Pith. sign in

REVIEW 3 major objections 6 minor 48 references

This paper argues that the identity switches in multi-object tracking of highly similar objects come from a mismatch between instance-level re-ID training and the frame-level global assignment used at inference, and that fixing this by trai

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

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

Reformulating re-ID as frame-level trajectory-to-detection association with shared-appearance suppression improves identity preservation in dense, visually similar bee tracking (BEE24) by ~1 HOTA and 28% fewer ID switches.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Worth a careful look, but test-set selection and no error bars mean the reported gains are not yet established; the idea is sound and the ablation is honest. the 3 major comments →

arxiv 2607.17157 v1 pith:C4MLOXOB submitted 2026-07-19 cs.CV cs.AI

VLA-ReID: Video-Level Association for Re-Identification in Multi-Object Tracking with Highly Similar Objects

Yanrong Qin , Xiaoyan Cao , Yao Yao This is my paper
classification cs.CV cs.AI
keywords multi-object trackingre-identificationvideo-level associationdata associationappearance modelinghigh appearance similarityinsect tracking
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper's central claim is that re-identification models for multi-object tracking should be trained the way they are used: not by comparing one query patch against one or a few candidates, but by assigning every historical trajectory to all detections present in the current frame. The authors build VLA-ReID on top of a base tracker, changing only the appearance cost. They introduce a learned 'common appearance direction' estimated from all current-frame detections, subtract its projection from both trajectory and detection features through a per-dimension gate, and re-score the resulting similarity matrix relative to the current candidate set. On the BEE24 benchmark, this cut identity switches from 606 to 437 (28%) and improved HOTA, MOTA, AssA, AssR, and IDF1, all without extra annotations. The implication is that appearance-based association in dense, visually homogenous scenes is better treated as set-relative re-ranking than as independent pairwise matching.

Core claim

VLA-ReID shows that the standard way of training re-ID for MOT—comparing a query patch to a few positive and negative patches—is misaligned with the inference-time task of jointly matching all trajectories against all detections. Training on video-level episodes that reproduce this global assignment raises association recall (AssR) from 54.9 to 61.9 while cutting fragmentations. Adding frame-common appearance suppression, which estimates a shared direction via attention pooling and subtracts its component from all features, recovers precision and further cuts identity switches. The full method reaches 49.2 HOTA, 67.0 MOTA, 46.6 AssA, 61.7 AssR, and 64.9 IDF1 on BEE24 under shared detections,

What carries the argument

Two modules carry the argument. Frame-Common Appearance Estimation (FCAE) takes all current-frame detection features, lets a learnable query attend to them, and outputs a single unit-norm vector B representing the appearance common to the frame's objects. Common-Appearance Suppression (CAS) applies the same B to both detection and trajectory features by subtracting α⊙(x·B^T)B, where α is a learnable per-dimension sigmoid gate shared across samples, and then re-normalizes. The residual features form a similarity matrix that is min–max normalized within the current trajectory-to-detection block, producing a set-relative appearance cost. The training objective is a row- and column-wise candidat

Load-bearing premise

The method assumes that a single learned direction, computed from the current frame's detections, isolates appearance components shared by all objects, and that removing it from trajectory and detection features does not discard the identity-specific cues needed to tell similar objects apart.

What would settle it

Measure the angle between the learned B and the first principal component of the same-frame detection features on a diverse-appearance tracking set; if they are close, the 'common appearance' interpretation is suspect. A direct falsifier is to fix the suppression gate to α=1 for all dimensions and show that on a heterogeneous benchmark the method performs no better—or worse—than the unmodified baseline, indicating that the learned gate, not the common-appearance prior, is responsible for the gain.

