Pith. sign in

REVIEW 3 major objections 3 minor 59 references

Visual domain shifts severely degrade referring multi-object tracking, and the paper argues the dominant failure is unstable language-conditioned temporal association and target selection, not detection errors.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 02:59 UTC pith:ZPLXFQGC

load-bearing objection A genuinely useful benchmark with a well-documented pipeline, but the core annotation-transfer mapping is unverified without the dataset, and the method comparison needs stronger baselines before I'd trust the headline claims. the 3 major comments →

arxiv 2607.25239 v1 pith:ZPLXFQGC submitted 2026-07-28 cs.CV

CD-RMOT-Bench: Benchmarking the Cross-Domain Referring Multi-Object Tracking

Xiangqun Zhang , Likai Wang , Zekun Qian , Ruize Han , Wei Feng This is my paper
classification cs.CV
keywords cross-domain adaptationreferring multi-object trackinglanguage-guided trackingdomain shiftbenchmarktemporal associationquery-centric adaptationdigital twin
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.

This paper introduces and studies cross-domain referring multi-object tracking (CD-RMOT): can a model trained on labeled source video keep following a natural-language expression when the target video has different weather, viewpoint, or realism? To answer it, the paper builds a unified benchmark whose controlled core transfers real referring annotations onto aligned synthetic scene variants, so weather and viewpoint can change while scene structure and expression semantics stay fixed. Experiments show domain shift sharply degrades tracking accuracy, and the dominant cause is not missed detections but unstable expression-conditioned temporal association and target selection. The paper then proposes a query-centric adaptation method that stabilizes and aligns language-conditioned temporal queries, recovering a large fraction of the lost accuracy. If the analysis is right, cross-domain robustness needs to be treated as a distinct evaluation axis for language-guided tracking.

Core claim

The central discovery is that under a changed visual domain, a referring tracker can still detect plausible objects but often fails to keep the selected trajectory consistent with the language expression over time: the language-conditioned temporal queries drift, causing wrong identity association and target selection. This is established by measuring HOTA, DetA, and AssA drops against an in-domain reference on the benchmark, where association accuracy degrades more sharply than detection accuracy under fog and rain. The paper further shows that a query-centric adaptation method — stabilizing target queries with teacher-student consistency, aligning task-relevant queries adversarially, and u

What carries the argument

The load-bearing object is the benchmark's controlled digital-twin core: real referring annotations are carried over to aligned synthetic scene variants through per-scene identity correspondence, so that the same expression and target semantics reappear across clear, fog, rain, morning, overcast, sunset, and multiple camera viewpoints. The method side relies on the language-conditioned temporal query as the central entity: three mechanisms — selective teacher-student consistency, task-relevant domain-adversarial alignment, and language-anchored prototype refinement — all operate on this query space to keep it stable, transferable, and semantically faithful.

Load-bearing premise

The benchmark's controlled conclusions rest on the per-scene identity mapping between real and synthetic objects being correct, so that each transferred expression still points at the semantic target it described in the real domain; if that mapping misfires, the measured association failures are not cleanly attributable to domain shift.

What would settle it

Sampling a few hundred of the 251 mapped identity correspondences and checking, in the target synthetic frames, whether the referred object still matches the expression's attributes (color, type, position, motion) would settle the benchmark's validity; a large mismatch rate would mean the controlled domain-shift diagnosis is an artifact of annotation transfer errors.

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

If this is right

  • A tracker can have healthy detections and still fail referring multi-object tracking under domain shift, so evaluation must report association separately from detection.
  • Controlled paired domain shifts are necessary to attribute RMOT failure to the visual domain rather than to unrelated scene changes.
  • Language-conditioned temporal query stability is a bottleneck, and methods that stabilize it recover a large share of cross-domain accuracy.
  • Mask-based referring video segmentation methods remain far behind identity-aware RMOT trackers under this trajectory protocol.
  • The remaining gap to in-domain performance shows that cross-domain referring multi-object tracking is far from solved even with the proposed adaptation.

