Pith. sign in

REVIEW 3 major objections 4 minor 65 references

YOLOv14: Adaptive Real-Time Object Detection for Diverse Imaging Conditions

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read YOLOv14 claims that a single real-time detector can keep high accuracy on fisheye, panorama, drone, and game-rendered inputs, reaching 49.1 mAP on COCO at 2.91 ms and adding +4.1 to +26.1 mAP on four cross-domain benchmarks.

desk verdict The COCO result is plausible but the cross-domain gains are train-on-test numbers, so the headline claims outrun the evidence. read the letter →

arxiv 2608.04720 v2 pith:E27RWFSJ submitted 2026-08-05 cs.CV

Jinling Jia , Jian Lu , Jone Yawl , Chenbin Zhang This is my paper · ORCID
classification cs.CV
keywords real-timeobjectdetectiondeformablearea-attentiondomainadaptationfisheyedistortionpanoramicimagingaerialdroneadaptiveaugmentationgame-to-realalignment
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

The paper sets out to show that a single real-time object detector can remain accurate when the input stops looking like an ideal pinhole photo, covering fisheye distortion, equirectangular panoramas, aerial drone views, and game-rendered imagery. It claims that four input-adaptive mechanisms, windowed deformable area-attention, multi-level game-to-real alignment, view-aware contrastive learning, and scene-adaptive augmentation routing, each target one source of variation and together add +4.1, +6.6, +6.4, and +26.1 mAP on the four constructed cross-domain benchmarks while keeping 49.1 mAP on COCO val2017 at 2.91 ms on a T4 GPU. A reader should care because deployed detectors in cars, drones, and AR/VR face exactly these non-ideal inputs, and the paper argues that adaptation can be built in rather than handled by separate specialist models. The evidence rests on synthetic benchmarks the paper constructs itself, which gives the claims a clear test protocol and also places the main risk on how well those synthetic transformations represent real-world conditions.

What carries the argument

The load-bearing mechanism is windowed Deformable Area-Attention (D-AAttn): a three-layer CNN predicts a dense two-channel offset field, and self-attention is computed inside local windows whose sampling points follow the predicted deformation, with shifted windows alternating across layers to allow cross-window information flow. A regularization term pulls the offsets toward an initialization estimated from synthetic fisheye transformations. The other three modules, Game2Real alignment (data-level stylization, AdaIN feature normalization, and a DANN-style adversarial loss), view-aware contrastive learning with adaptive temperature, and scene-adaptive augmentation routing with dynamic loss balancing, condition the same backbone on domain, viewpoint, and scene type. Together they implement the paper's hypothesis that distortion, domain shift, viewpoint, and scene diversity can be handled by input-adaptive sampling and normalization rather than by separate models per domain.

What would settle it

Run the released model on real fisheye data such as WoodScape, on real equirectangular panoramas, and on game frames from engines or titles not used for stylization-parameter estimation, measuring COCO-protocol mAP against YOLOv12s; if the reported advantages shrink to near zero or reverse, the synthetic benchmarks are not faithful proxies and the claimed generalization to real deployment conditions fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that input-adaptive feature sampling and normalization can remove the implicit pinhole-camera assumption in a modern YOLO detector. YOLOv14 is claimed to surpass YOLOv12s by +1.5 mAP on COCO val2017 and to beat both the standard detector and two unsupervised domain-adaptation baselines on every one of the four cross-domain benchmarks. The mechanism is a set of four complementary modules: D-AAttn (windowed deformable area-attention with shifted windows) for geometric distortion, multi-level Game2Real alignment (data stylization, AdaIN, and adversarial confusion with progressive training) for domain shift, view-aware contrastive learning with adaptive temperature for viewpoint invariance, and scene-adaptive augmentation routing with dynamic loss weighting for scene diversity. The paper reports the combination is super-additive, with ablations showing each module contributes most on its own domain, and the full model reaches 49.1 COCO mAP at 2.91 ms on T4.

Load-bearing premise

The synthetic transformations used to build the fisheye, panorama, and game benchmarks come from the same distribution families YOLOv14 trains on, so the claimed cross-domain gains depend on those synthetic transformations faithfully representing real fisheye, panorama, and game-rendered imagery.

