Pith. sign in

REVIEW 3 major objections 4 minor 282 references

Constraining super-resolution with hydrostatic primitive equations improves both the physical consistency and the pixel-level accuracy of super-resolved atmospheric fields—and sharpens detection of heatwaves and extreme winds.

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

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

T0 review · deepseek-v4-flash

2026-08-01 14:03 UTC pith:BD2E6HQ2

load-bearing objection A real multi-scale HPE idea, but the hydrostatic constraint assumes constant sea-level pressure and the evaluation leaks ground-truth latent fields, so the central claim doesn't hold yet. the 3 major comments →

arxiv 2607.18877 v1 pith:BD2E6HQ2 submitted 2026-07-21 cs.LG

Physics-Informed Super-Resolution of Atmospheric Data

Chang Xu , Gencer Sumbul , Hugo Porta , Manon B\'echaz , Sebastian Schemm , Devis Tuia This is my paper
classification cs.LG
keywords atmospheric super-resolutiondownscalingphysics-informed neural networkshydrostatic primitive equationsphysical consistency metricextreme event detectionclimate datamulti-scale physics loss
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper confronts a trust problem in data-driven downscaling: neural super-resolution can reconstruct fine-grained weather fields efficiently, but nothing forces those fields to respect the atmosphere's governing equations. It proposes PISR, a physics-informed super-resolution method that adds the hydrostatic primitive equations as multi-scale soft constraints to the standard reconstruction loss, and introduces the Normalized Physical Consistency (NPC) metric to quantify how far a super-resolved state departs from the ground truth's physical residuals. The central claim, supported on global and regional datasets at three resolutions, is that encoding these multi-variable physical relationships does not sacrifice fidelity—it improves it, lowering both NPC and per-variable reconstruction error. The paper further claims these gains carry over to downstream tasks: super-resolved temperature and wind fields detect heatwaves and extreme wind events more accurately than physics-agnostic baselines, with foreground overlap improving by 24% on heatwaves.

Core claim

The paper's discovery is that the hydrostatic primitive equations—hydrostatic balance, mass continuity, horizontal momentum, thermodynamics—can be turned from a description of atmospheric dynamics into a practical super-resolution training signal. PISR separates variables into observable ones from the SR output and latent ones (vertical velocity, friction, diabatic heating) computed from high-resolution ground truth, so the equations close without predicting every field. A multi-scale physics loss enforces these residuals at target and downsampled coarser scales; the NPC metric measures the relative discrepancy between predicted and ground-truth residuals. Across global and regional data at

What carries the argument

The carrying mechanism is the physics residual operator R_HPEs = [R_hydro, R_mass, R_mom, R_thermo], built from the hydrostatic primitive equations and minimized as a soft constraint. Missing latent fields are not modeled; they are numerically resolved from ground-truth high-resolution fields, giving physically meaningful proxies for vertical velocity, friction, and diabatic heating. The hydrostatic constraint is encoded through the hypsometric relation: minimizing the spatial variance of ln p_s + g z_s/(R_d T_v) enforces a constant inferred sea-level reference pressure. Multi-scale pooling applies the same residuals at downsampled scales to capture relationships that are more robust at coar

Load-bearing premise

The load-bearing premise is that the unobserved physical fields—vertical velocity, friction, and diabatic heating—can be reliably computed from high-resolution ground truth and used as proxies during training and evaluation, and that a spatially constant sea-level reference pressure makes the hydrostatic constraint valid; if either fails, the reported physical consistency and its downstream benefits are conditional on information that is absent at inference.

What would settle it

Compute the spatial variance of ln p_s + g z_s/(R_d T_v) on high-resolution ground-truth fields across flat and mountainous terrain: if it is far from zero wherever non-hydrostatic motions are active, the constraint is punishing real physics. Independently, recompute NPC for PISR outputs using only observable variables and an independently estimated latent vertical velocity/friction/heating—rather than ground truth—and check whether the metric still separates PISR from baselines; if the ordering flips, the claimed physical consistency is an artifact of ground-truth-informed latents.

