformer: High- 11 resolution vision transformer for dense predict

    Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. Hrformer: High- 11 resolution vision transformer for dense predict. In Advances in Neural Information Processing Systems 34: Annual Con- ference on Neural Information Processing Systems 202...

    2021

  61. [69]

    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 Eur. Conf. Comput. Vis. (ECCV), pages 659–675, Cham, 2022. Springer Nat. Switz. 2, 3, 5, 6, 4, 8

    2022

  62. [70]

    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. Int. J. Comput. Vis., 129(11):3069–3087, 2021. 2, 6

    2021

  63. [71]

    ByteTrack: Multi-object tracking by associating ev- ery detection box

    Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. ByteTrack: Multi-object tracking by associating ev- ery detection box. In Eur. Conf. Comput. Vis. (ECCV), pages 1–21. Springer Nat. Switz., 2022. 1, 2, 6, 3, 4

    2022

  64. [72]

    MOTRv2: Bootstrapping end-to-end multi-object tracking by pre- trained object detectors

    Yuang Zhang, Tiancai Wang, and Xiangyu Zhang. MOTRv2: Bootstrapping end-to-end multi-object tracking by pre- trained object detectors. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 22056–22065, Vancouver, Can., 2023. Inst. Electr. Electron. Eng. (IEEE). 2, 3, 5, 6, 8

    2023

  65. [73]

    Scalable person re-identification: A benchmark

    Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jing- dong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7- 13, 2015, pages 1116–1124. IEEE Computer Society, 2015. 4

    2015

  66. [74]

    Omni-scale feature learning for person re- identification

    Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Omni-scale feature learning for person re- identification. In ICCV, 2019. 4

    2019

  67. [75]

    Tracking objects as points

    Xingyi Zhou, Vladlen Koltun, and Philipp Kr ¨ahenb¨uhl. Tracking objects as points. In Eur. Conf. Comput. Vis. (ECCV), pages 474–490. Springer, 2020. 6

    2020

  68. [76]

    Global tracking transformers

    Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Kra- henbuhl. Global tracking transformers. In IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , pages 8761–8770, New Orleans, LA, USA, 2022. Inst. Electr. Electron. Eng. (IEEE). 3, 6 12 CAMELTrack: Context-Aware Multi...

    2022

  69. [78]

    CAMEL follows a fundamentally different approach based on deep metric learning

    N×M Edge Tokens vs N+M Object Tokens: Trans- MOT and TransTAM adopt graph-inspired approaches using N×M edge tokens (one edge token for each tracklet-detection pair) and employ token-wise binary classifiers to predict an assocation score for each token and thereby approximate ...

  70. [79]

    Explicit Attention Bias: TransMOT and TransTAM both introduce an explicit spatial bias into the attention mechanism of their Spatial Graph Transformer, artifi- cially restricting communication between spatially ad- jacent detections having a non-null IoU. We found that such ex...

  71. [80]

    On the other hand, CAMEL first embeds each cue independently into a higher- dimensional space with an end-to-end learnable FFN be- fore summing them

    Cue Fusion: TransMOT naively concatenates a re- identification embedding with a vector of 4 scalars rep- resenting the bounding box. On the other hand, CAMEL first embeds each cue independently into a higher- dimensional space with an end-to-end learnable FFN be- fore summing ...

  72. [81]

    Association-centric Training (ACT):TransMOT trains on short time windows reducing their potential to solve long-term tracking, does not perform data augmentation on the tracklet-detection pairs, and do not build syn- thetic association training examples combining multiple vide...

  73. [82]

    Heuristic Dependency: First, TransMOT is not heuristic-free. Indeed, it relies on a hand-crafted multi- stage matching pipeline where their learned transformer module is only used in the second stage, while the first and third stages remain purely based on heuristics, with a b...

  74. [83]

    We perform Hungarian matching with IoU between the ground truth bounding boxes and a detector, in order to give each predicted bounding box a ground truth identity

  75. [84]

    Every resulting detection is then passed through every cue-specific model (reid, pose estimation)

  76. [85]

    This information is later required by some data augmentations

    We compute the bbox overlap between detections in the same frame. This information is later required by some data augmentations

  77. [86]

    All resulting information is saved on disk. E.2. Training Loop During training we sample from our pre-generated set to build training batches of B training samples. One sam- ple for training with P tracklet-detection pairs is created through the following steps

  78. [87]

    Selecting a random frame from a random video

  79. [88]

    Gathering all detections from that frame and all detec- tions from previous frames

  80. [89]

    Performing data augmentation on the tracklets and de- tections (see Appendix E.3)

  81. [90]

    Only keeping the W last detections per tracklet (W=50 in most experiments)

  82. [91]

    CAMEL then receives all the detections and tracklets for the samples in a batch, and outputs one embedding for each detection and each tracklet

    Repeating this procedure with a new frame until we ob- tain P tracklet-detection pairs. CAMEL then receives all the detections and tracklets for the samples in a batch, and outputs one embedding for each detection and each tracklet. The InfoNCE [44] loss is then computed using...

  83. [92]

    Current detections are matched to ground truth bounding boxes using the Hungarian algorithm

  84. [93]

    IoU score is used as the matching metric with a mini- mum threshold of 0.5

  85. [94]

    Feature Fusion Oracle (Exp

    eEach matched detection inherits the track identity of its corresponding ground truth. Feature Fusion Oracle (Exp. 11). This oracle demon- strates the potential of optimal feature fusion while high- lighting current limitations of heuristic-based association rules. For each in...

  86. [95]

    A single weight factor linearly combines appearance and motion costs into a unified cost matrix

  87. [96]

    The resulting cost matrix is processed by the Hungarian algorithm for final matching

  88. [97]

    Limitations and Future Extensions

    The optimal weight is determined by maximizing the association accuracy (percentage of correct tracklet- detection matches), thanks to privileged access to ground-truth annotations. Limitations and Future Extensions. While our simple im- plementation sufficiently illustrates t...

  89. [98]

    and threshold-based gating [3, 24]. GHOST [49] also conducts an extensive study examining how different ”Mo- tion Weight” values (weighting factors combining motion and appearance cost matrices) impact tracking performance across various datasets. (iii) Multi-stage Matching: t...

    2020

  90. [3997]

    Inst. Electr. Electron. Eng. (IEEE), 2019. 3, 6

    2019

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.