Pith. sign in

REVIEW 5 major objections 4 minor 74 references

YOLOv14:Unified Cross-Domain Real-Time Object Detectionwith Adaptive Multi-View Representation

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a single real-time detector, YOLOv14, can keep high accuracy on standard photos while also detecting objects in fisheye, panoramic, drone, and game-rendered imagery, using four adaptive mechanisms rather than…

desk verdict Plausible COCO gains and a clean integration, but the flagship cross-domain results are train/test distribution matches rather than evidence of robustness to real game-rendered or fisheye data. read the letter →

arxiv 2608.04720 v1 pith:E27RWFSJ submitted 2026-08-05 cs.CV

Jinling Jia , Jian Lu , Jone Yawl , Chenbin Zhang This is my paper
classification cs.CV
keywords real-timeobjectdetectioncross-domaindeformableattentiondomainadaptationfisheyedistortionpanoramicimagesdroneimagerygame-to-realtransfer
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

YOLOv14 is a real-time object detector built to survive inputs that break the usual pinhole-camera assumption: fisheye lenses, 360-degree panoramas, aerial drone views, and game-rendered characters. The paper argues that one network, not a collection of specialist models, can cover all of these domains at real-time speed. On COCO val2017 it reports 49.1 mAP at 2.91 ms on a T4 GPU, and on synthetic proxies for the four hard domains it reports gains of +4.1, +6.6, +6.4, and +26.1 mAP over strong baselines. The largest claimed gain, +26.1 mAP on game characters, comes from a Game2Real mechanism that aligns rendered and photographic feature distributions so game characters are detected as people.

What carries the argument

The central mechanism is Deformable Area-Attention (D-AAttn), which replaces rigid attention grids with a learned dense 2D deformation field that warps query, key, and value features before attention is computed. It is supported by Game2Real Domain Adaptation, which combines procedural stylization with Adaptive Instance Normalization (AdaIN), an operation that aligns feature means and variances, and an adversarial domain-confusion loss; by ViewEmbedding with a cross-view NT-Xent contrastive loss that pulls same-class features from different views closer; and by DynamicScaleRouter, a gating network that learns per-input feature pyramid weights. The theoretical engine is the HΔH-divergence bound, a measure of how much two domains differ as seen by the hypothesis class, used to argue that adversarial confusion plus AdaIN reduces expected target risk.

What would settle it

Take the trained YOLOv14s and evaluate it on held-out real data that was never synthesized by its own pipeline: raw frames captured from several commercial game engines not used in training, and photographs from a real fisheye lens with factory calibration, then compare against the reported 50.2 mAP and 45.3 mAP. If the game gain collapses toward the 24 mAP baseline or the fisheye gain disappears on the real lens, the cross-domain claims are an artifact of proxy matching.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-domain robustness can be engineered directly into a real-time YOLO-style detector without sacrificing speed. Four mechanisms carry the claim: Deformable Area-Attention, which warps the feature grid with a learned 2D offset field so attention samples adaptively under geometric distortion; Game2Real Domain Adaptation, a three-tier mix of procedural game stylization, AdaIN feature alignment, and adversarial domain confusion that makes game characters detectable as real humans; Multi-View Conditioning with a cross-view contrastive loss for viewpoint invariance; and an Adaptive Augmentation Policy plus DynamicScaleRouter that routes each input to suitable augmentations and re-weights feature pyramid scales per image. Together these are claimed to give 49.1 mAP at 2.91 ms on COCO and consistent gains across fisheye, panorama, drone, and game benchmarks.

Load-bearing premise

The load-bearing premise is that the four benchmark proxies—COCO images passed through the same procedural game stylization used in training, COCO with simulated barrel distortion, SUN360 panoramas, and VisDrone footage—actually represent the real target domains; if a real game render or a real fisheye photo does not look like the training transform, the headline cross-domain gains, especially +26.1 mAP on game characters, could be an artifact of train/test distribution match.

Editorial extensions

If this is right

  • If the claims hold, a single model can replace separate fisheye, aerial, panoramic, and game detectors, simplifying deployment in vehicles, drones, and content-moderation pipelines.
  • Game-character detection as person means game analytics and moderation could reuse a photographic person detector without retraining on each game engine.
  • The per-input DynamicScaleRouter is consistent with the reported concentration of gains on small objects, where small-object AP rises by +1.9 mAP (+6.7%).
  • The architecture adds roughly 27% FLOPs over the YOLOv12s baseline but only 0.49 ms of latency, keeping the robustness features inside a real-time budget on a T4 GPU.
  • The theoretical domain-adaptation bound makes a quantitative prediction: lowering the adversarial loss and strengthening AdaIN alignment should lower target-domain error.