Where Pith is reading between the lines

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

  • If the association-failure diagnosis generalizes, then test-time or continual adaptation for language-guided tracking should prioritize query and embedding consistency over detector retraining — a testable prediction.
  • The same controlled-transfer recipe could be applied to other language-conditioned video tasks to check whether their domain failures are also association-dominated rather than detection-dominated.
  • A next natural step is to vary expression difficulty under fixed domain shift, testing whether spatial, relational, or motion-heavy expressions are more or less robust to query drift.
  • The protocol leaves open source-free or test-time variants; if query instability is the bottleneck, those settings should show even larger association drops.

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 / 3 minor

Summary. The paper introduces CD-RMOT-Bench, a benchmark for cross-domain referring multi-object tracking (RMOT), combining a real clear-domain anchor (Refer-KITTI-V2), a controlled digital-twin core (Refer-vKITTI), and a real adverse-domain anchor (Refer-BDD). Refer-vKITTI is constructed by transferring referring expressions and track identities from aligned KITTI scenes to Virtual KITTI 2 variants across weather and viewpoint shifts. The paper also proposes a Query-Centric Adaptation (QCA) method with three components: cross-domain temporal mean teacher (CDT-MT), adaptive curricular domain adversarial alignment (ACDA), and language-guided continuous adaptation (LGCA). Experiments report severe performance drops under domain shift, larger in association than in detection metrics, and claim that QCA establishes a strong baseline by recovering part of the drop.

Significance. If valid, CD-RMOT-Bench addresses a genuine gap: existing RMOT benchmarks are largely in-domain, and existing UDA benchmarks do not evaluate language-conditioned trajectories. The controlled digital-twin design, with shared scene structure and transferred expressions, is a principled way to isolate domain shift from semantic change. The paper documents the construction pipeline in unusual detail (identity mapping, expression transfer, quality control) and provides a clean protocol in which target-domain expressions are task inputs but target annotations are used only for evaluation. QCA targets a plausible failure mode—unstable language-conditioned temporal queries—and the metric decomposition is a useful diagnostic. However, the strength of both the benchmark and the method claims currently depends on two things the paper does not yet provide: external verification of the transferred labels and comparison against standard UDA baselines.

major comments (3)
  1. [Table 2 and §5.2] The claim that QCA 'establishes a strong baseline' is not supported by the comparison set. All baselines are RMOT or RVOS methods; no existing UDA method (e.g., DANN [28], DA-Faster R-CNN [32], Unbiased Mean Teacher [35], Adaptive Teacher [36]) is evaluated, despite these being cited in §2. Without standard UDA baselines under the same backbone and protocol, it is impossible to tell whether QCA's gains come from its query-centric design or from generic adversarial/mean-teacher adaptation. Please add such baselines; if they are incompatible with the RMOT protocol, justify why and provide the closest feasible comparisons.
  2. [Appendix A.1, Eq. (11) and Table 3] The correctness of Refer-vKITTI, and therefore every downstream conclusion about expression-conditioned association failure, rests on the KITTI-to-vKITTI identity mapping. The mapping uses hand-set score weights in Eq. (11), a deliberately permissive IoU evidence threshold of 0.02, and yields only 251 identity correspondences reused across 10 domains. The paper explicitly states that manual quality control is not logged as a statistic ('manual inspection is used only as residual quality control and is therefore described in the text rather than reported as a separate logged statistic'). The dataset is not released. If a nontrivial fraction of these correspondences are wrong—especially in the two held-out scenes used for fog, rain, and 30°L evaluation—the transferred expressions point to the wrong synthetic targets and the benchmark's central conclusions collapse. Please release the datas
  3. [§5.2 and Table 5] The central diagnostic finding—'failure is not merely caused by object detection errors but more critically by unstable expression-conditioned temporal association and target selection'—is inferred from the HOTA/DetA/AssA decomposition. While AssA drops are larger than DetA drops, this is an indirect test. A direct test would be an oracle experiment in which ground-truth target detections (or cached source-domain detections) are provided to the tracker, so that the remaining degradation is purely association/selection. Without such a test, the 'not merely detection' phrasing is stronger than the evidence. Please either add an oracle-detection experiment or temper the claim accordingly.
minor comments (3)
  1. [Table 2] Several entries are run together (e.g., '11.475.4424.57' in the last row), making the table hard to read. Please fix spacing and alignment.
  2. [Appendix C.3, Table 6] The ablation gains for single components are small (e.g., CDT-MT only: +0.25 HOTA on fog). No standard deviation or multiple-seed results are reported. Since the source code is not yet released, please report variance or state that the results are from a single seed.
  3. [Appendix C.1] The RVOS baseline adaptation converts masks to boxes by taking the tight bounding box of all foreground pixels; this is a weak proxy that can degrade association. The paper acknowledges this, but it would be helpful to state in §5.2 that the RVOS numbers are lower bounds for the RVOS family rather than a fair comparison.