Editorial extensions

If this is right

  • A single model can replace per-domain specialists: the same weights serve pinhole, fisheye, panorama, and drone views at 344 FPS on T4.
  • Source-domain augmentation that uses target-domain visual statistics can beat UDA baselines on the constructed benchmarks, offering a different trade-off for practitioners who have a few reference images of the deployment style.
  • The four components are complementary and super-additive: each ablation contributes most on its own target domain, while the full model improves on all domains at once.
  • The constructed benchmark suite (COCO-Fisheye, COCO-Panorama, Game-Stylized COCO, plus VisDrone) provides a repeatable evaluation protocol with explicit box-warping and panorama-projection steps, making cross-domain detection comparisons reproducible.
  • The measured overhead of the adaptive modules is 0.49 ms on T4, so the accuracy gains do not break real-time operation.

Reading between the lines

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

  • The +26.1 mAP gain on the game benchmark is likely an upper bound for real-world game frames: the stylization parameters are estimated from reference game images of the same style used in the synthetic test set, and the paper itself notes real-game performance is lower.
  • If the synthetic-to-real gap proves small, the same design pattern could extend to other non-ideal inputs such as thermal, low-light, or lidar-projected imagery by re-estimating the offset initialization and stylization statistics and adding a new entry to the scene-routing table.
  • The discrete six-class view embedding (pinhole, fisheye, panoramic, drone, BEV, ground) could be generalized to a continuous pose or distortion-strength embedding, which would handle mixed or intermediate distortions more gracefully.
  • The scene classifier's routing suggests a generic 'mixture of augmentations' recipe for dense prediction tasks, since the router selects per-input augmentation and loss weights rather than requiring a fixed global policy.
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

3 major / 4 minor

Summary. The paper proposes YOLOv14, an extension of YOLOv12 with four adaptive components: windowed deformable area-attention (D-AAttn), multi-level Game2Real alignment, view-aware contrastive learning with adaptive temperature, and scene-adaptive augmentation with dynamic loss balancing. The authors report 49.1 mAP on COCO val2017 at 2.91 ms on a T4 GPU and claim substantial gains on synthetic fisheye (+4.1 mAP), panorama (+6.6 mAP), drone (+6.4 mAP), and game-stylized (+26.1 mAP) benchmarks. The paper also releases code and models.

Significance. If the cross-domain results held on independent real-world data, the integrated framework would be a meaningful step toward real-time detection under non-ideal imaging conditions. The standard COCO result is plausible and consistent with incremental YOLO-style improvements, and the paper ships a code-release commitment that supports reproducibility. However, the headline cross-domain claims rest on self-generated synthetic benchmarks whose generating distributions are used during training, and the paper itself concedes in §5 that real-game performance is lower. Consequently, the claimed generalization to fisheye, panorama, and game domains is unverified, which limits the significance of the contributions as stated.