Reading between the lines

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

  • If the synthetic proxies generalize to real game renders and real lenses, the same procedural stylization recipe could extend to other rendering styles, such as cel-shading or ray tracing, and to simulation-to-real transfer in robotics.
  • The scene-routing and per-input scale weighting imply an adaptive-compute strategy, where scene-type estimation could budget latency dynamically across a video stream.
  • A direct test of the Game2Real claim would be detection on a game engine never seen during training; the paper's own limitation section concedes that real game pipelines are not fully covered by the stylization.
  • The reported COCO gains from the deformable modules suggest these components could be evaluated independently as plug-in improvements for other real-time detectors, not only inside YOLOv14.
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

5 major / 4 minor

Summary. The paper proposes YOLOv14, a real-time object detector that extends YOLOv12 with four mechanisms: Deformable Area-Attention (D-AAttn), Game2Real domain adaptation via AdaIN and adversarial confusion, Multi-View Conditioning with a cross-view contrastive loss, and an Adaptive Augmentation Policy with DynamicScaleRouter. It reports 49.1 mAP on COCO val2017 at 2.91 ms on a T4 GPU and claims large cross-domain gains on fisheye (+4.1), panorama (+6.6), drone (+6.4), and game-character (+26.1) benchmarks. The paper also presents theoretical bounds in Section 3.8 and Appendix B, an ablation study in Table 3, and comparisons with 23 prior detectors.

Significance. If the cross-domain results held on genuine target-domain data, the paper would be a valuable step toward a single real-time detector for nonstandard inputs. The COCO result is plausible as an incremental architecture contribution, and the paper deserves credit for reporting error bars, a clear inference protocol, per-component latency overhead, and a limitations section. However, the two largest cross-domain gains are measured on benchmarks constructed from the same augmentations used during training, so they do not establish generalization to real game-rendered or real lens-distorted images. The theoretical proofs contain uncontrolled steps, and the central ablation table is internally inconsistent. These issues directly affect the paper's headline claims.

major comments (5)
  1. [Section 3.3, Table 9, Table 4] The Game benchmark (Delta-Syn) is generated by applying 'game stylization' to COCO images, and Section 3.3 trains with GameCharacterStylization using the same posterization (bit depth 3–6), unsharp masking, saturation boost (×1.5–1.8), and contrast adjustment (α∈[1.2,1.8]). The +26.1 mAP gain on this benchmark therefore reflects a train/test distribution match, not evidence that YOLOv14 detects real game-rendered characters. Section 5 itself acknowledges that real game pipelines (deferred shading, ray tracing, cel-shading) are 'not fully covered' by the stylization. Please evaluate on real game-rendered data or explicitly reframe the claim as robustness to the specific augmentation distribution.
  2. [Appendix A.3, Section 5] The fisheye benchmark is COCO with simulated barrel distortion k=0.15, and the limitations paragraph says training uses simulated fisheye transforms. Without evaluation on a real lens dataset such as WoodScape, the +4.1 mAP gain may reflect matching the synthetic distortion model rather than true geometric robustness. A real-world fisheye evaluation is needed to support the abstract's claim of robustness to fisheye distortion.
  3. [Table 3, Section 4.3] The ablation table is not interpretable as written. If the rows are cumulative, the Game column goes 24.1 → 35.8 → 42.3 → 49.6 → 25.0 after adding DeformableA2C2f, then to 50.2 for the full model; if the rows are individual additions to the baseline, the text's claim of a cumulative Game2Real effect (+11.7, +6.5, +7.3) is not supported by the table. Please provide a clearly labeled cumulative ablation with confidence intervals and explain the nonmonotonic Game column.
  4. [Section 3.8, Appendix B] The proof of Theorem 3.1 is not mathematically sound as presented. In Step 2, the equality JSD = 2 L*_adv is asserted without defining L*_adv relative to the cross-entropy minimax objective in Eq. (8). In Step 3, the stated constants are inconsistent: with JSD ≤ 2 L*_adv, Pinsker gives δ ≤ sqrt(L*_adv), hence d_HΔH ≤ 2 sqrt(L*_adv), not 2 sqrt(2 L*_adv). In Step 4, 'AdaIN reduces the effective distribution diameter from R to R/||γ||' is undefined because γ is a learned vector and no norm or distributional diameter is specified. Proposition 3.1's proof also contains an unexplained area term and concludes α>0 from |∂T|/|T| without showing the first term dominates. These theoretical claims should be corrected or removed.
  5. [NeurIPS Checklist items 5 and 13, Section 4.1] The paper states that code and weights are not released. Since the main cross-domain results rest on synthetic benchmarks and many nonstandard components, the absence of code or a detailed reproducibility package prevents independent verification of the COCO result and of the claimed cross-domain gains. Please provide code, weights, or at minimum a complete implementation specification and evaluation scripts.
minor comments (4)
  1. [Abstract] The phrase 'allowing game characters are detected as real humans' is grammatically incorrect; change 'are' to 'to be'.
  2. [Checklist item 7] The checklist says 'Table 1 and Table 2 report ±standard deviation over 3 runs,' but Table 1 is a complexity table without standard deviations; the relevant tables are Tables 2, 3, and 4.
  3. [Table 3 caption] The caption 'Each row adds a component to the YOLOv12s baseline' conflicts with the apparent cumulative interpretation in the text; clarify the experimental protocol.
  4. [Section 4.5] The text states YOLOv14s achieves '>340 FPS' on T4; given the reported 2.91 ms latency, the FPS value is 344, which is fine, but the reader must reconcile this with the Table 2 entry of 344 FPS. Please keep the units consistent.

