Pith. sign in

REVIEW 2 major objections 7 minor 39 references

Why Domain Matters: Domain-Aware Benchmarking of Underwater Object Detection and Annotation Quality

T0 review · 2 major / 7 minor · reviewed 2026-07-14 · grok-4.5

Pith's one-line read Domain labels from appearance, scene, and camera geometry expose large, systematic gaps in both human annotation quality and detector performance that a single mAP number hides.

desk verdict Solid empirical paper that makes domain shift measurable for underwater detection and annotation; co-occurrence is the real soft spot, not a fatal one. read the letter →

arxiv 2607.10575 v1 pith:27MOQDS4 submitted 2026-07-12 cs.CV

classification cs.CV
keywords underwaterobjectdetectiondomainshiftannotationqualitydomain-awarebenchmarkingscenecompositionvisibilitymAPgaps
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

Underwater images change dramatically with water clarity, light, habitat clutter, object size, and camera angle, so a detector that looks strong on an average score can still fail in whole environments. The paper shows that the same physical factors also make human annotators miss far more objects in some conditions than in others. It introduces a simple labeling scheme that tags every image by visibility, illumination, color cast, object layout and scale, background complexity, and viewpoint. Using those tags on real datasets, the authors find that annotation correction rates and detector accuracy both swing by large, consistent margins across domain extremes. The practical point is that domain shift becomes something you can measure and act on: you can budget annotation effort, design stress-test splits, and report robustness per condition instead of trusting one aggregate number.

What carries the argument

An underwater domain labeling framework that scores each image on three axes—appearance (visibility, illumination, color), scene composition (layout, scale, background), and acquisition geometry (orientation, perspective)—and turns continuous metrics into categorical domain labels for domain-wise evaluation.

What would settle it

Re-label the same images with substantially different metrics or thresholds, or control for co-occurrence (for example, compare blue versus natural water only inside matched visibility bins); if the large annotation-correction and mAP gaps disappear or reverse, the claim that these domain labels isolate the drivers of difficulty is falsified.

Watch

Extended reading notes

Core claim

Physically meaningful domain labels assigned from image appearance, scene composition, and acquisition geometry reveal substantial domain-dependent discrepancies in both human annotation quality and deep-learning detector performance; these gaps are hidden by ordinary dataset-level mAP and are largely consistent across detector architectures.

Load-bearing premise

The chosen image statistics, object measures, depth cues, and fixed thresholds cleanly isolate the intended physical factors rather than being dominated by natural co-occurrences between them.

Editorial extensions

If this is right

  • Annotation budgets and quality-control effort should be allocated preferentially to small-object and crowded scenes, which show the highest human correction rates.
  • Detector reports should include domain-wise mAP and the gap between best and worst domain properties, not only a single dataset score.
  • Future collection and balancing can deliberately target under-represented or high-difficulty domains (low visibility, dark, nadir, sparse) rather than adding more of the same easy conditions.
  • Domain labels enable construction of real-condition domain-shift benchmarks that isolate specific underwater factors instead of relying on synthetic style transfer.
  • Noise-aware training and selective preprocessing can be guided by domain tags so that harder images receive more verification or enhancement before labeling.

Reading between the lines

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

  • The same labeling pipeline could be applied to instance segmentation or multi-species surveys to test whether the human-versus-detector difficulty mismatch generalizes beyond the four benthic classes studied here.
  • If co-occurrence is the real driver of some counter-intuitive results (blue water, complex backgrounds), continuous multi-factor models or causal graphs may be needed before domain tags are used for automated data selection.
  • Deployed marine monitoring systems could flag live images that fall into historically hard domains and trigger human review or on-site adaptation, turning the labels into an operational risk signal.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 7 minor

Summary. The paper proposes a three-axis domain labeling framework (image appearance, scene composition, acquisition geometry) that assigns categorical labels to underwater images from image statistics, object-level metrics, and monocular-depth cues. Using these labels, it presents the first systematic analysis of domain-dependent human annotation difficulty (via RUOD vs. re-annotated RUOD-R correction rates) and detector performance (YOLO26n, Faster R-CNN, RT-DETR trained on a mixed DUO/RUOD-R/UTDAC set). Non-parametric tests show large annotation correction-rate differences driven mainly by scale and layout; multi-architecture mAP gaps of ~0.20–0.33 mAP50 appear across visibility, color, scale, layout, background, and perspective. The authors argue that physically meaningful domain labels turn domain shift into a measurable evaluation dimension for annotation QC, data collection, and robustness benchmarking.