major comments (3)
  1. [§3.9.1, §3.9.2, §3.9.4; §3.2 Eq. (2); §3.3; §3.5] The headline cross-domain gains on COCO-Fisheye, COCO-Panorama, and Game-Stylized COCO are measured on benchmarks generated by the same transformation families used during training, making the numbers in Table 2 train-on-test-distribution results rather than evidence of generalization. The D-AAttn offset initializer Δ_init in Eq. (2) is estimated from synthetic fisheye transformations, the GameCharacterStylization parameters in §3.3 define both the training augmentation and the test renderer, and the scene classifier in §3.5 is trained on synthetic images from these families. The paper's own Limitations section (§5) concedes that real-game performance is lower, confirming that the +26.1 mAP Game result does not transfer. Independent real-world evaluation (e.g., WoodScape for fisheye, a genuine panorama dataset, and real game footage) is required to support the abstract's claim of substantial gains.
  2. [§4.1 vs. Appendix A.8] The description of the UDA baselines is self-contradictory. §4.1 states that MCD and SWDA are run without the Game2Real data-level stylization to keep them as pure UDA methods, while Appendix A.8 specifies that for Game dataset evaluation SWDA uses Game stylization as its strong augmentation. Because the Game column of Table 2 is the basis for the +26.1 mAP claim, this inconsistency either handicaps SWDA unfairly or gives YOLOv14s an unfair advantage. The protocol must be stated unambiguously and the comparison rerun under a single consistent set of conditions.
  3. [§3.9.2 and §A.3] The COCO-Panorama construction is not a physically valid projection of a single pinhole image into a 360° equirectangular panorama: a single 640×640 perspective image cannot define the content of a 1280×640 field of view, and the sampling procedure in A.3 effectively tiles or extrapolates unknown regions. This makes the +6.6 mAP panorama gain an artifact of the specific synthetic rendering rather than a measure of performance under real equirectangular capture. A real 360° dataset with ground-truth boxes, or a well-posed generation protocol that explicitly models the scene, is needed.
minor comments (4)
  1. [Table 1] The YOLOv14s row is misaligned: the latency and mAP values appear concatenated as “2.9149.1344,” which obscures the reported 2.91 ms and 49.1 mAP; the table should be reformatted.
  2. [§5] The Limitations section mentions the SUN360 dataset as a heuristic proxy, but SUN360 is never described or evaluated in the paper; either add the corresponding experiment or remove the mention.
  3. [§4.4] The per-module latency sum (0.58 ms) is larger than the observed full-model increase of 0.49 ms, and the 0.09 ms difference is attributed to kernel fusion without explanation; please clarify how kernel fusion accounts for the discrepancy.
  4. [§3.8 and Table 3] The ablation study reports each component added to the baseline but does not report leave-one-out ablations, so the claim that the components are complementary and yield super-additive gains is not directly supported by the presented data.

Circularity Check

2 steps flagged · score 6.0 of 10

Cross-domain gains on fisheye and game benchmarks are self-referential: the test generators are the same transformations used to fit D-AAttn offsets and to train Game2Real stylization; COCO val2017 and VisDrone results remain independent.

  1. fitted input called prediction [§3.2 (Eq. 2) and §3.9.1]
    "Loffset = λreg · 1/HW Σ∥∆i,j −∆init,i,j∥2², where ∆init is estimated from synthetic fisheye transformations and λreg = 0.01. ... We apply barrel distortion k ∈ {0.10,0.15,0.20} to COCO val2017. The transformation is rout = rin(1 + k∥rin∥2)."

    Eq. 2 pulls the learned deformation field toward an initializer fit to the same synthetic fisheye barrel-distortion family that generates the COCO-Fisheye test set. The reported fisheye improvements (+4.1 mAP overall; +3.6 mAP D-AAttn ablation) are therefore measurements of a prior computed from the test-generating transform, not an independent test on real fisheye imagery; the paper cites WoodScape but does not evaluate on it.

  2. self definitional [§3.3 (Level 1), §3.9.4, Limitations]
    "GameCharacterStylization applies: posterization (bit depth b∈[3,6]), unsharp masking (strength 0.5–1.0), saturation boost (×1.5–1.8), and contrast adjustment (α∈[1.2,1.8]). The parameters are estimated from a small set (N=50) of reference game images ... We apply global game-style rendering to COCO val2017 images. ... Limitations: The Game benchmark is synthetic; real-game performance is lower."

    The Game-Stylized COCO benchmark is produced by the same parameterized GameCharacterStylization used as the Level-1 data-level training augmentation, with stylization parameters estimated from reference game images of the same visual family. The +26.1 mAP gain is thus a train-on-test-distribution number, not evidence of adaptation to real game-rendered inputs. The paper's own limitation states real-game performance is lower, confirming that the headline game delta is a self-generated stress test rather than an independent cross-domain prediction.