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

If this is right

  • If re-ID for tracking is trained on full candidate sets rather than isolated pairs, identity scores become comparable across frames, so any tracker with a pairwise cosine appearance cost could swap in this set-relative cost without changing its motion model or matching logic.
  • The common-appearance suppression is fully self-supervised; it requires no pose, part, or identity annotations beyond what the tracker already has, making it applicable to other species or objects with uniform appearance.
  • The paper shows that the training-inference mismatch itself—not just weak features—drives identity switches in dense scenes; this reframes how appearance models for MOT should be evaluated and trained.
  • The ablation indicates the two components are complementary: video-level episode training boosts recall of ground-truth associations, while common-appearance suppression restores precision; a future tracker could potentially get the same benefit by tuning the balance between row and column losses differently.

Where Pith is reading between the lines

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

  • The use of a single shared suppression direction is a strong simplifying assumption; in scenes where the 'common' appearance is itself multi-modal (varying lighting, multiple poses, or sub-species markings), a mixture of directions or a per-candidate-set basis might be needed, and the paper's single-direction mechanism may not transfer directly.
  • One could test whether the learned B is genuinely capturing common appearance or simply the direction of maximum variance: if B aligns with the first principal component of the detection features, the gain may come from decorrelation rather than from subtracting a semantically shared appearance.
  • Because the FCAE attention pool operates on all detections in the frame, it is vulnerable to detector false positives; a robust variant that weights detections by confidence or excludes low-quality boxes might make the mechanism degrade more gracefully in noisier detection regimes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes VLA-ReID, a re-identification method for multi-object tracking in scenes with highly similar objects (e.g., bee swarms). It reformulates re-ID training from instance-level matching to video-level association: historical trajectory features serve as queries and all current-frame detections as candidates, enabling direct optimization of the global trajectory-to-detection association matrix. Two additional modules, FCAE and CAS, estimate a common appearance direction from the current detection set and suppress that component from both trajectory and detection features, with the goal of amplifying identity-discriminative differences without extra annotations. The method is integrated into the TrackTrack tracker by replacing only the pairwise appearance cost. Experiments on the BEE24 dataset report improvements over TrackTrack and other state-of-the-art trackers in HOTA, MOTA, AssR, AssA, IDF1, and identity switches.

Significance. If the empirical claims hold, the video-level training/inference alignment is a conceptually clean and potentially transferable idea for MOT in high-similarity settings, and the annotation-free common-appearance suppression is an attractive alternative to pose/part supervision. The paper includes a systematic ablation, a sensitivity sweep, and ranking/gap diagnostics, and promises source code and dataset release. However, the current experimental protocol — in particular test-set selection of the injection weight and the absence of variance estimates — prevents the headline numbers from being interpreted as a generalizable result. The contribution is therefore significant in potential but not yet established.