Circularity Check

0 steps flagged

No significant circularity: the benchmark, diagnosis, and QCA baseline are empirically grounded in held-out target-domain annotations rather than by-construction identities.

full rationale

The paper's central claims are empirical findings from a purpose-built benchmark, not consequences of the benchmark's construction equations. Refer-vKITTI is built by transferring Refer-KITTI-V2 expressions to aligned Virtual KITTI 2 scenes via the matching score in Eq. (11); this is a dataset-construction step, not a prediction about RMOT behavior. The conclusion that domain shift degrades expression-conditioned temporal association and target selection is drawn from HOTA/DetA/AssA evaluations against target-domain boxes, identities, masks, and expression-object associations that are explicitly withheld from training (Table 4 and Sec. 3.1). QCA is trained without target-domain supervised labels and is compared with external RMOT/RVOS baselines, so its gains are not encoded by the benchmark design. The identity-mapping threshold and scoring function determine which synthetic targets receive transferred expressions, but they do not determine whether a tracker will select those targets; the failure diagnosis therefore remains an empirical result, conditional on the correctness of the transferred annotations. That correctness is a data-quality/verifiability concern, not circularity. The paper's self-citations ([44,45]) appear only as related-work motivation and are not load-bearing for the benchmark or the QCA method. No prediction reduces by construction to a fitted parameter, and no uniqueness theorem or ansatz is imported from the authors' prior work. The benchmark's internal consistency is a limitation of external auditability rather than a circular derivation, and a non-finding is therefore appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claims rest on the correctness of the digital-twin identity mapping and on the assumption that transferred referring expressions preserve their semantics. No new physical entities or forces are introduced. QCA contains several free hyperparameters whose values are not disclosed, and the benchmark construction relies on hand-chosen matching weights and thresholds.

free parameters (5)
  • QCA loss trade-off weights (lambda_mt, lambda_acda, lambda_lgca)
    Introduced in Eq. (2); numeric values are never reported in the main text or appendix, so they are tuned or selected rather than derived. QCA's reported gains depend on them.
  • CDT-MT EMA decay alpha and agreement balance gamma = alpha=0.999, gamma in [0,1]
    Hyperparameters for teacher-student query selection in Eqs. (3) and (17); no sensitivity analysis is provided.
  • ACDA retention top-M and adversarial schedule (lambda_min, lambda_max, warmup) = M=16; lambda_min/lambda_max not given
    Controls how many target queries are adversarially aligned; values are hand-set and affect the ACDA contribution.
  • LGCA prototype momentum m and quality gate = m=0.9
    Momentum update in Eq. (9) and gate parameters are hand-set; no ablation on m is reported.
  • Identity-matching score weights and IoU threshold = w=(4,30,6,2,1,3); IoU>=0.02
    Appendix A.1 Eq. (11) uses hand-chosen weights and a permissive IoU threshold to build KITTI-to-vKITTI identity mappings. These choices determine every transferred annotation in Refer-vKITTI.
axioms (5)
  • domain assumption Virtual KITTI 2 scenes are aligned digital twins of KITTI scenes, enabling a per-scene KITTI-to-vKITTI identity mapping.
    The entire Refer-vKITTI construction (Section 3.2, Stage 2) relies on this; vKITTI is a re-rendered proxy, not a pixel-perfect twin.
  • domain assumption The identity mapping inferred from IoU and geometric scoring is correct enough to transfer referring expressions to synthetic targets.
    Appendix A.1 Eq. (11) uses hand-set weights and IoU=0.02; errors would attach expressions to wrong synthetic objects and invalidate benchmark labels.
  • domain assumption Referring expressions remain semantically valid after transfer to fog, rain, and viewpoint variants.
    Stage 4 residual QC filters some ambiguous cases, but the benchmark treats expression semantics as preserved across domains (Section 3.2).
  • domain assumption Target-domain referring expressions are task inputs and do not leak target boxes, track identities, or expression-object associations.
    Section 3.1 asserts this distinction; QCA's unsupervised status and the UDA protocol depend on it.
  • standard math HOTA/DetA/AssA decomposition isolates detection failures from association/selection failures.
    Standard tracking metrics are used for diagnosis in Section 5.2, but the 'more critically' claim is an interpretation of metric drops, not a formal causal test.