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

If this is right

  • If the claims hold, any existing SR backbone—deterministic or generative—can be made more physically consistent and more accurate by adding the multi-scale HPE loss, with no added cost at inference.
  • NPC gives the field a single, equation-derived number for comparing physical consistency across downscaling methods, beyond single-law checks like mass conservation.
  • Physics-informed downscaling improves rare-event detection even where per-variable RMSE gains are modest, suggesting physical consistency is a structured prior that matters most for extremes.
  • The approach is resolution-dependent: gains are strongest for hydrostatic balance at coarse scales and weaken at fine scales, where non-hydrostatic motions dominate.
  • Prognostic equations (continuity, thermodynamics) are harder to enforce than balance equations, so further gains would require better treatment of unobserved heating and vertical motion.

Where Pith is reading between the lines

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

  • Editorial extension: because latent fields are taken from ground truth, a practical deployment would need a separate estimator of vertical velocity, friction, and heating from coarse inputs; that estimator could itself be learned with the same HPE residuals as supervision.
  • Editorial extension: the hydrostatic constraint's constant-sea-level-pressure assumption could over-regularize regions with strong mesoscale pressure gradients (mountainous or convective areas); comparing the Var term's magnitude on flat vs. complex terrain would test this.
  • Editorial extension: the heatwave foreground IoU gain (24%) being much larger than the corresponding T2m RMSE gain suggests the physics loss regularizes spatial structure rather than just pixel values; a natural test is whether similar disproportionate gains appear for other spatially coherent hazards such as heavy precipitation or fog.
  • Editorial extension: the same multi-scale residual framework could be adapted to non-hydrostatic equations for convection-permitting resolutions, where the hydrostatic approximation itself is the limiting assumption.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes Physics-Informed Super-Resolution (PISR), a multivariate SR method that augments a standard MSE reconstruction loss with multi-scale soft constraints derived from the hydrostatic primitive equations. It further introduces a Normalized Physical Consistency (NPC) metric that compares the residual of a reconstructed field against the residual of the ground truth. Experiments on ERA5, CERRA, and COSMO, with deterministic and generative baselines, report improvements in NPC, RMSE, and downstream heatwave/extreme-wind detection. The central claim is that incorporating HPEs as multi-scale constraints improves both physical consistency and pixel-level reconstruction fidelity.

Significance. If the central claim were sound, the paper would make a useful contribution to atmospheric downscaling: it targets multivariate coupling rather than single-variable SR, uses a multi-scale physics loss, and evaluates downstream extreme-event detection, which is a relevant and often neglected test. The experimental design is broad, with three datasets, several baselines, architectural generalization, and ablations. However, the central physical-consistency claim is not supported: the hydrostatic residual is based on a false assumption that sea-level reference pressure is spatially constant, and the mass/momentum/thermodynamic residuals rely on latent fields (w, F, H) that are taken from the ground truth in both training and evaluation. These are load-bearing issues, not presentation issues.