major comments (3)
  1. [§III-D, Eq. (14), Tables I–II] The injection weight c is selected on the BEE24 test set. The text states: 'We selected c=0.50 because it removes the mixing ratio at inference and achieved the lowest IDs while remaining close to the best HOTA value in the sweep.' All final results in Tables I and II are then reported with this c. This is test-set model selection, not independent evaluation. No validation split or nested selection is provided, and no error bars or multiple-seed runs are reported. Given that the main deltas are small (HOTA 48.1→49.2, AssA 45.9→46.6) while FP changes from 6026 to 11956 and FN from 51093 to 43096, the reported gains could lie within run-to-run variation. Please provide a validation-based selection of c (or fix c a priori) and report variance across seeds for the final configuration.
  2. [Table I] The ablation does not fully support the claim that each component contributes monotonically. 'Baseline + Episode' alone achieves MOTA 67.5, which is higher than the final 'Baseline + Episode + FCAE + CAS' MOTA of 67.0, and FP 14856 vs 11956. The final model improves AssR, AssA, IDF1, and IDs, but the MOTA regression relative to the episode-only variant is not discussed. Since the abstract highlights MOTA gains (+0.3 over TOPICTrack, +1.3 over TrackTrack), the paper should explicitly address this trade-off and clarify why the final model is preferred overall. Without such discussion, the ablation is ambiguous about the contribution of FCAE+CAS.
  3. [§II-C, §II-D, Eq. (6)–(9)] FCAE and CAS are the only genuinely new modules, and their core assumption is that a single learned linear direction B, obtained via attention pooling over all current detections, captures the dominant common appearance without discarding identity-specific cues. This assumption is weakly constrained: the anti-collapse regularizer anchors B only with weight w_anc=0.1, and no analysis shows what B represents or whether it is contaminated by detector false positives or the direction of largest variance. Please add diagnostics such as (i) alignment of B with the top PCA component of Z, (ii) norm of the suppressed component and per-dimension gate values after training, and (iii) sensitivity of association quality to the choice of w_anc. These would directly address the correctness risk of the central mechanism.
minor comments (6)
  1. [Abstract] The abstract reports MOTA +0.3, but Table II shows +1.3 over TrackTrack and +0.3 over TOPICTrack. Please state the comparison baseline explicitly to avoid ambiguity.
  2. [Fig. 5] The caption says 'dashed lines denote the original TrackTrack baseline,' but the text does not refer to the dashed lines in the figure. Add a sentence in §III-D explaining what the dashed lines show.
  3. [Eq. (6)] The attention temperature τ_f is set to 1, but this is not mentioned in the implementation details. Please state it explicitly alongside the other hyperparameters.
  4. [§II-D] The initialization 'θ to a positive constant' is vague. Specify the numerical value used (e.g., θ0 = 4.0, giving α ≈ 0.98).
  5. [§II-B] Minor typo: 'exponentially aggregated historical trajectory features' should likely be 'exponentially weighted' or 'exponential moving average' to match Eq. (2).
  6. [References] References [18], [32], and [33] appear unrelated to MOT or re-ID (they concern RL, flood modeling, and nowcasting). Please verify whether these citations are necessary or replace them with more relevant work.

Circularity Check

1 steps flagged

Method derivation is self-contained, but the final c=0.50 model is selected on the BEE24 test set and then evaluated on the same set, making the headline metrics partially in-sample.

specific steps
  1. fitted input called prediction [Section III-D (Sensitivity Analysis), Eq. (14), Fig. 5; Table II]
    "We selected c=0.50 because it removes the mixing ratio at inference and achieved the lowest IDs while remaining close to the best HOTA value in the sweep. This full-replacement setting was used for the final model."

    The blending coefficient c in Eq. (14) is a free hyperparameter controlling how much of the proposed set-relative cost replaces the original pairwise appearance cost. It is chosen by sweeping over the BEE24 test set (selecting for lowest IDs and near-best HOTA), and the same test set is then used for all reported final metrics in Tables I and II. Thus the headline improvements (e.g., IDs 606->437, a 28% reduction) are the result of selecting the best-looking c on the evaluation set, not an independent out-of-sample prediction. The final reported performance is therefore partly an in-sample selected statistic rather than a predicted result of the method.

full rationale

The core derivation of VLA-ReID is not circular: FCAE estimates a direction B from current-frame detections by attention pooling (Eqs. 6-7), CAS removes a gated projection along B (Eq. 9), the residual similarity matrix S is then a deterministic cosine matrix (Eq. 11), and training uses candidate-set cross-entropy on video-level episodes (Eqs. 15-17). These steps do not reduce to their inputs; the components are optimizable and the training objective is standard supervised ranking. No uniqueness theorem or load-bearing self-citation chain is used; the BEE24 dataset [8] and TrackTrack [29] are concrete external artifacts that others can run against. The one genuine circularity-adjacent flaw is the test-set model selection of c (Section III-D): a hyperparameter is chosen on the BEE24 test set and the same test set is then reported as the evaluation, so the headline numbers are partially in-sample and no confidence intervals or validation split are provided. Because the main mechanism (episode training) already improves over baseline at c=0, the central claim retains independent content; the selection issue makes the exact reported numbers less trustworthy but does not collapse the derivation into a fit. Score 5 reflects this partial empirical circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 1 invented entities