Significance. If the domain labels are accepted as useful (even if imperfectly isolating) descriptors, the work supplies a practical, data-centric complement to aggregate mAP and to synthetic style-transfer domain-generalization benchmarks. Strengths include: (i) paired original/revised annotations as a proxy for annotation difficulty with appropriate non-parametric tests (Kruskal–Wallis, BH-corrected Mann–Whitney, Spearman, bootstrap); (ii) consistent performance gaps across three detector families and five random splits; (iii) explicit discussion of counter-intuitive trends and co-occurrence; (iv) actionable recommendations for annotation budget, QC, and stress-testing; (v) planned public release of labels, splits, and analysis code. That combination is useful for underwater CV and marine monitoring even without a new detector architecture.

major comments (2)
  1. Section 5.2.2 and Limitations: several headline gaps (color blue vs natural ~0.33 mAP50; complex vs simple background; crowded vs sparse layout) are explicitly attributed to co-occurrence with visibility/illumination rather than the named factor alone. The central claim that labels “characterize” and isolate physically meaningful factors therefore rests on interpretation of confounded categories. Please add, at least for the counter-intuitive axes (Color, Layout, Background), a stratified or partial analysis (e.g., mAP within high-visibility only, or co-occurrence-conditioned gaps) in the main text so readers can judge residual effects after the dominant confounders. Without this, Table 3’s per-category gaps overstate factor-specific difficulty.
  2. Section 3 and free parameters: domain assignment depends on hand-chosen metrics and categorical thresholds deferred to the Supplementary Material, plus monocular depth for Orientation/Perspective. Because every statistical comparison in §§4–5 is conditioned on these bins, the main paper should report a brief sensitivity check (e.g., alternate thresholds or continuous rank correlations with the same metrics) showing that the large effects (scale/layout for annotation; visibility/scale/perspective for detection) are not artifacts of a single cut. A one-paragraph main-text summary plus pointer to full supp results would make the load-bearing premise auditable.
minor comments (7)
  1. Fig. 5: two nearly identical mAP panels appear stacked; clarify whether both are needed or merge into one multi-metric figure with a single legend.
  2. Table 1 caption: “boosted … by 294%” for Scallop is striking; consider also reporting absolute added counts or FN rate to avoid over-reading percentage inflation from a small base.
  3. Correction rate Ci = (Ai+Ri)/Oi uses IoU 0.5 matching; state briefly whether results are stable at IoU 0.3/0.7 (or point to supp).
  4. §5.1: “YOLO26n” and arXiv:2606.03748 are very recent; ensure naming and citation match the released model the community can reproduce.
  5. Fig. 4 color properties marked with *; the non-ordered treatment is good—apply the same visual cue consistently in Fig. 5 axis labels.
  6. Typos/style: “i.e. more scattering” (§5.2.2) needs comma after i.e.; “e.g. crowded” in Table 4 caption same; “top–bottom difference” vs “left–right” in Fig. 2 could use units or estimator name once.
  7. Related Work: synthetic S-URPC/S-UTDAC critique is fair; a short sentence on whether any real multi-site splits (beyond style transfer) exist would situate the contribution more tightly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: purely empirical measurement of annotation and detector gaps under independently computed domain labels.

full rationale

The paper defines domain labels from image statistics, object counts, and monocular-depth cues (Section 3), then measures human correction rates (Section 4) and detector mAP (Section 5) on those groups. Labels are never fitted to, or defined from, the performance numbers later reported; the two are independent. There are no equations equating a fitted parameter to a claimed prediction, no uniqueness theorems imported from the authors, and no ansatz smuggled via self-citation. Self-citations (e.g., prior dataset or disparity papers) are ordinary background and do not close any logical loop. Aggregate metrics are simply disaggregated by the proposed labels; the observed gaps are empirical observations, not tautologies. The work is therefore self-contained against external benchmarks and exhibits no circular reduction.