major comments (3)
  1. [Appendix A.2, Eq. (36)] The hydrostatic residual is derived incorrectly. Integrating the hypsometric equation gives ln p_s + g z_s/(R_d T_v) ≈ ln p_0, where p_0 is the pressure at the reference elevation (sea level). The paper states that 'the right-hand side is constant over the spatial domain' and therefore minimizes Var(ln p_s + g z_s/(R_d T_v)). This is physically false: p_0 is the sea-level pressure, which varies spatially with synoptic systems and orography. Minimizing this variance drives the reconstructed surface fields toward a spatially constant sea-level pressure, not toward hydrostatic balance. Since R_hydro is used both in training and in the NPChydro evaluations in Tables 1–3, the reported hydrostatic consistency improvements do not establish physical consistency; they may reflect an artificial flattening of sea-level pressure.
  2. [Section 3.3, Table 8, Eq. (18)] The latent fields w, F, and H are 'numerically resolved from the corresponding ground-truth high-resolution fields' and used in the physics losses and in NPC. At inference these fields are not available, so the physics residual for a prediction is undefined unless the ground-truth latent fields are supplied. The NPC metric therefore compares a prediction residual that is conditional on ground truth with a ground-truth residual computed from the same latent fields. This does not measure whether the SR output alone respects the HPEs; it measures consistency conditional on information that is unavailable in the operational setting. The conclusion that PISR 'improves the physical consistency of the predicted fields' is thus overstated, especially for the mass, momentum, and thermodynamic components.
  3. [Section 4.4, Tables 1–3] The paper interprets the large improvements in NPChydro as evidence that neural networks 'easily reduce the errors of the hydrostatic relation.' Because the hydrostatic residual itself enforces an unphysical spatially constant p0, the numerical gains in NPChydro and the associated qualitative statements in Section 4.4 do not support the paper's physical-consistency claim. At minimum, the authors would need to replace Eq. (36) with a spatially varying p0 treatment or remove the hydrostatic constraint and re-evaluate all reported conclusions.
minor comments (4)
  1. [Section 4.4] The sentence 'NPC mom is reduced from 0.601 to near 0.000 while NPC mom is reduced from 0.435 to 0.275' contains a duplication; the second quantity appears to refer to NPCthermo. Please correct.
  2. [Figure 1 / Section 3.3] The figure caption states that latent fields are estimated from ground truth, but the main text should more prominently state that these fields are unavailable at inference and that the physics losses are therefore not applicable to deployment-time evaluation.
  3. [Eq. (13) vs. Eq. (36)] R_hydro in Eq. (36) is not of the form R_k = F_k(â_o,k, a_l,k) used in Eq. (13). It is a variance operator on observable variables only. The authors should explain this special case or reformulate it consistently.
  4. [Table 3] The COSMO table lists a column labeled 'PISR' with checkmarks but no explicit column header for the model variants; the formatting makes it hard to read. Please clarify.

Circularity Check

0 steps flagged

No significant circularity found; the PISR derivation chain does not reduce to its own inputs.

full rationale

The paper's central claim is that adding hydrostatic primitive equation residuals as soft, multi-scale constraints improves both physical consistency (NPC) and pixel-level reconstruction (RMSE) on ERA5, CERRA, and COSMO. The derivation chain from HPEs to the physics loss (Eqs. 12-17) is standard: differential residual operators are constructed, latent variables (w, F, H) are explicitly stated to be resolved from ground-truth fields (Section 3.3), and the physics loss is RMSE of the residuals. The NPC metric (Eq. 18) compares prediction residual to ground-truth residual; since both use the same ground-truth latent variables, this is an oracle-conditioned diagnostic, not a hidden re-use of the output as input. This is a limitation in scope (the metric and loss do not evaluate the model's ability to infer latent fields), but it is transparently disclosed and does not make the reported quantities equal by construction. The hydrostatic constraint in Eq. 36 assumes a spatially constant sea-level reference pressure p0; even if this is meteorologically questionable and a correctness risk, it is not circular. The only self-citation (GeoFAR, Xu et al. 2026) appears in related work and is not load-bearing. No uniqueness theorem is imported from the authors, no ansatz is smuggled in via a prior self-citation, and no known result is merely renamed. The improvements over baselines and on downstream extreme-event detection are therefore not forced by the paper's own definitions. The self-reported limitations about hydrostatic assumptions and latent-field coverage further confirm that the claimed physics consistency is conditional rather than definitionally guaranteed, but this does not constitute circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced; w, F, and H are real atmospheric quantities, though their estimation from ground-truth surface fields is not specified. The ledger shows that the central claim rests on several domain assumptions, including two that are questionable or false: the constant sea-level pressure assumption and the availability of ground-truth latent fields at inference.