The method rests on a learned single-direction appearance decomposition and a set of hand-chosen loss/temperature/injection hyperparameters, none of which are externally benchmarked. The most load-bearing assumption is that suppressing one learned common direction cannot remove identity-specific signal; the paper provides only a weak regularizer (w_anc=0.1) to keep B aligned with the detection distribution.

free parameters (7)
  • Candidate-set softmax temperature τ_r = 0.1
    Chosen by hand for episode training; controls sharpness of row/column cross-entropy in Eq. (15)-(16).
  • Loss weights λ_col, λ_com, w_sep, w_anc = 0.1, 0.5, 1.0, 0.1
    Hand-set in Eq. (17); no ablation or sensitivity analysis is reported for these values.
  • Injection weight c = 0.50
    Selected on the BEE24 test set (Section III-D, Fig. 5) as the value with lowest IDs; no validation set is used, so this is test-set tuning.
  • EMA decay μ = 0.95
    Fixed in Eq. (2); the quality of historical trajectory features depends on this choice.
  • Attention temperature τ_f = 1
    Fixed in Eq. (6); no sensitivity analysis.
  • Suppression gate initializer θ = not specified (chosen so α=σ(θ)≈1)
    Ad hoc initialization to make CAS start close to full suppression; the constant is not reported.
  • AGW encoder + FCAE attention (q, W_K, W_V) + CAS gate θ = learned on BEE24 training split from random initialization
    All discriminative power comes from fitting these parameters on the same benchmark; no external pretraining or independent dataset is used, so generalization beyond BEE24 is unmeasured.
axioms (5)
  • ad hoc to paper One learned linear direction B captures the dominant common appearance of all current-frame detections
    FCAE/CAS Eq. (6)-(9) posit a single direction; the paper states this 'limits removal of identity-bearing information' but provides no analysis for multi-modal appearance distributions.
  • domain assumption Subtracting B from both trajectory and detection features does not remove identity-discriminative information
    CAS Eq. (9) suppresses the same component from queries and candidates; if identity cues align with B, suppression would hurt discrimination. The anti-collapse regularizer only weakly constrains B (w_anc=0.1).
  • domain assumption Current-frame detections are reliable enough that attention pooling over all of them yields a meaningful common direction
    FCAE pools over all N detections including false positives and unmatched boxes; no sensitivity to detector noise is reported.
  • domain assumption Exponential moving average with μ=0.95 is a sufficient trajectory history representation
    Eq. (2) collapses all past observations into one vector; the method claims 'video-level' but has no temporal attention or uncertainty modeling.
  • domain assumption BEE24 is representative of highly-similar-object MOT and the shared-detection protocol yields fair comparison
    All conclusions are drawn from one dataset; the unshaded top block of Table II uses non-shared detections and is not directly comparable.
invented entities (1)
  • Frame-Common Appearance direction B no independent evidence
    purpose: A learned unit vector representing the appearance shared by all detections in a frame; used by CAS for suppression.
    B is a latent feature direction trained on BEE24 and has no external falsifiable handle; it could be dataset-specific.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of VLA-ReID: Video-Level Association for Re-Identification in Multi-Object Tracking with Highly Similar Objects." pith.science (2026). https://pith.science/paper/C4MLOXOB