pith-pipeline@v1.3.0-alltime-deepseek · 16421 in / 10210 out tokens · 113214 ms · 2026-08-01T02:59:27.783135+00:00 · methodology

0 comments
read the original abstract

Referring multi-object tracking (RMOT) extends tracking from category-driven perception to language-guided understanding by grounding object trajectories in natural-language expressions. Despite recent progress, existing RMOT studies are largely conducted under in-domain settings, leaving the robustness of language-conditioned tracking under inevitable visual domain shifts unexplored. In this paper, we study Cross-Domain Referring Multi-Object Tracking (CD-RMOT), a new and challenging problem that evaluates whether an RMOT model trained on a labeled source domain can reliably follow natural-language expressions in an unlabeled target domain with different visual conditions. To support systematic study, we construct CD-RMOT-Bench, a unified benchmark that combines real clear-domain referring tracking data, aligned digital-twin variants, and real adverse-domain videos. CD-RMOT-Bench enables both controlled weather/viewpoint shift analysis and realistic synthetic-real transfer evaluation under a shared RMOT protocol. Further, we provide a Query-Centric Adaptation (QCA) framework, designed to stabilize the query space that bridges visual trajectories and referring expressions. Extensive experiments reveal that domain shifts severely degrade RMOT performance, where the failure is not merely caused by object detection errors but more critically by unstable expression-conditioned temporal association and target selection. QCA establishes a strong baseline, while CD-RMOT-Bench opens a new direction for robust language-guided tracking across visual domains.

Figures

Figures reproduced from arXiv: 2607.25239 by Likai Wang, Ruize Han, Wei Feng, Xiangqun Zhang, Zekun Qian.

Figure 1
Figure 1. Figure 1: Controlled domain shift in RMOT. (a) Under the same referring expression and target [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Refer-vKITTI construction and supported protocols. Refer-KITTI-V2 provides real-domain [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Benchmark characterization of CD-RMOT-Bench. Refer-vKITTI provides the controlled [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of QCA as a query-centric reference method for language-conditioned UDA [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Additional qualitative cases illustrating common cross-domain RMOT failures. The [PITH_FULL_IMAGE:figures/full_fig_p019_5.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

59 extracted references · 5 linked inside Pith

  1. [1]

    Bootstrapping referring multi-object tracking.arXiv preprint arXiv:2406.05039, 2024

    Yani Zhang, Dongming Wu, Wencheng Han, and Xingping Dong. Bootstrapping referring multi-object tracking.arXiv preprint arXiv:2406.05039, 2024. 2, 3, 4, 5, 9

    arXiv 2024

  2. [2]

    Simple online and realtime tracking

    Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. InICIP, 2016. 3

    2016

  3. [3]

    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. InICIP, 2017

    2017

  4. [4]

    Tracking without bells and whistles

    Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixé. Tracking without bells and whistles. InICCV, 2019