free parameters (3)
  • physics loss weights omega_k = 1 (all equations and all scales)
    Set to one to avoid dataset-specific tuning (Eq. 16, Appendix D); an arbitrary hyperparameter choice that affects the balance between physical constraints.
  • NPC stabilizer epsilon = 1e-12
    Chosen for numerical stability in Eq. 18; directly controls NPC values when ground-truth residuals are near zero.
  • multi-scale pooling factors S = {1, 2}
    The authors use the target resolution and one 2x-downsampled scale (Eq. 15-16); other scale choices could change results.
axioms (5)
  • domain assumption Hydrostatic primitive equations sufficiently describe the near-surface atmospheric variables used in super-resolution.
    The method and NPC metric are built on HPEs; the paper itself notes hydrostatic balance becomes less accurate at fine scales (Section 4.4, Conclusion).
  • ad hoc to paper Latent fields w, F, H can be reliably derived from ground-truth high-resolution fields.
    Section 3.3 uses these latent fields for training and evaluation but does not specify how they are computed; their availability at inference is assumed for the physical-consistency claims.
  • ad hoc to paper Sea-level reference pressure p0 is constant over the spatial domain.
    Appendix A.2 (Eq. 35-36) derives R_hydro by assuming ln p0 is constant; sea-level pressure is a spatially varying field, so this assumption is physically false.
  • domain assumption 2-m virtual temperature approximates the layer-mean virtual temperature in the hypsometric equation.
    Appendix A.1 and A.2 replace the layer-mean T_v with 2-m virtual temperature; this is a rough approximation for a soft constraint.
  • domain assumption Finite-difference derivatives with dataset-specific dx, dy, dt are adequate for evaluating HPE residuals.
    Appendix D sets dx, dy, dt to grid spacing and time step; numerical differentiation of reanalysis fields may be noisy and resolution-dependent.

pith-pipeline@v1.3.0-alltime-deepseek · 19339 in / 17999 out tokens · 159019 ms · 2026-08-01T14:03:16.055867+00:00 · methodology

0 comments
read the original abstract

In the context of global warming, extreme events have become more frequent and intense, making their trustworthy detection and forecasting more important than ever. Yet, atmospheric observations lack sufficient spatial resolution, motivating atmospheric data downscaling as a way to reconstruct high-resolution data from coarse observations. This task is now being formulated as a super-resolution (SR) problem with machine learning methods featuring high efficiency. Nevertheless, it remains unclear whether the super-resolved atmospheric data still satisfies fundamental physics governing the Earth system, raising concerns about their trustworthiness in climate-related applications. In this work, we address this challenge by constraining SR models to respect hydrostatic primitive equations that represent multivariate atmospheric physics. First, we propose a Physics-Informed Super-Resolution (PISR) method involving multi-scale physics-informed objectives based on primitive equations. PISR favors the SR outputs to respect these equations and therefore naturally encodes inter-variable relationships. In addition, we propose a metric called Normalized Physical Consistency (NPC) derived from said primitive equations to measure the physical consistency of super-resolved data. Experiments on ERA5, CERRA, and COSMO demonstrate that PISR enhances the reconstruction fidelity by improving physical consistency, SR accuracy, and downstream detection of extreme events, as demonstrated by case studies in heatwaves and extreme winds.

Figures

Figures reproduced from arXiv: 2607.18877 by Chang Xu, Devis Tuia, Gencer Sumbul, Hugo Porta, Manon B\'echaz, Sebastian Schemm.

Figure 1
Figure 1. Figure 1: Overview of the proposed PISR method. Given coarse-resolution atmo￾spheric fields at multiple time steps, a backbone network predicts high-resolution atmo￾spheric fields. Differential operators are applied to the prediction to get observable fields ˆao, while the numerical solver estimates latent fields al from the ground truth. The predic￾tions are supervised by a reconstruction loss and further constrain… view at source ↗
Figure 2
Figure 2. Figure 2: Schematic illustration of the hydrostatic primitive equations on a unit parcel. (a) hydrostatic balance between the vertical pressure gradient and gravity: ∂p ∂z = −ρg; (b) mass continuity relating the density tendency to the divergence mass fluxes: ∂ρ ∂t + ∂(ρu) ∂x + ∂(ρv) ∂y + ∂(ρw) ∂z = 0; (c) horizontal momentum balance among wind acceleration, the pressure gradient, the Coriolis force, and external fo… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of the super-resolution results produced by the baseline [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of detected heatwaves based on CERRA. Red pixels [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of extreme wind detections. Blue pixels indicate detected [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