Assumptions & free parameters 2 free parameters · 3 assumptions · 1 invented entities

The central claims rest on a small set of domain assumptions about which image statistics proxy physical factors, plus a collection of free thresholds that discretize continuous metrics into categories. No new physical entities are postulated; the framework itself is an organizational invention rather than an ontological one.

free parameters (2)
  • visibility/illumination/color/layout/scale/background/orientation/perspective category thresholds
    All continuous metrics are binned into low/moderate/high (or equivalent) by thresholds whose exact values appear only in the Supplementary Material; different cut-offs would re-assign images and could change effect sizes.
  • IoU = 0.5 matching threshold for added/removed annotations
    Correction-rate definition treats an annotation as added or removed if it fails IoU 0.5 match; the choice is conventional but free and affects the numeric rates.
assumptions (3)
  • domain assumption Selected low-level image statistics (Tenengrad, Laplacian variance, color ratios, ORB density, monocular depth gradients, etc.) are faithful proxies for the intended physical domain factors.
    Invoked throughout Section 3; if the proxies are dominated by unrelated image properties the domain labels lose physical meaning.
  • domain assumption Differences between original RUOD and professionally re-annotated RUOD-R primarily reflect annotation difficulty rather than systematic protocol changes or new object definitions.
    Section 4 treats the correction rate as a pure difficulty proxy; any residual protocol shift would inflate or deflate the reported domain effects.
  • standard math Standard non-parametric tests (Kruskal–Wallis, Mann–Whitney with BH correction) and bootstrap resampling are sufficient to establish domain dependence.
    Used in Section 4.1–4.2; conventional and appropriate for the data distribution.
invented entities (1)
  • three-axis underwater domain labeling framework (appearance / scene composition / acquisition geometry)
    purpose: to assign every image a set of categorical domain labels that enable domain-wise evaluation
    The framework is the paper’s main methodological invention; it has no independent existence outside the proposed metrics and thresholds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Why Domain Matters: Domain-Aware Benchmarking of Underwater Object Detection and Annotation Quality." pith.science (2026). https://pith.science/paper/27MOQDS4

@misc{pith2026260710575,
  author       = {Pith},
  title        = {Pith review of: Why Domain Matters: Domain-Aware Benchmarking of Underwater Object Detection and Annotation Quality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/27MOQDS4}},
  note         = {Machine review of arXiv:2607.10575}
}
read the original abstract

Underwater object detection is strongly affected by domain shift, where performance can vary significantly across different locations, habitats, and deployment conditions. However, detector performance is typically evaluated using aggregate metrics that hide failures in specific environments, while existing domain generalization benchmarks often rely on synthetic variations that do not reflect real-world conditions. We introduce a framework that characterizes underwater images by appearance, scene composition, and acquisition geometry to assign domain labels. Using this framework, we perform the first systematic study of how domain factors influence both human annotation quality in underwater object detection datasets and deep learning-based detector performance, revealing substantial domain-dependent discrepancies. By incorporating physically meaningful domain labels, domain shift becomes something we can characterize, measure, benchmark, and act on. We highlight how this can be used to guide data collection and annotation, design more informative benchmarks, and assess detector robustness across diverse underwater environments.

Figures

Figures reproduced from arXiv: 2607.10575 by the authors.

Figure 1
Figure 1. Example underwater images from visually distinct un [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our underwater domain labeling framework. Images are characterized along three complementary axes: [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Boxplots of correction rates grouped by domain category. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Detection performance (mAP50) across domains and [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 4 linked inside Pith

  1. [1]

    Ali Awad, Ashraf Saleem, Yaman Aljnadi, Evan Lucas, Sidike Paheding, and Timothy C. Havens. RUOD-R: A high- fidelity re-annotated benchmark for underwater object detec- tion.IEEE Access, 14:60030–60046, 2026. 3, 4, 6

  2. [2]

    Underwater object detection in noisy imbalanced datasets.Pattern Recognition, 155:110649, 2024

    Long Chen, Tengyue Li, Andy Zhou, Shengke Wang, Junyu Dong, and Huiyu Zhou. Underwater object detection in noisy imbalanced datasets.Pattern Recognition, 155:110649, 2024. 2

  3. [3]

    Underwater optical object detection in the era of artificial intelligence: current, challenge, and future.ACM Computing Surveys, 58(3):1–34,