Circularity Check

1 steps flagged · score 7.0 of 10

The +26.1 mAP Game2Real gain is measured on COCO images passed through the same GameCharacterStylization used as training augmentation, so the flagship cross-domain result is a train/test distribution match rather than evidence of adaptation to real game-rendered frames.

  1. self definitional [Section 3.3 (Data-level augmentation); Appendix A.3, Table 9 (Game benchmark); Section 4.2 (main cross-domain result)]
    ""Data-level augmentation. GameCharacterStylization applies game-engine rendering effects to training images: posterization (bit depth b∈[3,6]), unsharp masking, saturation boost (×1.5–1.8), and contrast adjustment (α∈[1.2,1.8])." ... "Game (Delta-Syn) 2,000 80 COCO + game stylization Bbox" ... "The largest gain is on the Game benchmark: 50.2 mAP vs. 24.1 mAP (YOLOv12s), a +26.1 mAP (+108%) improvement driven by the three-tier Game2Real mechanism (Section 3.3).""

    The Game benchmark is defined as COCO images plus 'game stylization', and Section 3.3's training-time GameCharacterStylization is the same procedural transform with the same parameter ranges (posterization bit depth 3–6, saturation boost ×1.5–1.8, contrast α∈[1.2,1.8]). Train and test images are therefore generated by the same distribution, so the +26.1 mAP gain measures the model's fit to its own augmentation distribution, not adaptation to a distinct game-rendered domain. The paper itself concedes that 'Real games employ diverse pipelines (deferred shading, ray tracing, cel-shading) not fully covered by our stylization', confirming the synthetic benchmark does not represent the claimed target domain. The headline cross-domain result is thus guaranteed partly by construction.

full rationale

The strongest circularity is the Game benchmark. Table 9 defines Game (Delta-Syn) as COCO images with 'game stylization', and Section 3.3's GameCharacterStylization training augmentation applies the same procedural effects with the same parameter ranges, so the +26.1 mAP game-character gain is a train/test distribution match rather than evidence of robustness to real game-rendered frames. This is the paper's flagship cross-domain result and the main support for the Game2Real claim. The COCO val2017 core result (49.1 mAP), the SUN360 panorama benchmark, and the VisDrone drone benchmark are real, independent evaluations and are not circular, which prevents a higher score. The fisheye benchmark is also synthetic (COCO plus simulated barrel distortion k=0.15) and no real lens dataset such as WoodScape is used, so the +4.1 mAP fisheye claim shares the same external-validity weakness; however, the text does not explicitly tie the training transform to the exact same simulation, so I do not count it as a construction-level circular step. No load-bearing self-citation chain was found, and the theoretical bounds are standard domain-adaptation results rather than self-citations. Overall, the paper is partial rather than fully circular: one of the four headline cross-domain predictions reduces by construction, while the backbone result and two real-data benchmarks remain independent.