full rationale

The core COCO val2017 result (49.1 mAP) is a standard, independent benchmark comparison: the model is trained on COCO train2017 and evaluated on COCO val2017, with no circularity. The VisDrone result is also real data. However, the two headline cross-domain gains that dominate the abstract are self-referential by construction. In §3.2, Eq. (2) regularizes the deformable offsets toward Δ_init 'estimated from synthetic fisheye transformations', and the COCO-Fisheye test set is generated by the same barrel-distortion family; the reported +4.1/+3.6 fisheye improvements are therefore partly a fit to the test-generating transform. In §3.3, GameCharacterStylization defines both the training augmentation and, via §3.9.4, the Game-Stylized COCO test set; the +26.1 mAP game delta is a within-distribution synthetic result, and the paper's own limitations concede that 'real-game performance is lower'. The COCO-Panorama benchmark is also synthetic, and while the paper does not explicitly state that the panorama generator was used in training, the scene classifier is trained on 2,000 synthetic panorama images, so the panorama route is likewise not an external test. The SWDA/MCD baseline comparison is also internally inconsistent (§4.1 says they receive no Game2Real stylization, while Appendix A.8 says SWDA uses Game stylization as its strong augmentation for Game evaluation), which further weakens the game comparison. No self-citation chain or imported uniqueness theorem is load-bearing. I score 6 rather than 8 because the COCO val2017 and VisDrone results are real independent benchmarks, so only the synthetic cross-domain deltas are substantially circular.

Assumptions & free parameters 8 free parameters · 3 assumptions · 0 invented entities

The central claims depend on several hand-chosen hyperparameters (loss weights, temperature schedule, window sizes, embedding dimensions) and on the assumption that synthetic benchmark transformations are faithful proxies for real domains. No new physical or conceptual entities are introduced; the modules are software constructs.

free parameters (8)
  • Loss weights lambda_reg, lambda_1, lambda_2 = 0.01, 0.1 (annealed), 0.05
    Loss weights chosen by hand with limited sensitivity analysis in Table 6.
  • Temperature schedule tau_min, tau_max, beta = 0.05, 0.5, 0.01
    Adaptive temperature parameters for contrastive loss; beta selected by grid search over {0.005, 0.01, 0.02} per Appendix A.2.
  • Window size W_A = 8
    Window size for D-AAttn; no ablation on window size is reported.
  • View embedding dimension d_emb = 64
    Hand-chosen dimension for the six-view embedding table.
  • Game stylization parameters = bit depth b in [3,6], unsharp 0.5-1.0, saturation x1.5-1.8, contrast alpha in [1.2,1.8]
    Estimated from N=50 reference game images; these parameters define the synthetic game benchmark and the training stylization.
  • Fisheye distortion levels k = 0.10, 0.15, 0.20
    Distortion levels used to synthesize the COCO-Fisheye benchmark.
  • SphereAAttn head allocation = 4 polar heads, 8 equatorial heads
    Fixed based on validation performance, as stated in Appendix A.1.
  • Class-balanced sampling counts = 3 samples per class across 2-3 views
    Batch sampling hyperparameter for the cross-view contrastive loss.
assumptions (3)
  • domain assumption Synthetic transformations (barrel distortion, equirectangular projection, game stylization) applied to COCO images are faithful proxies for real fisheye, panorama, and game-rendered inputs.
    Section 3.9 assumes these transformations simulate real deployment conditions; if false, the benchmarks measure performance on synthetic artifacts only.
  • ad hoc to paper The method may use target-domain statistics at training time via source-domain augmentation.
    Section 3.3 uses reference game images to stylize source data, changing the problem setting from unsupervised domain adaptation to a source-augmentation paradigm and making the comparison to UDA baselines unfair.
  • ad hoc to paper The D-AAttn offset initializer estimated from synthetic fisheye transformations improves performance on real fisheye inputs.
    Section 3.2 initializes the offset regularization target from synthetic fisheye transformations, which are the same family as the fisheye test benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of YOLOv14: Adaptive Real-Time Object Detection for Diverse Imaging Conditions." pith.science (2026). https://pith.science/paper/E27RWFSJ

@misc{pith2026260804720,
  author       = {Pith},
  title        = {Pith review of: YOLOv14: Adaptive Real-Time Object Detection for Diverse Imaging Conditions},
  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-renderedcharacters, aerial viewpoints, and 360{\deg}panoramas. We present YOLOv14, a detection framework with four adaptive mechanisms designed for specific types of inputvariation:(1) Deformable Area-Attention with windowed computation and shiftedwindows for geometric distortion;(2) Multi-level Game2Real Alignment with progressive adversarial training for domain shift;(3) View-Aware Contrastive Learning with adaptive temperature for viewpoint invariance; and (4) Scene-Adaptive Augmentation with dynamic loss balancing for scene diversity. Together, YOLOv14 achieves 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), andour synthesized game-character benchmark (+26.1 mAP). We release code and models to facilitate reproducible research.