@misc{pith2026260717157,
  author       = {Pith},
  title        = {Pith review of: VLA-ReID: Video-Level Association for Re-Identification in Multi-Object Tracking with Highly Similar Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4MLOXOB}},
  note         = {Machine review of arXiv:2607.17157}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multi-object tracking (MOT) aims to localize multiple objects in videos while preserving their identities over time. Long-term identity preservation remains difficult when objects are small, densely distributed, and highly similar in appearance, as in bee swarm scenes. Existing trackers rely on re-identification (re-ID) models trained through single-instance assignment (instance-level querying). At inference, however, MOT requires global assignment between multiple trajectories and detections, corresponding to video-level querying. This training-inference mismatch can cause identity switches among visually similar objects. Existing approaches also often require substantial additional annotations to enhance appearance discrimination. We propose Video-Level Association re-ID (VLA-ReID), which reformulates re-ID as video-level association modeling. It uses aggregated historical trajectory features as queries and all current-frame detections as candidates, enabling direct optimization of their global association at each frame. In addition, Frame-Common Appearance Estimation (FCAE) estimates a common appearance direction from current-frame detections, while Common-Appearance Suppression (CAS) removes the corresponding component along this direction from trajectory and detection features. This amplifies discriminative differences among highly similar objects without additional annotations. Experiments on BEE24 show that VLA-ReID improves HOTA by 1.1, MOTA by 0.3, AssR by 2.6, AssA by 0.7, and IDF1 by 0.8 over state-of-the-art trackers, while reducing identity switches by 28%. These results demonstrate the effectiveness of video-level re-ID modeling for appearance-based association in MOT.

Figures

Figures reproduced from arXiv: 2607.17157 by Xiaoyan Cao, Yanrong Qin, Yao Yao.

Figure 1
Figure 1. Figure 1: Motivation and overview of VLA-ReID. (a) Conventional re-ID training learns appearance embeddings through instance-level querying. (b) During [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Detailed workflow of FCAE and CAS. FCAE estimates the common appearance direction [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of the per-query similarity gap between the true candidate [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Candidate-ranking performance. Rank-1, Top-3, and Top-5 report [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Parameter-free commonality suppression moves the true candidate from rank 5 to rank 1 among 33 detections in a dense frame from BEE24-34. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Parameter-free commonality suppression moves the true candidate from rank 2 to rank 1 among 15 detections in BEE24-36. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The learned FCAE + CAS model moves the true candidate from rank 17 to rank 1 among 27 detections in BEE24-13. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

48 extracted references · 4 linked inside Pith

  1. [1]

    Markerless tracking of an entire honey bee colony,

    K. Bozek, L. Hebert, Y . Portugal, A. S. Mikheyev, and G. J. Stephens, “Markerless tracking of an entire honey bee colony,”Nature communi- cations, vol. 12, no. 1, p. 1733, 2021

    2021

  2. [2]

    Idtracker.ai: tracking all individuals in small or large collectives of unmarked animals,

    F. Romero-Ferrero, M. G. Bergomi, R. C. Hinz, F. J. Heras, and G. G. De Polavieja, “Idtracker.ai: tracking all individuals in small or large collectives of unmarked animals,”Nature methods, vol. 16, no. 2, pp. 179–182, 2019

    2019

  3. [3]

    Online tracking of ants based on deep association metrics: method, dataset and evaluation,

    X. Cao, S. Guo, J. Lin, W. Zhang, and M. Liao, “Online tracking of ants based on deep association metrics: method, dataset and evaluation,” Pattern Recognition, vol. 103, p. 107233, 2020

    2020

  4. [4]

    A dataset of ant colonies’ motion trajectories in indoor and outdoor scenes to study clustering behavior,

    M. Wu, X. Cao, M. Yang, X. Cao, and S. Guo, “A dataset of ant colonies’ motion trajectories in indoor and outdoor scenes to study clustering behavior,”GigaScience, vol. 11, p. giac096, 2022

    2022

  5. [5]

    Swarm behavior tracking based on a deep vision algorithm,

    M. Wu, X. Cao, and S. Guo, “Swarm behavior tracking based on a deep vision algorithm,”arXiv preprint arXiv:2204.03319, 2022

    Pith/arXiv arXiv 2022

  6. [6]

    Intelligent beehive monitoring system based on internet of things and colony state analysis,