    Long Chen, Yuzhi Huang, Junyu Dong, Qi Xu, Sam Kwong, Huimin Lu, Huchuan Lu, and Chongyi Li. Underwater optical object detection in the era of artificial intelligence: current, challenge, and future.ACM Computing Surveys, 58(3):1–34,

  4. [4]

    Achieving domain generalization for underwater object detection by domain mixup and contrastive learning.Neurocomputing, 528:20–34,

    Yang Chen, Pinhao Song, Hong Liu, Linhui Dai, Xiaochuan Zhang, Runwei Ding, and Shengquan Li. Achieving domain generalization for underwater object detection by domain mixup and contrastive learning.Neurocomputing, 528:20–34,

  5. [5]

    Edge-guided representation learning for underwater object detection.CAAI Transactions on Intelli- gence Technology, 9(5):1078–1091, 2024

    Linhui Dai, Hong Liu, Pinhao Song, Hao Tang, Runwei Ding, and Shengquan Li. Edge-guided representation learning for underwater object detection.CAAI Transactions on Intelli- gence Technology, 9(5):1078–1091, 2024. 6

  6. [6]

    Training marine species object detectors with synthetic images and unsupervised domain adaptation.Frontiers in Marine Science, 12:1581778, 2025

    Heather Doig, Oscar Pizarro, and Stefan Williams. Training marine species object detectors with synthetic images and unsupervised domain adaptation.Frontiers in Marine Science, 12:1581778, 2025. 1

  7. [7]

    Comparison of image annotation data generated by multiple investigators for benthic ecology.Marine Ecology Progress Series, 552: 61–70, 2016

    Jennifer M Durden, Brian J Bett, Timm Schoening, Kirsty J Morris, Tim W Nattkemper, and Henry A Ruhl. Comparison of image annotation data generated by multiple investigators for benthic ecology.Marine Ecology Progress Series, 552: 61–70, 2016. 2

  8. [8]

    Advancing underwater vision: a survey of deep learning models for underwater object recognition and tracking.IEEE Access,

    Mahmoud Elmezain, Lyes Saad Saoud, Atif Sultan, Mohamed Heshmat, Lakmal Seneviratne, and Irfan Hussain. Advancing underwater vision: a survey of deep learning models for underwater object recognition and tracking.IEEE Access,

Show all 39 references
  1. [9]

    Rethinking general underwater object detection: Datasets, challenges, and solutions.Neurocomputing, 517:243–256, 2023

    Chenping Fu, Risheng Liu, Xin Fan, Puyang Chen, Hao Fu, Wanqi Yuan, Ming Zhu, and Zhongxuan Luo. Rethinking general underwater object detection: Datasets, challenges, and solutions.Neurocomputing, 517:243–256, 2023. 1, 4

  2. [10]

    See you somewhere in the ocean: few-shot domain adaptive underwa- ter object detection.Frontiers in Marine Science, 10:1151112,

    Lu Han, JiPing Zhai, Zhibin Yu, and Bing Zheng. See you somewhere in the ocean: few-shot domain adaptive underwa- ter object detection.Frontiers in Marine Science, 10:1151112,

  3. [11]

    Under- water species detection using channel sharpening attention

    Lihao Jiang, Yi Wang, Qi Jia, Shengwei Xu, Yu Liu, Xin Fan, Haojie Li, Risheng Liu, Xinwei Xue, and Ruili Wang. Under- water species detection using channel sharpening attention. InProceedings of the 29th ACM International Conference on Multimedia, pages 4259–4267, 2021. 1

  4. [12]

    Ultralytics yolo26: Unified real-time end-to-end vision models.arXiv preprint arXiv:2606.03748, 2026

    Glenn Jocher, Jing Qiu, Mengyu Liu, Shuai Lyu, Fatih Ca- gatay Akyon, and Muhammet Esat Kalfaoglu. Ultralytics yolo26: Unified real-time end-to-end vision models.arXiv preprint arXiv:2606.03748, 2026. 6

  5. [13]

    Flaws of imagenet, computer vision’s favourite dataset.arXiv preprint arXiv:2412.00076,

    Nikita Kisel, Illia V olkov, Katerina Hanzelkova, Klara Janouskova, and Jiri Matas. Flaws of imagenet, computer vision’s favourite dataset.arXiv preprint arXiv:2412.00076,

  6. [14]

    Estimating label quality and errors in semantic segmentation data via any model.arXiv preprint arXiv:2307.05080, 2023