    2019

  5. [5]

    Tracking objects as points

    Xingyi Zhou, Vladlen Koltun, and Philipp Krähenbühl. Tracking objects as points. InECCV, 2020

    2020

  6. [6]

    FairMOT: On the fairness of detection and re-identification in multiple object tracking.International Journal of Computer Vision, 129:3069–3087, 2021

    Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. FairMOT: On the fairness of detection and re-identification in multiple object tracking.International Journal of Computer Vision, 129:3069–3087, 2021

    2021

  7. [7]

    ByteTrack: Multi-object tracking by associating every 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 every detection box. InECCV, 2022

    2022

  8. [8]

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

    Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation- centric SORT: Rethinking SORT for robust multi-object tracking. InCVPR, 2023. 3

    2023

  9. [9]

    TransTrack: Multiple object tracking with transformer.arXiv preprint arXiv:2012.15460, 2020

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

    Pith/arXiv arXiv 2012

  10. [10]

    Track- Former: Multi-object tracking with transformers

    Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixé, and Christoph Feichtenhofer. Track- Former: Multi-object tracking with transformers. InCVPR, 2022

    2022

  11. [11]

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

    Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. MOTR: End-to-end multiple-object tracking with transformer. InECCV, 2022

    2022

  12. [12]

    MOTRv2: Bootstrapping end-to-end multi- object tracking by pretrained object detectors

    Yuang Zhang, Tiancai Wang, and Xiangyu Zhang. MOTRv2: Bootstrapping end-to-end multi- object tracking by pretrained object detectors. InCVPR, 2023

    2023

  13. [13]

    Towards grand unification of object tracking

    Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. InECCV, 2022. 3

    2022

  14. [14]

    Referring multi-object tracking

    Dongming Wu, Wencheng Han, Tiancai Wang, Xingping Dong, Xiangyu Zhang, and Jianbing Shen. Referring multi-object tracking. InCVPR, 2023. 3, 9

    2023

  15. [15]

    iKUN: Speak to trackers without retraining

    Yunhao Du, Cheng Lei, Zhicheng Zhao, and Fei Su. iKUN: Speak to trackers without retraining. InCVPR, 2024. 3

    2024

  16. [16]

    EchoTrack: Auditory referring multi-object tracking for autonomous driving.IEEE Transactions on Intelligent Transportation Systems, 25(11):18964–18977, 2024

    Jiacheng Lin, Jiajun Chen, Kunyu Peng, Xuan He, Zhiyong Li, Rainer Stiefelhagen, and Kailun Yang. EchoTrack: Auditory referring multi-object tracking for autonomous driving.IEEE Transactions on Intelligent Transportation Systems, 25(11):18964–18977, 2024. 3, 4, 5

    2024

  17. [17]

    Visual-linguistic representation learning with deep cross-modality fusion for referring multi-object tracking

    Wenyan He, Yajun Jian, Yang Lu, and Hanzi Wang. Visual-linguistic representation learning with deep cross-modality fusion for referring multi-object tracking. InICASSP, 2024. 3

    2024

  18. [18]

    Multigranularity local- ization transformer with collaborative understanding for referring multiobject tracking.IEEE Transactions on Instrumentation and Measurement, 74:1–13, 2025

    Jiajun Chen, Jiacheng Lin, Guojin Zhong, You Yao, and Zhiyong Li. Multigranularity local- ization transformer with collaborative understanding for referring multiobject tracking.IEEE Transactions on Instrumentation and Measurement, 74:1–13, 2025. 3, 5, 9

    2025

  19. [19]

    Language decoupling with fine-grained knowledge guidance for referring multi-object tracking

    Guangyao Li, Siping Zhuang, Yajun Jian, Yan Yan, and Hanzi Wang. Language decoupling with fine-grained knowledge guidance for referring multi-object tracking. InICCV, 2025. 3

    2025

  20. [20]

    ReaMOT: A benchmark and framework for reasoning-based multi-object tracking.arXiv preprint arXiv:2505.20381, 2025