282 extracted references · 3 canonical work pages

  1. [1]

    2017 , volume=

    Sun, Chen and Shrivastava, Abhinav and Singh, Saurabh and Gupta, Abhinav , booktitle=ICCV, title=. 2017 , volume=

    2017

  2. [2]

    Image and Vision Computing , volume=

    Recent advances in small object detection based on deep learning: A review , author=. Image and Vision Computing , volume=. 2020 , publisher=

    2020

  3. [3]

    Ninth International Conference on Graphic and Image Processing (ICGIP 2017) , volume=

    Feature-fused SSD: Fast detection for small objects , author=. Ninth International Conference on Graphic and Image Processing (ICGIP 2017) , volume=. 2018 , organization=

    2017

  4. [4]

    Object detection via a multi-region and semantic segmentation-aware cnn model , author=

  5. [5]

    Relation networks for object detection , author=

  6. [6]

    arXiv preprint arXiv:1902.07296 , year=

    Augmentation for small object detection , author=. arXiv preprint arXiv:1902.07296 , year=

    Pith/arXiv arXiv 1902

  7. [7]

    2018 , organization=

    Kim, Yonghyun and Kang, Bong-Nam and Kim, Daijin , title =. 2018 , organization=

    2018

  8. [8]

    Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection , author=

  9. [9]

    2020 , pages=

    Probabilistic anchor assignment with iou prediction for object detection , author=. 2020 , pages=

    2020

  10. [10]

    2021 , pages=

    OTA: Optimal Transport Assignment for Object Detection , author=. 2021 , pages=

    2021

  11. [11]

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

  12. [12]

    Pattern Recognition , volume=

    Iou-uniform r-cnn: Breaking through the limitations of rpn , author=. Pattern Recognition , volume=. 2021 , publisher=

    2021

  13. [13]

    Dot distance for tiny object detection in aerial images , author=

  14. [14]

    Rethinking rotated object detection with gaussian wasserstein distance loss , author=

  15. [15]

    , booktitle=ICCVW, title=

    Du, Dawei and Zhu, Pengfei and Wen, Longyin and et al. , booktitle=ICCVW, title=. 2019 , booktitle=ICCVW, volume=

    2019

  16. [16]

    Wider face: A face detection benchmark , author=

  17. [17]

    Cascade RPN: Delving into High-Quality Region Proposal Network with Adaptive Convolution , author=

  18. [18]

    Sun, Ke and Xiao, Bin and Liu, Dong and Wang, Jingdong , title =

  19. [19]

    Precise detection in densely packed scenes , author=

  20. [20]

    S3fd: Single shot scale-invariant face detector , author=

  21. [21]

    2021 , volume=

    Oksuz, Kemal and Cam, Baris Can and Kalkan, Sinan and Akbas, Emre , journal=TPAMI, title=. 2021 , volume=

    2021

  22. [22]

    Group sampling for scale invariant face detection , author=

  23. [23]

    ISPRS Journal of Photogrammetry and Remote Sensing , volume=

    FAIR1M: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery , author=. ISPRS Journal of Photogrammetry and Remote Sensing , volume=. 2022 , publisher=

    2022

  24. [24]

    2020 , issn =

    HyNet: Hyper-scale object detection network framework for multiple spatial resolution remote sensing imagery , journal = ISPRSJ, volume =. 2020 , issn =. doi:https://doi.org/10.1016/j.isprsjprs.2020.04.019 , author =

    work page doi:10.1016/j.isprsjprs.2020.04.019 2020

  25. [25]