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

No new physical entities are introduced. The uncharged inputs are free hyperparameters that shape the synthetic benchmarks, plus several unproved assumptions in the theoretical analysis. The most consequential is the use of the same procedural stylization for training augmentation and for the Game test set.

free parameters (4)
  • lambda_1, lambda_2 (loss weights) = 0.1, 0.05
    Weights on adversarial and cross-view losses in Eq. 17; chosen by hand, no sweep reported.
  • GameCharacterStylization ranges and probability = bit depth 3-6, saturation x1.5-1.8, contrast 1.2-1.8, probability 0.3
    These ranges define both the training augmentation and the Game test benchmark, so they directly set the difficulty of the claimed game-to-real transfer.
  • Fisheye distortion k = 0.15
    The COCO-Fisheye test benchmark is generated with simulated barrel distortion k=0.15; no real fisheye data is used.
  • View taxonomy and contrastive temperature tau = six views, tau not specified
    Multi-view conditioning depends on the hand-defined view set and temperature; no sensitivity analysis is given.
assumptions (5)
  • standard math Ben-David et al. domain adaptation bound: epsilon_T(h) <= epsilon_S(h) + 1/2 d_HDeltaH + lambda
    Inherited as the starting point of Theorem 3.1 without local re-derivation.
  • ad hoc to paper At the minimax equilibrium of the adversarial loss, JSD <= 2 L*_adv and d_HDeltaH <= 2 sqrt(2 L*_adv)
    Invoked in Appendix B.2; the factor and the JSD bound are asserted, not proven.
  • ad hoc to paper AdaIN reduces the effective feature distribution diameter from R to R/||gamma||
    Used to claim the bound improves by 1 - ||gamma||^-1/2; no proof or experiment supports it.
  • domain assumption The deformable offset network learns the exact inverse distortion Delta = -delta_distort + epsilon with ||epsilon|| <= epsilon for fisheye
    Appendix B.1 assumes the learned offsets are the inverse of the radial distortion, which is the conclusion the architecture would need to establish.
  • ad hoc to paper Cross-view NT-Xent loss bounds the HDeltaH distance as d <= (2/rho) E||zi - zj||
    Proposition 3.2 is stated without proof and assumes a Lipschitz constant rho relating feature distance to domain divergence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of YOLOv14:Unified Cross-Domain Real-Time Object Detectionwith Adaptive Multi-View Representation." pith.science (2026). https://pith.science/paper/E27RWFSJ

@misc{pith2026260804720,
  author       = {Pith},
  title        = {Pith review of: YOLOv14:Unified Cross-Domain Real-Time Object Detectionwith Adaptive Multi-View Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E27RWFSJ}},
  note         = {Machine review of arXiv:2608.04720}
}
read the original abstract

Real-time object detectors achieve remarkable accuracy under controlled conditions, yet degrade sharply on non-ideal inputs: fisheye distortion, game-rendered characters, aerial viewpoints, and 360{\deg} panoramas. We present YOLOv14, aunified detection framework addressing these challenges through four synergisticinnovations. (1) Deformable Area-Attention (D-AAttn) replaces rigid attentiongrids with learned 2D deformation fields, enabling adaptive sampling under geometric distortion. (2) Game2Real Domain Adaptation aligns rendered-game and photographic feature distributions via Adaptive Instance Normalization (AdaIN)and adversarial domain confusion, allowing game characters are detected as realhumans. (3) Multi-View Conditioning injects learned viewpoint embeddings intothe backbone with a cross-view contrastive loss that pulls same-class features fromdifferent perspectives closer. (4) An Adaptive Augmentation Policy automaticallyclassifies each input' scene type and routes to optimal augmentations, while a DynamicScaleRouter learns per-input feature pyramid weights. Together, YOLOv14achieves 49.1 mAP on COCO val2017 at 2.91 ms (T4 GPU), and delivers substantial gains on fisheye (+4.1 mAP), panorama (+6.6 mAP), drone (+6.4 mAP), andgame-character (+26.1 mAP) benchmarks

Figures

Figures reproduced from arXiv: 2608.04720 by the authors.