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 2
Figure 2. Comparison of standard area-attention and deformable area-attention. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Latency-accuracy trade-off on COCO val2017. YOLOv14 [PITH_FULL_IMAGE:figures/full_fig_p019_3.png]
Figure 3
Figure 3. Figure 3: Examples from our four multi-domain benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]
Figure 4
Figure 4. Figure 4: Performance comparison across four challenging domains. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Radar chart of performance gains across domains. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

65 extracted references · 57 canonical work pages

  1. [1]

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

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. “You only look once: Unified, real-time object detection.” InCVPR, 2016

    work page 2016

  2. [2]

    YOLOv8: A new state-of-the-art real-time object detection model

    Ultralytics. “YOLOv8: A new state-of-the-art real-time object detection model.” https:// github.com/ultralytics/ultralytics, 2023

    work page 2023

  3. [3]

    YOLOv12: Attention-centric real-time object detectors

    Y . Tian, Q. Ye, and D. Doermann. “YOLOv12: Attention-centric real-time object detectors.” arXiv:2406.00430, 2024

    arXiv 2024

  4. [4]

    YOLOv3: An incremental improvement

    J. Redmon and A. Farhadi. “YOLOv3: An incremental improvement.”arXiv:1804.02767, 2018

    arXiv 2018

  5. [5]

    YOLOv4: Optimal speed and accuracy of object detection

    A. Bochkovskiy, C.-Y . Wang, and H.-Y . M. Liao. “YOLOv4: Optimal speed and accuracy of object detection.”arXiv:2004.10934, 2020

    arXiv 2004

  6. [6]

    YOLOX: Exceeding YOLO series in 2021

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun. “YOLOX: Exceeding YOLO series in 2021.” arXiv:2107.08430, 2021

    arXiv 2021

  7. [7]

    YOLOv6: A single-stage object detection framework for industrial applications

    C. Li, L. Li, H. Jiang, K. Weng, Y . Geng, L. Li, Z. Ke, Q. Li, M. Cheng, W. Nie, et al. “YOLOv6: A single-stage object detection framework for industrial applications.”arXiv:2209.02976, 2022

    arXiv 2022

  8. [8]

    YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors

    C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao. “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors.” InCVPR, 2023

    work page 2023

Show all 65 references
  1. [9]

    PP-YOLOE: An evolved version of YOLO

    S. Xu, X. Wang, W. Lv, Q. Chang, C. Cui, K. Deng, G. Wang, Q. Dang, S. Du, and Y . Wei. “PP-YOLOE: An evolved version of YOLO.”arXiv:2203.16250, 2022

    2022 arXiv

  2. [10]

    Gold-YOLO: Efficient object detector via gather-and-distribute mechanism

    C. Wang, W. He, Y . Nie, J. Guo, C. Liu, Y . Wang, and K. Han. “Gold-YOLO: Efficient object detector via gather-and-distribute mechanism.” InNeurIPS, 2023

    2023

  3. [11]

    YOLOv9: Learning what you want to learn using programmable gradient information

    C.-Y . Wang, I.-H. Yeh, and H.-Y . M. Liao. “YOLOv9: Learning what you want to learn using programmable gradient information.”arXiv:2402.13616, 2024

    2024 arXiv

  4. [12]

    YOLOv10: Real-time end-to-end object detection

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, and G. Ding. “YOLOv10: Real-time end-to-end object detection.”arXiv:2405.14458, 2024

    2024 arXiv

  5. [13]