    Sijia Chen, Yanqiu Yu, En Yu, and Wenbing Tao. ReaMOT: A benchmark and framework for reasoning-based multi-object tracking.arXiv preprint arXiv:2505.20381, 2025. 3, 4 11

    Pith/arXiv arXiv 2025

  21. [21]

    ReferGPT: Towards zero-shot referring multi-object tracking

    Tzoulio Chamiti, Leandro Di Bella, Adrian Munteanu, and Nikos Deligiannis. ReferGPT: Towards zero-shot referring multi-object tracking. InCVPRW, 2025. 3

    2025

  22. [22]

    End-to-end referring video object segmentation with multimodal transformers

    Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. InCVPR, 2022. 3

    2022

  23. [23]

    Language as queries for referring video object segmentation

    Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. InCVPR, 2022. 3

    2022

  24. [24]

    OnlineRefer: A simple online baseline for referring video object segmentation

    Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. OnlineRefer: A simple online baseline for referring video object segmentation. InICCV, 2023. 3

    2023

  25. [25]

    Referred by multi-modality: A unified temporal transformer for video object segmentation

    Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. InAAAI, 2024. 3, 9

    2024

  26. [26]

    ReferDINO: Referring video object segmentation with visual grounding foundations

    Tianming Liang, Kun-Yu Lin, Chaolei Tan, Jianguo Zhang, Wei-Shi Zheng, and Jian-Fang Hu. ReferDINO: Referring video object segmentation with visual grounding foundations. InICCV,

  27. [27]

    Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I. Jordan. Learning transferable features with deep adaptation networks. InICML, 2015. 3

    2015

  28. [28]

    Domain-adversarial training of neural networks.Journal of Machine Learning Research, 17(59):1–35, 2016

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks.Journal of Machine Learning Research, 17(59):1–35, 2016. 3

    2016

  29. [29]

    Adversarial discriminative domain adaptation

    Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. InCVPR, 2017. 3

    2017

  30. [30]

    Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I. Jordan. Conditional adversarial domain adaptation. InNeurIPS, 2018. 3

    2018

  31. [31]

    Maximum classifier discrepancy for unsupervised domain adaptation

    Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. InCVPR, 2018. 3

    2018

  32. [32]

    Domain adaptive Faster R-CNN for object detection in the wild

    Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive Faster R-CNN for object detection in the wild. InCVPR, 2018. 3

    2018

  33. [33]

    Strong-weak distribution alignment for adaptive object detection

    Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, and Kate Saenko. Strong-weak distribution alignment for adaptive object detection. InCVPR, 2019. 3

    2019

  34. [34]

    Exploring object relation in mean teacher for cross-domain detection

    Qi Cai, Yingwei Pan, Chong-Wah Ngo, Xinmei Tian, Lingyu Duan, and Ting Yao. Exploring object relation in mean teacher for cross-domain detection. InCVPR, 2019. 3

    2019

  35. [35]

    Unbiased mean teacher for cross-domain object detection

    Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. InCVPR, 2021. 3

    2021

  36. [36]

    Cross-domain adaptive teacher for object detection

    Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. Cross-domain adaptive teacher for object detection. InCVPR,

  37. [37]

    Learning domain adaptive object detection with probabilistic teacher

    Meilin Chen, Weijie Chen, Shicai Yang, Jie Song, Xinchao Wang, Lei Zhang, Yunfeng Yan, Donglian Qi, Yueting Zhuang, Di Xie, et al. Learning domain adaptive object detection with probabilistic teacher. InICML, 2022. 3

    2022

  38. [38]

    UPRE: Zero- shot domain adaptation for object detection via unified prompt and representation enhancement

    Xiao Zhang, Fei Wei, Yong Wang, Wenda Zhao, Feiyi Li, and Xiangxiang Chu. UPRE: Zero- shot domain adaptation for object detection via unified prompt and representation enhancement. InICCV, 2025. 3

    2025

  39. [39]

    Temporal attentive alignment for large-scale video domain adaptation

    Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal attentive alignment for large-scale video domain adaptation. InICCV, 2019. 3