    2018 , issn =

    Multi-scale object detection in remote sensing imagery with convolutional neural networks , journal = ISPRSJ, volume =. 2018 , issn =. doi:https://doi.org/10.1016/j.isprsjprs.2018.04.003 , url =

    work page doi:10.1016/j.isprsjprs.2018.04.003 2018

  26. [26]

    2021 , pages =

    Sun, Peize and Zhang, Rufeng and Jiang, Yi and Kong, Tao and Xu, Chenfeng and Zhan, Wei and Tomizuka, Masayoshi and Li, Lei and Yuan, Zehuan and Wang, Changhu and Luo, Ping , title =. 2021 , pages =

    2021

  27. [27]

    2020 , organization=

    End-to-end object detection with transformers , author=. 2020 , organization=

    2020

  28. [28]

    Deformable DETR: Deformable Transformers for End-to-End Object Detection , author=

  29. [29]

    2020 , publisher=

    RFP-Net: Receptive field-based proposal generation network for object detection , author=. 2020 , publisher=

    2020

  30. [30]

    Understanding the effective receptive field in deep convolutional neural networks , author=

  31. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Prime sample attention in object detection , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  32. [32]

    arXiv preprint arXiv:2106.06072 , year=

    Gaussian Bounding Boxes and Probabilistic Intersection-over-Union for Object Detection , author=. arXiv preprint arXiv:2106.06072 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2110.13389 , year=

    A Normalized Gaussian Wasserstein Distance for Tiny Object Detection , author=. arXiv preprint arXiv:2110.13389 , year=

    Pith/arXiv arXiv

  34. [34]

    Foundations and Trends in Machine Learning , volume=

    Computational optimal transport: With applications to data science , author=. Foundations and Trends in Machine Learning , volume=. 2019 , publisher=

    2019

  35. [35]

    2020 , publisher=

    Small-object detection in remote sensing images with end-to-end edge-enhanced GAN and object detector network , author=. 2020 , publisher=

    2020

  36. [36]

    2021 , publisher=

    Small object detection in remote sensing images with residual feature aggregation-based super-resolution and object detector network , author=. 2021 , publisher=

    2021

  37. [37]

    2020 , publisher=

    Small object detection in remote sensing images based on super-resolution with auxiliary generative adversarial networks , author=. 2020 , publisher=

    2020

  38. [38]

    1996 , publisher=

    Digital image processing , author=. 1996 , publisher=

    1996

  39. [39]

    You only look one-level feature , author=

  40. [40]

    arXiv preprint arXiv:2007.03496 , year=

    Autoassign: Differentiable label assignment for dense object detection , author=. arXiv preprint arXiv:2007.03496 , year=

    Pith/arXiv arXiv 2007

  41. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Iqdet: Instance-wise quality distribution sampling for object detection , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  42. [42]

    ISPRS Journal of Photogrammetry and Remote Sensing , volume=

    Detecting Tiny Objects in Aerial Images: A Normalized Wasserstein Distance and A New Benchmark , author=. ISPRS Journal of Photogrammetry and Remote Sensing , volume=

  43. [43]

    ISPRS Journal of Photogrammetry and Remote Sensing , volume=

    Visual object tracking by correlation filters and online learning , author=. ISPRS Journal of Photogrammetry and Remote Sensing , volume=. 2018 , publisher=

    2018

  44. [44]

    Entropy , volume=

    On a generalization of the Jensen--Shannon divergence and the Jensen--Shannon centroid , author=. Entropy , volume=. 2020 , publisher=

    2020

  45. [45]

    arXiv preprint arXiv:1805.00123 , year=

    Crowdhuman: A benchmark for detecting human in a crowd , author=. arXiv preprint arXiv:1805.00123 , year=

    Pith/arXiv arXiv

  46. [46]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Precise detection in densely packed scenes , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  47. [47]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Adaptive nms: Refining pedestrian detection in a crowd , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  48. [48]

    Dynamic refinement network for oriented and densely packed object detection , author=

  49. [49]

    Beyond bounding-box: Convex-hull feature adaptation for oriented and densely packed object detection , author=