Figure 1
Figure 1. YOLOv14 system pipeline. The architecture comprises six stages: scene analysis, adaptive augmentation (training-only), Game2Real domain adaptation via AdaIN and adversarial loss, multi-view conditioning through ViewEmbedding, deformable feature pyramid with DynamicScaleRouter, and multi-scale detection heads. addresses surround-view perception. YOLOv14 handles diverse inputs through adaptive conditioning modules wit… view at source ↗
Figure 2
Figure 2. Qualitative detection results across challenging scenarios. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Latency-accuracy trade-off on COCO val2017. YOLOv14 [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 42 canonical work pages

  1. [1]

    Analysis of representations for domain adaptation.Advances in neural information processing systems, 19, 2006

    Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation.Advances in neural information processing systems, 19, 2006. 11

    work page 2006

  2. [2]

    A theory of learning from different domains.Machine learning, 79(1): 151–175, 2010

    Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jen- nifer Wortman Vaughan. A theory of learning from different domains.Machine learning, 79(1): 151–175, 2010

    work page 2010

  3. [3]

    Yolov4: Optimal speed and accuracy of object detection.arXiv preprint arXiv:2004.10934, 2020

    Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection.arXiv preprint arXiv:2004.10934, 2020

    arXiv 2004

  4. [4]

    Visdrone-det2021: The vision meets drone object detection challenge results

    Yaru Cao, Zhijian He, Lujia Wang, Wenguan Wang, Yixuan Yuan, Dingwen Zhang, Jinglin Zhang, Pengfei Zhu, Luc Van Gool, Junwei Han, et al. Visdrone-det2021: The vision meets drone object detection challenge results. InProceedings of the IEEE/CVF International confer- ence on computer vision, pages 2847–2854, 2021

    2021

  5. [5]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean conference on computer vision, pages 213–229. Springer, 2020

    2020

  6. [6]

    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. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3339–3348, 2018

    work page 2018

  7. [7]

    Yolo-ms: Rethinking multi-scale representation learning for real-time object detection

    Yuming Chen, Xinbin Yuan, Jiabao Wang, Ruiqi Wu, Xiang Li, Qibin Hou, and Ming-Ming Cheng. Yolo-ms: Rethinking multi-scale representation learning for real-time object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(6):4240–4252, 2025

    work page 2025

  8. [8]

    Domain adaptation in regression

    Corinna Cortes and Mehryar Mohri. Domain adaptation in regression. InInternational conference on algorithmic learning theory, pages 308–323. Springer, 2011

    work page 2011

Show all 74 references
  1. [9]

    Autoaugment: Learning augmentation strategies from data

    Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113–123, 2019

    2019

  2. [10]

    Randaugment: Practical automated data augmentation with a reduced search space

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702–703, 2020

    2020

  3. [11]

    Deformable convolutional networks

    Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. InProceedings of the IEEE international conference on computer vision, pages 764–773, 2017

    2017

  4. [12]

    Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation

    Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. InProceedings of the European conference on computer vision (ECCV), pages 447–463, 2018

    2018

  5. [13]

    Object detection in aerial images: A large-scale benchmark and challenges.IEEE transactions on pattern analysis and machine intelligence, 44 (11):7778–7796, 2021

    Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges.IEEE transactions on pattern analysis and machine intelligence, ...

    2021

  6. [14]

    Cswin transformer: A general vision transformer backbone with cross-shaped windows

    Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pag...

    2022

  7. [15]

    Domain-adversarial training of neural networks

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1–35, 2016

    2016

  8. [16]

    Ota: Optimal transport assignment for object detection

    Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 303–312, 2021

    2021

  9. [17]

    Yolox: Exceeding yolo series in 2021.arXiv preprint arXiv:2107.08430, 2021

    Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021.arXiv preprint arXiv:2107.08430, 2021. 12

    2021 arXiv

  10. [18]

    Siou loss: More powerful learning for bounding box regression.arXiv preprint arXiv:2205.12740, 2022

    Zhora Gevorgyan. Siou loss: More powerful learning for bounding box regression.arXiv preprint arXiv:2205.12740, 2022

    2022 arXiv

  11. [19]

    Arbitrary style transfer in real-time with adaptive instance normalization

    Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. InProceedings of the IEEE international conference on computer vision, pages 1501–1510, 2017

    2017

  12. [20]

    Ultralytics yolo26: Unified real-time end-to-end vision models, 2026

    Glenn Jocher, Jing Qiu, Mengyu Liu, Shuai Lyu, Fatih Cagatay Akyon, and Muhammet Esat Kalfaoglu. Ultralytics yolo26: Unified real-time end-to-end vision models, 2026. URL https://arxiv.org/abs/2606.03748

    2026 arXiv

  13. [21]

    Analyzing and improving the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110–8119, 2020

    2020

  14. [22]

    Omnidet: Surround view cameras based multi-task visual perception network for autonomous driving.IEEE Robotics and Automation Letters, 6 (2):2830–2837, 2021