    Y . Zheng, X. Cao, S. Xu, S. Guo, R. Huang, Y . Li, Y . Chen, L. Yang, X. Cao, and H. Sun, “Intelligent beehive monitoring system based on internet of things and colony state analysis,”Smart Agricultural Technology, 2024

    2024

  7. [7]

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

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

    2022

  8. [8]

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

    X. Cao, Y . Zheng, Y . Yao, H. Qin, X. Cao, and S. Guo, “TOPIC: A parallel association paradigm for multi-object tracking under complex motions and diverse scenes,”IEEE Transactions on Image Processing, vol. 34, pp. 743–758, 2025

    2025

  9. [9]

    Simple online and realtime tracking,

    A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft, “Simple online and realtime tracking,” in2016 IEEE international conference on image processing (ICIP). IEEE, 2016, pp. 3464–3468

    2016

  10. [10]

    Simple online and realtime tracking with a deep association metric,

    N. Wojke, A. Bewley, and D. Paulus, “Simple online and realtime tracking with a deep association metric,” in2017 IEEE international conference on image processing (ICIP). IEEE, 2017, pp. 3645–3649

    2017

  11. [11]

    Quasi- dense similarity learning for multiple object tracking,

    J. Pang, L. Qiu, X. Li, H. Chen, Q. Li, T. Darrell, and F. Yu, “Quasi- dense similarity learning for multiple object tracking,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 164–173

    2021

  12. [12]

    Adaptive sparse pairwise loss for object re-identification,

    X. Zhou, Y . Zhong, Z. Cheng, F. Liang, and L. Ma, “Adaptive sparse pairwise loss for object re-identification,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 19 691–19 701

    2023

  13. [13]

    Deep quadruplet appearance learning for vehicle re-identification,

    J. Hou, H. Zeng, J. Zhu, J. Hou, J. Chen, and K.-K. Ma, “Deep quadruplet appearance learning for vehicle re-identification,”IEEE Transactions on V ehicular Technology, vol. 68, no. 9, pp. 8512–8522, 2019

    2019

  14. [14]

    In defense of the triplet loss for person re-identification,

    A. Hermans, L. Beyer, and B. Leibe, “In defense of the triplet loss for person re-identification,”arXiv preprint arXiv:1703.07737, 2017

    Pith/arXiv arXiv 2017

  15. [15]

    Multiple object tracking as id predic- tion,

    R. Gao, J. Qi, and L. Wang, “Multiple object tracking as id predic- tion,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 27 883–27 893

    2025

  16. [16]

    Deconfusetrack: Dealing with confusion for multi-object tracking,

    C. Huang, S. Han, M. He, W. Zheng, and Y . Wei, “Deconfusetrack: Dealing with confusion for multi-object tracking,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 19 290–19 299

    2024

  17. [17]

    Towards discriminative representation: Multi- view trajectory contrastive learning for online multi-object tracking,

    E. Yu, Z. Li, and S. Han, “Towards discriminative representation: Multi- view trajectory contrastive learning for online multi-object tracking,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8834–8843

    2022

  18. [18]

    Sample efficient re- inforcement learning via model-ensemble exploration and exploitation,

    Y . Yao, L. Xiao, Z. An, W. Zhang, and D. Luo, “Sample efficient re- inforcement learning via model-ensemble exploration and exploitation,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 4202–4208

    2021

  19. [19]

    Honeybee re-identification in video: New datasets and impact of self- supervision

    J. Chan, H. Carri ´on, R. M ´egret, J. L. Agosto-Rivera, and T. Giray, “Honeybee re-identification in video: New datasets and impact of self- supervision.” inVISIGRAPP (5: VISAPP), 2022, pp. 517–525

    2022

  20. [20]

    A purely visual re- id approach for bumblebees (bombus terrestris),

    P. Borlinghaus, F. Tausch, and L. Rettenberger, “A purely visual re- id approach for bumblebees (bombus terrestris),”Smart Agricultural Technology, vol. 3, p. 100135, 2023

    2023

  21. [21]