    FCOS: Fully convolutional one-stage object detection

    Z. Tian, C. Shen, H. Chen, and T. He. “FCOS: Fully convolutional one-stage object detection.” InICCV, 2019

    2019

  6. [14]

    EfficientDet: Scalable and efficient object detection

    M. Tan, R. Pang, and Q. V . Le. “EfficientDet: Scalable and efficient object detection.” InCVPR, 2020

    2020

  7. [15]

    End-to-end object detection with transformers

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko. “End-to-end object detection with transformers.” InECCV, 2020

    2020

  8. [16]

    DINO: DETR with improved denoising anchor boxes for end-to-end object detection

    H. Zhang, F. Li, S. Liu, L. Zhang, H. Su, J. Zhu, L. M. Ni, and H.-Y . Shum. “DINO: DETR with improved denoising anchor boxes for end-to-end object detection.” InICLR, 2022

    2022

  9. [17]

    SSD: Single shot multibox detector

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg. “SSD: Single shot multibox detector.” InECCV, 2016

    2016

  10. [18]

    Feature pyramid networks for object detection

    T.-Y . Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie. “Feature pyramid networks for object detection.” InCVPR, 2017

    2017

  11. [19]

    Deformable DETR: Deformable transformers for end-to-end object detection

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai. “Deformable DETR: Deformable transformers for end-to-end object detection.” InICLR, 2021

    2021

  12. [20]

    Deformable convolutional networks

    J. Dai, H. Qi, Y . Xiong, Y . Li, G. Zhang, H. Hu, and Y . Wei. “Deformable convolutional networks.” InICCV, 2017. 9

    2017

  13. [21]

    Deformable ConvNets v2: More deformable, better results

    X. Zhu, H. Hu, S. Lin, and J. Dai. “Deformable ConvNets v2: More deformable, better results.” InCVPR, 2019

    2019

  14. [22]

    Vision transformer with deformable attention

    Z. Xia, X. Pan, S. Song, L. Li, and G. Huang. “Vision transformer with deformable attention.” InCVPR, 2022

    2022

  15. [23]

    CSWin Trans- former: A general vision transformer backbone with cross-shaped windows

    X. Dong, J. Bao, D. Chen, W. Zhang, N. Yu, L. Yuan, D. Chen, and B. Guo. “CSWin Trans- former: A general vision transformer backbone with cross-shaped windows.” InCVPR, 2022

    2022

  16. [24]

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

    Y . Chen, W. Li, C. Sakaridis, D. Dai, and L. Van Gool. “Domain adaptive faster R-CNN for object detection in the wild.” InCVPR, 2018

    2018

  17. [25]

    Domain-adversarial training of neural networks

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand, and V . Lempitsky. “Domain-adversarial training of neural networks.”JMLR, 2016

    2016

  18. [26]

    Deep unsupervised domain adaptation for object detection

    S. Li, C.-H. Lin, B. Price, S. Cohen, and H. Yang. “Deep unsupervised domain adaptation for object detection.”IEEE TPAMI, 2020

    2020

  19. [27]

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

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros. “Unpaired image-to-image translation using cycle-consistent adversarial networks.” InICCV, 2017

    2017

  20. [28]

    Maximum classifier discrepancy for unsu- pervised domain adaptation

    K. Saito, K. Watanabe, Y . Ushiku, and T. Harada. “Maximum classifier discrepancy for unsu- pervised domain adaptation.” InCVPR, 2018

    2018

  21. [29]

    Conditional adversarial domain adaptation

    M. Long, Z. Cao, J. Wang, and M. I. Jordan. “Conditional adversarial domain adaptation.” In NeurIPS, 2018

    2018

  22. [30]

    DeepJDOT: Deep joint distribution optimal transport for unsupervised domain adaptation

    B. B. Damodaran, B. Kellenberger, R. Flamary, D. Tuia, and N. Courty. “DeepJDOT: Deep joint distribution optimal transport for unsupervised domain adaptation.” InECCV, 2018

    2018

  23. [31]