    Varun Ravi Kumar, Senthil Yogamani, Hazem Rashed, Ganesh Sitsu, Christian Witt, Isabelle Leang, Stefan Milz, and Patrick Mäder. Omnidet: Surround view cameras based multi-task visual perception network for autonomous driving.IEEE Robotics and Automation Letters, 6 (2):2830–2837, 2021

    2021

  15. [23]

    Yolov13: Real-time object detection with hypergraph-enhanced adaptive visual perception.arXiv preprint arXiv:2506.17733, 2025

    Mengqi Lei, Siqi Li, Yihong Wu, Han Hu, You Zhou, Xinhu Zheng, Guiguang Ding, Shaoyi Du, Zongze Wu, and Yue Gao. Yolov13: Real-time object detection with hypergraph-enhanced adaptive visual perception.arXiv preprint arXiv:2506.17733, 2025

    2025 arXiv

  16. [24]

    Yolov6 v3

    Chuyi Li, Lulu Li, Yifei Geng, Hongliang Jiang, Meng Cheng, Bo Zhang, Zaidan Ke, Xiaoming Xu, and Xiangxiang Chu. Yolov6 v3. 0: A full-scale reloading.arXiv preprint arXiv:2301.05586, 2023

    2023 arXiv

  17. [25]

    Deep domain adaptive object detection: A survey

    Wanyi Li, Fuyu Li, Yongkang Luo, Peng Wang, et al. Deep domain adaptive object detection: A survey. In2020 IEEE Symposium Series on Computational Intelligence (SSCI), pages 1808–1813. IEEE, 2020

    2020

  18. [26]

    Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection.Advances in neural information processing systems, 33:21002–21012, 2020

    Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection.Advances in neural information processing systems, 33:21002–21012, 2020

    2020

  19. [27]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean conference on computer vision, pages 740–755. Springer, 2014

    2014

  20. [28]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2117–2125, 2017

    2017

  21. [29]

    Ssd: Single shot multibox detector.arXiv preprint arXiv:1512.02325, 2015

    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector.arXiv preprint arXiv:1512.02325, 2015

    2015 arXiv

  22. [30]

    Conditional adversarial domain adaptation.Advances in neural information processing systems, 31, 2018

    Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael Jordan. Conditional adversarial domain adaptation.Advances in neural information processing systems, 31, 2018

    2018

  23. [31]

    Rt-detrv2: Improved baseline with bag-of-freebies for real-time detection transformer.arXiv preprint arXiv:2407.17140, 2024

    Wenyu Lv, Yian Zhao, Qinyao Chang, Kui Huang, Guanzhong Wang, and Yi Liu. Rt-detrv2: Improved baseline with bag-of-freebies for real-time detection transformer.arXiv preprint arXiv:2407.17140, 2024

    2024 arXiv

  24. [32]

    Domain adaptation: Learning bounds and algorithms.arXiv preprint arXiv:0902.3430, 2009

    Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. Domain adaptation: Learning bounds and algorithms.arXiv preprint arXiv:0902.3430, 2009

    2009 arXiv

  25. [33]

    MIT press, 2018

    Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar.Foundations of machine learning. MIT press, 2018

    2018

  26. [34]

    Trivialaugment: Tuning-free yet state-of-the-art data augmentation

    Samuel G Müller and Frank Hutter. Trivialaugment: Tuning-free yet state-of-the-art data augmentation. InProceedings of the IEEE/CVF international conference on computer vision, pages 774–782, 2021. 13

    2021

  27. [35]

    Yolov3: An incremental improvement.arXiv preprint arXiv:1804.02767, 2018

    Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement.arXiv preprint arXiv:1804.02767, 2018

    2018 arXiv

  28. [36]

    You only look once: Unified, real-time object detection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 779–788, 2016

    2016

  29. [37]

    Generalized intersection over union: A metric and a loss for bounding box regression

    Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658–666, 2019

    2019

  30. [38]

    Maximum classifier discrepancy for unsupervised domain adaptation

    Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3723–3732, 2018

    2018

  31. [39]

    Efficientdet: Scalable and efficient object detec- tion

    Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detec- tion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781–10790, 2020

    2020

  32. [40]

    Yolov12: Attention-centric real-time object detectors.Advances in neural information processing systems, 38:78433–78457, 2026

    Yunjie Tian, Qixiang Ye, and David Doermann. Yolov12: Attention-centric real-time object detectors.Advances in neural information processing systems, 38:78433–78457, 2026

    2026

  33. [41]

    Fcos: Fully convolutional one-stage object detection

    Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. InProceedings of the IEEE/CVF international conference on computer vision, pages 9627–9636, 2019