    Openanimals: Revisiting person re-identification for animals towards better generalization,

    S. Hou, P. Huang, Z. Wang, Y . Liu, Z. Li, M. Zhang, and Y . Huang, “Openanimals: Revisiting person re-identification for animals towards better generalization,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 14 369–14 379

    2025

  22. [22]

    Similarity learning net- works for animal individual re-identification: an ecological perspective,

    S. Schneider, G. W. Taylor, and S. C. Kremer, “Similarity learning net- works for animal individual re-identification: an ecological perspective,” Mammalian Biology, vol. 102, no. 3, pp. 899–914, 2022

    2022

  23. [23]

    Han- wooreid: Multi-view cattle re-identification with pose-aware transformer enhancements,

    J. Liu, A. Fuentes, S. Han, Y . Jeong, S. Yoon, and D. S. Park, “Han- wooreid: Multi-view cattle re-identification with pose-aware transformer enhancements,”Computers and Electronics in Agriculture, vol. 239, p. 111117, 2025

    2025

  24. [24]

    Detail enhancement-based vehicle re-identification with orientation-guided re-ranking,

    Z. Sun, X. Nie, X. Bi, S. Wang, and Y . Yin, “Detail enhancement-based vehicle re-identification with orientation-guided re-ranking,”Pattern Recognition, vol. 137, p. 109304, 2023

    2023

  25. [25]

    An improved architecture for part-based animal re- identification through semantic segmentation distillation,

    E. D. R. Neto, M. Chaumont, G. Subsol, M. de Garine-Wichatitsky, and H. Guis, “An improved architecture for part-based animal re- identification through semantic segmentation distillation,” inProceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2026, pp. 5280–5289

    2026

  26. [26]

    Species-agnostic patterned ani- mal re-identification by aggregating deep local features,

    E. Nepovinnykh, I. Chelak, T. Eerola, V . Immonen, H. K ¨alvi¨ainen, M. Kholiavchenko, and C. V . Stewart, “Species-agnostic patterned ani- mal re-identification by aggregating deep local features,”International Journal of Computer Vision, vol. 132, no. 9, pp. 4003–4018, 2024

    2024

  27. [27]

    Multi-branch feature learning network via global-local self-distillation for vehicle re-identification,

    X. Chen, H. Yu, C. Hu, and H. Wang, “Multi-branch feature learning network via global-local self-distillation for vehicle re-identification,” IEEE Transactions on V ehicular Technology, vol. 73, no. 9, pp. 12 415– 12 425, 2024

    2024

  28. [28]

    Overcoming fine-grained visual challenges in animal re-identification via semantic feature alignment,

    Y . Wu, D. Zhao, Y . Li, M. Alajas, A. S. Glen, J. Zhang, G. Dobbie, D. Wilson, and Y . S. Koh, “Overcoming fine-grained visual challenges in animal re-identification via semantic feature alignment,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2026, pp. 371–381

    2026

  29. [29]

    Focusing on tracks for online multi-object tracking,

    K. Shim, K. Ko, Y . Yang, and C. Kim, “Focusing on tracks for online multi-object tracking,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 11 687–11 696

    2025

  30. [30]

    Fastreid: A pytorch toolbox for general instance re-identification,

    L. He, X. Liao, W. Liu, X. Liu, P. Cheng, and T. Mei, “Fastreid: A pytorch toolbox for general instance re-identification,” inProceedings of the 31st ACM international conference on multimedia, 2023, pp. 9664– 9667

    2023

  31. [31]

    Deep learning for person re-identification: A survey and outlook,

    M. Ye, J. Shen, G. Lin, T. Xiang, L. Shao, and S. C. Hoi, “Deep learning for person re-identification: A survey and outlook,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 6, pp. 2872– 2893, 2021

    2021

  32. [32]

    Large- scale urban flood modeling and zero-shot high-resolution generalization with larno,