    Domain generalization: A survey

    K. Zhou, Z. Liu, Y . Qiao, T. Xiang, and C. C. Loy. “Domain generalization: A survey.”IEEE TPAMI, 2022

    2022

  24. [32]

    Object detection in aerial images: A large-scale benchmark and challenges

    J. Ding, N. Xue, G.-S. Xia, and D. Dai. “Object detection in aerial images: A large-scale benchmark and challenges.”IEEE TPAMI, 2021

    2021

  25. [33]

    Remote sensing object detection meets deep learning: A survey

    X. Zhang, Y . Zhou, and J. Pan. “Remote sensing object detection meets deep learning: A survey.” IEEE TGRS, 2023

    2023

  26. [34]

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

    S. Yogamani, C. Hughes, J. Horgan, G. Sistu, P. Varley, D. O’Dea, M. Uricár, J. Millet, P. D. L. Rivas, S. P. K. M. et al. “WoodScape: A multi-task, multi-camera fisheye dataset for autonomous driving.” InICCV Workshops, 2019

    2019

  27. [35]

    OmniDet: Surround view camera based multi-task perception system for autonomous driving

    V . R. Kumar, S. Milz, M. Witt, H. Lutz, and C. Maier. “OmniDet: Surround view camera based multi-task perception system for autonomous driving.” InIROS, 2021

    2021

  28. [36]

    AutoAugment: Learning augmentation strategies from data

    E. D. Cubuk, B. Zoph, D. Mane, V . Vasudevan, and Q. V . Le. “AutoAugment: Learning augmentation strategies from data.” InCVPR, 2019

    2019

  29. [37]

    RandAugment: Practical automated data augmentation with a reduced search space

    E. D. Cubuk, B. Zoph, J. Shlens, and Q. V . Le. “RandAugment: Practical automated data augmentation with a reduced search space.” InCVPR Workshops, 2020

    2020

  30. [38]

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

    S. G. Müller and F. Hutter. “TrivialAugment: Tuning-free yet state-of-the-art data augmentation.” InICCV, 2021

    2021

  31. [39]

    Training generative adversarial networks with limited data

    T. Karras, M. Aittala, J. Hellsten, S. Laine, J. Lehtinen, and T. Aila. “Training generative adversarial networks with limited data.” InNeurIPS, 2020

    2020

  32. [40]

    mixup: Beyond empirical risk mini- mization

    H. Zhang, M. Cisse, Y . N. Dauphin, and D. Lopez-Paz. “mixup: Beyond empirical risk mini- mization.” InICLR, 2018

    2018

  33. [41]

    CutMix: Regularization strategy to train strong classifiers with localizable features

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y . Yoo. “CutMix: Regularization strategy to train strong classifiers with localizable features.” InICCV, 2019

    2019

  34. [42]

    Learning data augmentation strategies for object detection

    B. Zoph, E. D. Cubuk, G. Ghiasi, T.-Y . Lin, J. Shlens, and Q. V . Le. “Learning data augmentation strategies for object detection.” InECCV, 2020. 10

    2020

  35. [43]

    Generalized intersec- tion over union: A metric and a loss for bounding box regression

    H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese. “Generalized intersec- tion over union: A metric and a loss for bounding box regression.” InCVPR, 2019

    2019

  36. [44]

    Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection

    X. Li, W. Wang, L. Wu, S. Chen, X. Hu, J. Li, J. Tang, and J. Yang. “Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection.” InNeurIPS, 2020

    2020

  37. [45]

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

    Z. Zheng, P. Wang, W. Liu, J. Li, R. Ye, and D. Ren. “Distance-IoU loss: Faster and better learning for bounding box regression.” InAAAI, 2020

    2020

  38. [46]

    SIoU loss: More powerful learning for bounding box regression

    Z. Gevorgyan. “SIoU loss: More powerful learning for bounding box regression.”arXiv, 2022

    2022

  39. [47]

    OTA: Optimal transport assignment for object detection

    Z. Ge, S. Liu, Z. Li, O. Yoshie, and J. Sun. “OTA: Optimal transport assignment for object detection.” InCVPR, 2021

    2021

  40. [48]