    Vedang Lad and Jonas Mueller. Estimating label quality and errors in semantic segmentation data via any model.arXiv preprint arXiv:2307.05080, 2023. 3

  7. [15]

    A dataset and bench- mark of underwater object detection for robot picking

    Chongwei Liu, Haojie Li, Shuchang Wang, Ming Zhu, Dong Wang, Xin Fan, and Zhihui Wang. A dataset and bench- mark of underwater object detection for robot picking. In 2021 IEEE international conference on multimedia & expo workshops (ICMEW), pages 1–6, 2021. 1, 6

  8. [16]

    Towards domain generalization in underwater object detection

    Hong Liu, Pinhao Song, and Runwei Ding. Towards domain generalization in underwater object detection. In2020 IEEE international conference on image processing (ICIP), pages 1971–1975. IEEE, 2020. 2

  9. [17]

    Evan Lucas, Ali Awad, Anthony Geglio, Shadi Moradi, Ashraf Saleem, Timothy Havens, Angus Galloway, and Sidike Paheding. Underwater image enhancement and object detec- tion: are poor object detection results on enhanced images due to missing human labels? InProceedings of the I...

  10. [18]

    Physics-coupled fre- quency dynamic adaptation network for domain generalized underwater object detection

    Linxuan Luo, Pan Mu, and Cong Bai. Physics-coupled fre- quency dynamic adaptation network for domain generalized underwater object detection. InProceedings of the 33rd ACM International Conference on Multimedia, page 2284–2293, New York, NY , USA, 2025. Association for Computi...

  11. [19]

    A structured review of underwater object detection challenges and solutions: From traditional to large vision language models.arXiv preprint arXiv:2509.08490, 2025

    Edwine Nabahirwa, Wei Song, Minghua Zhang, Yi Fang, and Zhou Ni. A structured review of underwater object detection challenges and solutions: From traditional to large vision language models.arXiv preprint arXiv:2509.08490, 2025. 1

  12. [20]

    Domain generalization for sea cucumber detection: Tackling back- ground color variability in aquaculture settings.Aquaculture International, 33(5), 2025

    Fangqun Niu, Yifan Sheng, Junyi Wang, Xinyu Zheng, Kexin Liu, Yuanshan Lin, Wei Wang, and GuoDong Li. Domain generalization for sea cucumber detection: Tackling back- ground color variability in aquaculture settings.Aquaculture International, 33(5), 2025. 2

  13. [21]

    Human-visual- system-inspired underwater image quality measures.IEEE Journal of Oceanic Engineering, 41(3):541–551, 2015

    Karen Panetta, Chen Gao, and Sos Agaian. Human-visual- system-inspired underwater image quality measures.IEEE Journal of Oceanic Engineering, 41(3):541–551, 2015. 3

  14. [22]

    Detection of marine animals in a new underwater dataset with varying visibility

    Malte Pedersen, Joakim Bruslund Haurum, Rikke Gade, and Thomas B Moeslund. Detection of marine animals in a new underwater dataset with varying visibility. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 18–26, 2019. 1

  15. [23]

    Revisiting oxford and paris: Large-scale image retrieval benchmarking

    Filip Radenovi ´c, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ond ˇrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5706–5715, 2018. 3

  16. [24]

    AI-driven marine robotics: emerging trends in underwater perception and ecosystem monitoring

    Scarlett Raine and Tobias Fischer. AI-driven marine robotics: emerging trends in underwater perception and ecosystem monitoring. InAAAI, pages 40981–40989, 2026. 1, 2

  17. [25]

    Urchin- bot: An open-source model for the rapid detection and clas- sification of habitat-modifying sea urchin species.Marine Environmental Research, page 107662, 2025

    Kelham Rawlinson, Arie JP Spyksma, Kelsey I Miller, Ariell Friedman, Caitlin Grosvenor, Shahrokh Heidari, John P Keane, Nicholas Perkins, and Katerina Taskova. Urchin- bot: An open-source model for the rapid detection and clas- sification of habitat-modifying sea urchin specie...