  50. [50]

    International Workshop on Machine Learning in Medical Imaging , pages=

    Detecting lesion bounding ellipses with gaussian proposal networks , author=. International Workshop on Machine Learning in Medical Imaging , pages=. 2019 , organization=

    2019

  51. [51]

    2023 , volume=

    Detecting Rotated Objects as Gaussian Distributions and Its 3-D Generalization , author=. 2023 , volume=

    2023

  52. [52]

    IEEE Transactions on Information theory , volume=

    Divergence measures based on the Shannon entropy , author=. IEEE Transactions on Information theory , volume=. 1991 , publisher=

    1991

  53. [53]

    Proceedings of the 30th ACM International Conference on Multimedia , pages=

    Mmrotate: A rotated object detection benchmark using pytorch , author=. Proceedings of the 30th ACM International Conference on Multimedia , pages=

  54. [54]

    2021 , publisher=

    Align deep features for oriented object detection , author=. 2021 , publisher=

    2021

  55. [55]

    Oriented reppoints for aerial object detection , author=

  56. [56]

    Redet: A rotation-equivariant detector for aerial object detection , author=

  57. [57]

    2018 , publisher=

    Arbitrary-oriented scene text detection via rotation proposals , author=. 2018 , publisher=

    2018

  58. [58]

    Oriented R-CNN for object detection , author=

  59. [59]

    Region proposal by guided anchoring , author=

  60. [60]

    arXiv preprint arXiv:2111.10780 , year=

    Fcosr: A simple anchor-free rotated detector for aerial object detection , author=. arXiv preprint arXiv:2111.10780 , year=

    Pith/arXiv arXiv

  61. [61]

    Dynamic anchor learning for arbitrary-oriented object detection , author=

  62. [62]

    Shape-Adaptive Selection and Measurement for Oriented Object Detection , author=

  63. [63]

    2022 , publisher=

    A general gaussian heatmap label assignment for arbitrary-oriented object detection , author=. 2022 , publisher=

    2022

  64. [64]

    2022 , publisher=

    Anchor-free oriented proposal generator for object detection , author=. 2022 , publisher=

    2022

  65. [65]

    Anchor detr: Query design for transformer-based detector , author=

  66. [66]

    DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR , author=

  67. [67]

    2022 , organization=

    RFLA: Gaussian Receptive Field based Label Assignment for Tiny Object Detection , author=. 2022 , organization=

    2022

  68. [68]

    2022 , pages=

    ObjectBox: From Centers to Boxes for Anchor-Free Object Detection , author=. 2022 , pages=

    2022

  69. [69]

    Feature selective anchor-free module for single-shot object detection , author=

  70. [70]

    2021 , publisher=

    Object detection in aerial images: A large-scale benchmark and challenges , author=. 2021 , publisher=

    2021

  71. [71]

    基于深度学习的遥感图像车辆目标检测 , author=

  72. [72]

    基于深度卷积神经网络的遥感图像车辆检测 , author=

  73. [73]

    武汉大学学报（信息科学版） , volume=

    基于多通道融合 HOG 特征的全天候运动车辆检测方法 , author=. 武汉大学学报（信息科学版） , volume=. 2015 , publisher=

    2015

  74. [74]

    分米级可见光遥感图像特定目标的在线检测 , author=

  75. [75]

    2004 , organization=

    Recognizing human actions: a local SVM approach , author=. 2004 , organization=

    2004

  76. [76]

    Statistics and its Interface , volume=

    Multi-class adaboost , author=. Statistics and its Interface , volume=. 2009 , publisher=

    2009

  77. [77]

    1999 , organization=

    Object recognition from local scale-invariant features , author=. 1999 , organization=

    1999

  78. [78]

    Thundernet: Towards real-time generic object detection on mobile devices , author=

  79. [79]

    G-cnn: an iterative grid based object detector , author=

  80. [80]

    Parallel feature pyramid network for object detection , author=

Showing first 80 references.