    2019

  34. [42]

    Yolov10: Real-time end-to-end object detection.Advances in neural information processing systems, 37:107984–108011, 2024

    Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Yolov10: Real-time end-to-end object detection.Advances in neural information processing systems, 37:107984–108011, 2024

    2024

  35. [43]

    Gold-yolo: Efficient object detector via gather-and-distribute mechanism.Advances in neural information processing systems, 36:51094–51112, 2023

    Chengcheng Wang, Wei He, Ying Nie, Jianyuan Guo, Chuanjian Liu, Yunhe Wang, and Kai Han. Gold-yolo: Efficient object detector via gather-and-distribute mechanism.Advances in neural information processing systems, 36:51094–51112, 2023

    2023

  36. [44]

    Yolov7: Trainable bag-of- freebies sets new state-of-the-art for real-time object detectors

    Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of- freebies sets new state-of-the-art for real-time object detectors. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7464–7475, 2023

    2023

  37. [45]

    Yolov9: Learning what you want to learn using programmable gradient information

    Chien-Yao Wang, I-Hau Yeh, and Hong-Yuan Mark Liao. Yolov9: Learning what you want to learn using programmable gradient information. InEuropean conference on computer vision, pages 1–21. Springer, 2024

    2024

  38. [46]

    Exploring dcn-like architecture for fast image generation with arbitrary resolution.Advances in Neural Information Processing Systems, 37:87959–87977, 2024

    Shuai Wang, Zexian Li, Tianhui Song, Xubin Li, Tiezheng Ge, Bo Zheng, and Limin Wang. Exploring dcn-like architecture for fast image generation with arbitrary resolution.Advances in Neural Information Processing Systems, 37:87959–87977, 2024

    2024

  39. [47]

    Vision transformer with deformable attention

    Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794–4803, 2022

    2022

  40. [48]

    Recognizing scene viewpoint using panoramic place representation

    Jianxiong Xiao, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Recognizing scene viewpoint using panoramic place representation. In2012 IEEE conference on computer vision and pattern recognition, pages 2695–2702. IEEE, 2012

    2012

  41. [49]

    Pp-yoloe: An evolved version of yolo

    Shangliang Xu, Xinxin Wang, Wenyu Lv, Qinyao Chang, Cheng Cui, Kaipeng Deng, Guanzhong Wang, Qingqing Dang, Shengyu Wei, Yuning Du, et al. Pp-yoloe: An evolved version of yolo. arXiv preprint arXiv:2203.16250, 2022

    2022 arXiv

  42. [50]

    Woodscape: A multi-task, multi-camera fisheye dataset for autonomous driving

    Senthil Yogamani, Ciarán Hughes, Jonathan Horgan, Ganesh Sistu, Sumanth Chennupati, Michal Uricar, Stefan Milz, Martin Simon, Karl Amende, Christian Witt, et al. Woodscape: A multi-task, multi-camera fisheye dataset for autonomous driving. In2019 IEEE/CVF Interna- tional Confe...

    2019

  43. [51]

    Cutmix: Regularization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023–6032, 2019

    2019

  44. [52]

    Dino: Detr with improved denoising anchor boxes for end-to-end object detection

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In The eleventh international conference on learning representations, 2022

    2022

  45. [53]

    Varifocalnet: An iou-aware dense object detector

    Haoyang Zhang, Ying Wang, Feras Dayoub, and Niko Sunderhauf. Varifocalnet: An iou-aware dense object detector. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8514–8523, 2021

    2021

  46. [54]

    mixup: Beyond empirical risk minimization.arXiv preprint arXiv:1710.09412, 2017

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization.arXiv preprint arXiv:1710.09412, 2017

    2017 arXiv

  47. [55]

    Remote sensing object detection meets deep learning: A metareview of challenges and advances.IEEE Geoscience and Remote Sensing Magazine, 11(4):8–44, 2023

    Xiangrong Zhang, Tianyang Zhang, Guanchun Wang, Peng Zhu, Xu Tang, Xiuping Jia, and Licheng Jiao. Remote sensing object detection meets deep learning: A metareview of challenges and advances.IEEE Geoscience and Remote Sensing Magazine, 11(4):8–44, 2023

    2023

  48. [56]

    On learning invariant representations for domain adaptation

    Han Zhao, Remi Tachet Des Combes, Kun Zhang, and Geoffrey Gordon. On learning invariant representations for domain adaptation. InInternational conference on machine learning, pages 7523–7532. PMLR, 2019