  18. [26]

    Faster r-cnn: Towards real-time object detection with region proposal networks.IEEE transactions on pattern analysis and machine intelligence, 39(6):1137–1149, 2016

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks.IEEE transactions on pattern analysis and machine intelligence, 39(6):1137–1149, 2016. 6

  19. [27]

    Underwater complex environment domain adaptation for few- shot object detection based on transfer learning.Neurocom- puting, 666:132341, 2026

    Shouyu Ren, Hongchi Hao, Yuxiang Zhang, and Zhibin Yu. Underwater complex environment domain adaptation for few- shot object detection based on transfer learning.Neurocom- puting, 666:132341, 2026. 2

  20. [28]

    Real-time and resource-efficient multi-scale adaptive robotics vision for underwater object detection and domain generalization

    Lyes Saad Saoud, Zhenwei Niu, Lakmal Seneviratne, and Irfan Hussain. Real-time and resource-efficient multi-scale adaptive robotics vision for underwater object detection and domain generalization. InIEEE International Conference on Image Processing, pages 3917–3923, 2024. 2

  21. [29]

    Recomia-recommendations for marine image annotation: Lessons learned and future directions.Frontiers in Marine Science, 3:59, 2016

    Timm Schoening, Jonas Osterloff, and Tim W Nattkem- per. Recomia-recommendations for marine image annotation: Lessons learned and future directions.Frontiers in Marine Science, 3:59, 2016. 2

  22. [30]

    EFCWM-Mamba-YOLO: Real-time underwater object detection with adaptive feature representation and domain adaptation

    Pan Sun, Yu Lu, Shijie Shi, Meng Li, Qiang Li, and Huilin Ge. EFCWM-Mamba-YOLO: Real-time underwater object detection with adaptive feature representation and domain adaptation. InIEEE/RSJ International Conference on Intelli- gent Robots and Systems, pages 9614–9619, 2025. 2

  23. [31]

    Walker, Zheng Zeng, Chengchen L

    Joseph L. Walker, Zheng Zeng, Chengchen L. Wu, Jules S. Jaffe, Kaitlin E. Frasier, and Stuart S. Sandin. Underwater object detection under domain shift.IEEE Journal of Oceanic Engineering, 49(4):1209–1219, 2024. 1, 2, 3

  24. [32]

    An imaging-inspired no- reference underwater color image quality assessment metric

    Yan Wang, Na Li, Zongying Li, Zhaorui Gu, Haiyong Zheng, Bing Zheng, and Mengnan Sun. An imaging-inspired no- reference underwater color image quality assessment metric. Computers & Electrical Engineering, 70:904–913, 2018. 3

  25. [33]

    Learning with noisy labels revisited: A study using real-world human annotations

    Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. InInternational Conference on Learning Representations, 2022. 3

  26. [34]

    Junjie Wen, Guidong Yang, Benyun Zhao, Lei Lei, Zhi Gao, Xi Chen, and Ben M. Chen. Joint image enhancement for underwater object detection in various domains.IEEE Journal of Oceanic Engineering, 51(1):807–825, 2026. 2

  27. [35]

    Are all marine species created equal? performance disparities in underwater object detection

    Melanie Wille, Tobias Fischer, and Scarlett Raine. Are all marine species created equal? performance disparities in underwater object detection. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4556–4565, 2026. 2

  28. [36]

    Zhuoran Xie, Miao Yang, Mengjiao Shen, Yuquan Qiu, and Xinyu Wang. FIOD-VUE: Focusing on invariant information in object detection of varying underwater environment.IEEE Transactions on Circuits and Systems for Video Technology, 34(11):10743–10752, 2024. 2

  29. [37]

    An underwater color im- age quality evaluation metric.IEEE Transactions on Image Processing, 24(12):6062–6071, 2015

    Miao Yang and Arcot Sowmya. An underwater color im- age quality evaluation metric.IEEE Transactions on Image Processing, 24(12):6062–6071, 2015. 3

  30. [38]

    A reference-free underwater image quality assessment metric in frequency domain.Signal Processing: Image Communication, 94:116218, 2021

    Ning Yang, Qihang Zhong, Kun Li, Runmin Cong, Yao Zhao, and Sam Kwong. A reference-free underwater image quality assessment metric in frequency domain.Signal Processing: Image Communication, 94:116218, 2021. 3

  31. [39]

    DETRs beat YOLOs on real-time object detection

    Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. DETRs beat YOLOs on real-time object detection. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16965–16974, 2024. 6

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.