    2019

  40. [40]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. InICLR, 2021. 3 12

    2021

  41. [41]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. InCVPR, 2022. 3

    2022

  42. [42]

    Efficient test-time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. InICML, 2022. 3

    2022

  43. [43]

    A probabilistic framework for lifelong test-time adaptation

    Dhanajit Brahma and Piyush Rai. A probabilistic framework for lifelong test-time adaptation. InCVPR, 2023. 3

    2023

  44. [44]

    Synthetic- to-real video person Re-ID.IEEE Transactions on Information F orensics and Security, 20: 5438–5450, 2025

    Xiangqun Zhang, Ruize Han, Likai Wang, Linqi Song, Junhui Hou, and Wei Feng. Synthetic- to-real video person Re-ID.IEEE Transactions on Information F orensics and Security, 20: 5438–5450, 2025. 3

    2025

  45. [45]

    A new bench- mark and algorithm for clothes-changing video person re-identification.IEEE Transactions on Information F orensics and Security, 20:1993–2005, 2025

    Likai Wang, Xiangqun Zhang, Ruize Han, Yanjie Wei, Song Wang, and Wei Feng. A new bench- mark and algorithm for clothes-changing video person re-identification.IEEE Transactions on Information F orensics and Security, 20:1993–2005, 2025. 3

    1993

  46. [46]

    Are we ready for autonomous driving? the KITTI vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. InCVPR, 2012. 3

    2012

  47. [47]

    MOT16: A benchmark for multi-object tracking.arXiv preprint arXiv:1603.00831, 2016

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

    Pith/arXiv arXiv 2016

  48. [48]

    The Cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. InCVPR, 2016. 3

    2016

  49. [49]

    BDD100K: A diverse driving dataset for heterogeneous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. InCVPR, 2020. 3, 5

    2020

  50. [50]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. InCVPR, 2020. 3

    2020

  51. [51]

    Virtual worlds as proxy for multi-object tracking analysis

    Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. InCVPR, 2016. 3

    2016

  52. [52]

    Virtual KITTI 2.arXiv preprint arXiv:2001.10773, 2020

    Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual KITTI 2.arXiv preprint arXiv:2001.10773, 2020. 3, 4, 5

    Pith/arXiv arXiv 2001

  53. [53]

    German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M. Lopez. The SYNTHIA dataset: A large collection of synthetic images for semantic segmentation of urban scenes. InCVPR, 2016. 3

    2016

  54. [54]

    Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun

    Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. InECCV, 2016. 3

    2016

  55. [55]

    SHIFT: A synthetic driving dataset for continuous multi-task domain adaptation

    Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. SHIFT: A synthetic driving dataset for continuous multi-task domain adaptation. InCVPR, 2022. 3

    2022

  56. [56]

    VisDA: The visual domain adaptation challenge.arXiv preprint arXiv:1710.06924, 2017

    Xingchao Peng, Ben Usman, Neela Kaushik, Dequan Wang, Judy Hoffman, and Kate Saenko. VisDA: The visual domain adaptation challenge.arXiv preprint arXiv:1710.06924, 2017. 3

    Pith/arXiv arXiv 2017

  57. [57]

    Moment matching for multi-source domain adaptation

    Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. InICCV, 2019. 3

    2019

  58. [58]

    WILDS: A benchmark of in-the-wild distribution shifts

    Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Sara Beery, et al. WILDS: A benchmark of in-the-wild distribution shifts. InICML, 2021. 3

    2021

  59. [59]

    AerialMind: Towards referring multi-object tracking in UA V scenarios

    Chenglizhao Chen, Shaofeng Liang, Runwei Guan, Xiaolou Sun, Haocheng Zhao, Haiyun Jiang, Tao Huang, Henghui Ding, and Qing-Long Han. AerialMind: Towards referring multi-object tracking in UA V scenarios. InAAAI, 2026. 4 13 A Benchmark Construction and Protocols A.1 Alignment, Matching, and Residual Calibration Refer-vKITTI is constructed from the five Ref...

    2026