    X. Cao, Y . Yao, Z. Wang, Z. Zhao, A. G. Borthwick, and H. Qin, “Large- scale urban flood modeling and zero-shot high-resolution generalization with larno,”Journal of Hydrology, p. 135686, 2026

    2026

  33. [33]

    U-rnn high-resolution spatiotemporal nowcasting of urban flooding,

    X. Cao, B. Wang, Y . Yao, L. Zhang, Y . Xing, J. Mao, R. Zhang, G. Fu, A. G. Borthwick, and H. Qin, “U-rnn high-resolution spatiotemporal nowcasting of urban flooding,”Journal of Hydrology, vol. 659, p. 133117, 2025

    2025

  34. [34]

    i-razor: A differentiable neural input razor for feature selection and dimension search in dnn-based recommender systems,

    Y . Yao, B. Liu, H. He, D. Sheng, K. Wang, L. Xiao, and H. Cao, “i-razor: A differentiable neural input razor for feature selection and dimension search in dnn-based recommender systems,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 9, pp. 4736–4749, 2023

    2023

  35. [35]

    Chained-tracker: Chaining paired attentive regres- sion results for end-to-end joint multiple-object detection and tracking,

    J. Peng, C. Wang, F. Wan, Y . Wu, Y . Wang, Y . Tai, C. Wang, J. Li, F. Huang, and Y . Fu, “Chained-tracker: Chaining paired attentive regres- sion results for end-to-end joint multiple-object detection and tracking,” inEuropean conference on computer vision. Springer, 2020, pp. 145– 161

    2020

  36. [36]

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

    Y . Zhang, C. Wang, X. Wang, W. Zeng, and W. Liu, “Fairmot: On the fairness of detection and re-identification in multiple object tracking,” International journal of computer vision, vol. 129, no. 11, pp. 3069– 3087, 2021

    2021

  37. [37]

    Track to detect and segment: An online multi-object tracker,

    J. Wu, J. Cao, L. Song, Y . Wang, M. Yang, and J. Yuan, “Track to detect and segment: An online multi-object tracker,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 12 352–12 361

    2021

  38. [38]

    Track- former: Multi-object tracking with transformers,

    T. Meinhardt, A. Kirillov, L. Leal-Taixe, and C. Feichtenhofer, “Track- former: Multi-object tracking with transformers,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8844–8854

    2022

  39. [39]

    Do different tracking tasks require different appearance models?

    Z. Wang, H. Zhao, Y .-L. Li, S. Wang, P. Torr, and L. Bertinetto, “Do different tracking tasks require different appearance models?”Advances in neural information processing systems, vol. 34, pp. 726–738, 2021

    2021

  40. [40]

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

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

    2022

  41. [41]

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

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

    2023

  42. [42]

    Yolox: Exceeding yolo series in 2021,

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “Yolox: Exceeding yolo series in 2021,”arXiv preprint arXiv:2107.08430, 2021. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11

    Pith/arXiv arXiv 2021

  43. [43]

    Fine-tuning cnn image retrieval with no human annotation,

    F. Radenovi ´c, G. Tolias, and O. Chum, “Fine-tuning cnn image retrieval with no human annotation,”IEEE transactions on pattern analysis and machine intelligence, vol. 41, no. 7, pp. 1655–1668, 2018

    2018

  44. [44]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  45. [45]

    Giaotracker: A comprehensive framework for mcmot with global information and optimizing strategies in visdrone 2021,

    Y . Du, J. Wan, Y . Zhao, B. Zhang, Z. Tong, and J. Dong, “Giaotracker: A comprehensive framework for mcmot with global information and optimizing strategies in visdrone 2021,” inProceedings of the IEEE/CVF International conference on computer vision, 2021, pp. 2809–2819

    2021

  46. [46]

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

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

    2021

  47. [47]

    Evaluating multiple object tracking performance: the clear mot metrics,

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

    2008

  48. [48]

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

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

    2016

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.