    2019

  49. [57]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16965–16974, 2024

    2024

  50. [58]

    Distance-iou loss: Faster and better learning for bounding box regression

    Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. InProceedings of the AAAI conference on artificial intelligence, volume 34, pages 12993–13000, 2020

    2020

  51. [59]

    Domain generalization: A survey.IEEE transactions on pattern analysis and machine intelligence, 45(4):4396–4415, 2022

    Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey.IEEE transactions on pattern analysis and machine intelligence, 45(4):4396–4415, 2022

    2022

  52. [60]

    Unpaired image-to-image translation using cycle-consistent adversarial networks

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. InProceedings of the IEEE international conference on computer vision, pages 2223–2232, 2017

    2017

  53. [61]

    Deformable convnets v2: More deformable, better results

    Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308–9316, 2019

    2019

  54. [62]

    Deformable detr: Deformable transformers for end-to-end object detection.arXiv preprint arXiv:2010.04159, 2020

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection.arXiv preprint arXiv:2010.04159, 2020

    2010 arXiv

  55. [63]

    Learning data augmentation strategies for object detection

    Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. InEuropean conference on computer vision, pages 566–583. Springer, 2020. A Additional Experimental Details A.1 Training Hyperpara...

    2020

  56. [64]

    The abstract and Section 1 list five contributions; each is substantiated in Sections 3 and 4 with quantitative evidence

    Claims.Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] . The abstract and Section 1 list five contributions; each is substantiated in Sections 3 and 4 with quantitative evidence

  57. [65]

    See Section 5 (Conclusion) for a dedicated Limitations paragraph

    Limitations.Question: Does the paper discuss the limitations of the work performed by the authors? [Yes] . See Section 5 (Conclusion) for a dedicated Limitations paragraph

  58. [66]

    Propositions 1–2 and Theorem 1 state all assumptions; proof sketches appear in Section 3.8, full proofs in Appendix B

    Theory Assumptions and Proofs.Question: For each theoretical claim, does the paper provide the full set of assumptions and a complete proof? [Yes] . Propositions 1–2 and Theorem 1 state all assumptions; proof sketches appear in Section 3.8, full proofs in Appendix B

  59. [67]

    Training hyperparameters (Table 8), architecture details (Appendix A.2), and evaluation protocols (Section 4.5) are specified

    Experimental Result Reproducibility.Question: Does the paper fully disclose all the information needed to reproduce the main experimental results? [Yes] . Training hyperparameters (Table 8), architecture details (Appendix A.2), and evaluation protocols (Section 4.5) are specified

  60. [68]

    [TODO: Provide anonymous GitHub repo link for review.]

    Open access to data and code.Question: Does the paper provide open access to the data and code? [No] . [TODO: Provide anonymous GitHub repo link for review.]

  61. [69]

    See Section 4.1 (Implementation Details) and Table 8

    Experimental Setting Details.Question: Does the paper specify all the training and test details? [Yes] . See Section 4.1 (Implementation Details) and Table 8

  62. [70]

    Table 1 and Table 2 report±standard deviation over 3 runs

    Experiment Statistical Significance.Question: Does the paper report error bars suitably and correctly defined? [Yes] . Table 1 and Table 2 report±standard deviation over 3 runs

  63. [71]

    Section 4.1 specifies 4×A100 GPUs for training, single T4 GPU for inference

    Experiments Compute Resources.Question: For each experiment, does the paper provide sufficient information on the computer resources? [Yes] . Section 4.1 specifies 4×A100 GPUs for training, single T4 GPU for inference. 9.Code Of Ethics.Question: Does the research conform with ...

  64. [72]

    COCO [27] and all prior YOLO works are cited

    Licenses for existing assets.Question: Are the original sources of existing assets credited? [Yes] . COCO [27] and all prior YOLO works are cited. 13.New Assets.Question: Are new assets (code, data, models) released? [No] . [TODO: Release via GitHub and Hugging Face at camera-ready.]

  65. [73]

    No human subjects were involved

    Crowdsourcing and Research with Human Subjects.Question: For crowdsourcing or research with human subjects? [NA] . No human subjects were involved

  66. [74]

    No human subjects were involved

    Institutional Review Board (IRB) Approvals.Question: Did the authors obtain approval from an IRB? [NA] . No human subjects were involved. 18 Figure 3: Latency-accuracy trade-off on COCO val2017. YOLOv14(red stars) consistently outperforms prior detectors across all model scale...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.