    VarifocalNet: An IoU-aware dense object detector

    H. Zhang, Y . Wang, F. Dayoub, and N. Sünderhauf. “VarifocalNet: An IoU-aware dense object detector.” InCVPR, 2021

    2021

  41. [49]

    Analysis of representations for domain adaptation

    S. Ben-David, J. Blitzer, K. Crammer, and F. Pereira. “Analysis of representations for domain adaptation.” InNeurIPS, 2006

    2006

  42. [50]

    A theory of learning from different domains

    S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan. “A theory of learning from different domains.”Machine Learning, 2010

    2010

  43. [51]

    Domain adaptation: Learning bounds and algorithms

    Y . Mansour, M. Mohri, and A. Rostamizadeh. “Domain adaptation: Learning bounds and algorithms.” InCOLT, 2009

    2009

  44. [52]

    Foundations of Machine Learning

    M. Mohri, A. Rostamizadeh, and A. Talwalkar. “Foundations of Machine Learning.” MIT Press, 2nd edition, 2018

    2018

  45. [53]

    On learning invariant representations for domain adaptation

    H. Zhao, R. T. Des Combes, K. Zhang, and G. Gordon. “On learning invariant representations for domain adaptation.” InICML, 2019

    2019

  46. [54]

    Domain adaptation in regression

    C. Cortes and M. Mohri. “Domain adaptation in regression.” InALT, 2011

    2011

  47. [55]

    Microsoft COCO: Common objects in context

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. “Microsoft COCO: Common objects in context.” InECCV, 2014

    2014

  48. [56]

    Recognizing scene viewpoint using panoramic place representation

    J. Xiao, K. A. Ehinger, A. Oliva, and A. Torralba. “Recognizing scene viewpoint using panoramic place representation.” InCVPR, 2012

    2012

  49. [57]

    VisDrone-DET2021: The vision meets drone object detection challenge results

    Y . Cao, Z. He, L. Wang, W. Wang, Y . Yuan, D. Zhang, J. Zhang, P. Zhu, L. Van Gool, J. Han, et al. “VisDrone-DET2021: The vision meets drone object detection challenge results.” InICCV Workshops, 2021

    2021

  50. [58]

    Arbitrary style transfer in real-time with adaptive instance normal- ization

    X. Huang and S. Belongie. “Arbitrary style transfer in real-time with adaptive instance normal- ization.” InICCV, 2017

    2017

  51. [59]

    DETRs beat YOLOs on real-time object detection

    Y . Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y . Liu, and J. Chen. “DETRs beat YOLOs on real-time object detection.” InCVPR, 2024

    2024

  52. [60]

    Swin Transformer: Hierarchical vision transformer using shifted windows

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo. “Swin Transformer: Hierarchical vision transformer using shifted windows.” InICCV, 2021. A Additional Implementation Details A.1 Network Architecture Details Deformable Offset Network: • Input: feature map...

    2021

  53. [61]

    Input: COCO val2017 images (5,000 images,640×640pixels)

  54. [62]

    Spherical projection: For each pixel (x, y)in the output equirectangular image (1280×640 ), compute spherical coordinates(θ, ϕ): θ= 2πx/1280, ϕ=πy/640−π/2.(14)

  55. [63]

    Boundary wrapping: x= 0 wraps to x= 1280 ; y= 0 maps to the pole (sampled from the top row)

    Sampling: For each output pixel, sample the input image at coordinates (θ, ϕ)using bilinear interpolation. Boundary wrapping: x= 0 wraps to x= 1280 ; y= 0 maps to the pole (sampled from the top row)

  56. [64]

    The new box is the axis-aligned rectangle of the transformed corners

    Bounding box transformation: Each box corner is projected to spherical coordinates and then to equirectangular coordinates. The new box is the axis-aligned rectangle of the transformed corners

  57. [65]

    weak" and Game stylization as

    Discard boxes that fall outside the valid region after transformation. The validation split uses the original COCO val2017 indices to avoid overlap. A.4 Scene Classifier Confusion Matrix We report the confusion matrix of the scene classifier on the 2,000 held-out test images: ...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.