Pith. sign in

REVIEW 4 major objections 5 minor 69 references

Measuring the Impact of Rotation Equivariance on Aerial Object Detection

T0 review · 4 major / 5 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 demonstrates that strict rotation equivariance, preserved by a tuning layer before each downsampling step, improves aerial object detection enough that an 18.1M-parameter detector outperforms its approximate counterpart and…

desk verdict The paper's central claim—measuring strict vs. approximate rotation equivariance—does not hold up, because the tuning layer used to enforce strictness is itself an even-sized convolution that breaks rotation equivariance on the discrete grid. read the letter →

arxiv 2507.09896 v1 pith:RH2AOOAD submitted 2025-07-14 cs.CV

Xiuyu Wu , Xinhao Wang , Xiubin Zhu , Lan Yang , Jiyuan Liu , Xingchen Hu This is my paper
classification cs.CV
keywords rotationequivarianceaerialobjectdetectionorientedboundingboxesgroup-equivariantconvolutiondownsamplingchannelattentionmulti-branchheadDOTA
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

This paper asks whether a detector for aerial imagery really needs strict rotation equivariance, or whether the approximate rotation equivariance learned through data augmentation is enough. The authors build MessDet, a single-stage oriented-object detector whose rotation-equivariant backbone and neck can be switched between approximate and strict equivariance by inserting or omitting a tuning layer before each 2x downsampling. On DOTA-v1.0, DOTA-v1.5 and DIOR-R, the strictly equivariant variant scores 79.12%, 73.14% and 68.19% mAP, edging out the approximate variant while using only 18.1M parameters. The paper also reports that rotation equivariance error in the approximate model decreases during training on aerial data, which it reads as evidence that aerial detection itself benefits from rotation equivariance.

What carries the argument

The load-bearing device is the tuning layer: a $4\times4$ kernel with stride 1 and padding 1 that maps an even $2n\times2n$ feature map to $(2n-1)\times(2n-1)$, so the following $3\times3$ stride-2 convolution always sees an odd-sized input and its sampling grid stays aligned under rotation. The paper relies on the parity condition $(i-k)\bmod s=0$ for equivariant downsampling, plus an explicit example showing that stride-2 sampling on even grids hits odd rows before rotation and even rows after a 90-degree rotation. Two supporting mechanisms are rotation-equivariant channel attention (RE-CA), which produces only $C/N$ per-orientation channel weights and repeats them $N$ times so weighting does not break the cyclic group structure, and the multi-branch head, which rearranges the $C$ channels into $N$ orientation groups and processes each group in its own branch before concatenation.

What would settle it

Train a strictly equivariant MessDet and an approximate counterpart matched for parameter count and FLOPs, for example by giving the approximate model equivalent extra channels or capacity, on DOTA-v1.0 and DOTA-v1.5; if strict equivariance is the active ingredient, the strict variant stays ahead, otherwise the gap should shrink or vanish. A complementary check would be to measure mAP after removing rotation data augmentation from training, since strict equivariance should absorb most of the loss.

Watch

Extended reading notes

Core claim

The central claim is that strict rotation equivariance, not just approximate equivariance learned from rotated training data, is measurably beneficial for oriented aerial object detection, and that this property can be kept without paying the usual cost in parameters. MessDet achieves this by inserting a tuning layer, a $4\times4$ stride-1 convolution, before every stride-2 downsampling layer, converting even-sized feature maps to odd sizes so that kernel sampling points align under rotation; by a rotation-equivariant channel attention that weights the orientation-group channels rather than breaking their structure; and by a multi-branch head that feeds features from each of $N$ orientations into separate branches, then concatenates the outputs, cutting head parameters. With $N=8$ orientation dimensions, the strictly equivariant MessDet outperforms its approximate twin on all three benchmarks and reaches the accuracy of much heavier detectors at roughly one third to one half of their parameter counts.

Load-bearing premise

The result that strict rotation equivariance helps aerial detection rests on comparing a strictly equivariant network with an approximate one that also differs in parameter count and FLOPs, so part of the mAP gap could come from the extra capacity of the tuning layers rather than from equivariance itself.

Editorial extensions

If this is right

  • If the measured gains hold, aerial detectors can be made strictly rotation-equivariant without single-sided padding, avoiding the feature misalignment that padding introduces.
  • A rotation-equivariant detector with 18.1M parameters can match or exceed the accuracy of detectors two to three times larger, so equivariance can substitute for parameter capacity in oriented detection.
  • The rotation equivariance error of an approximately equivariant aerial detector falls during training, a behavior opposite to what is reported for classification, making equivariance error a useful training diagnostic.
  • The grouped structure of rotation-equivariant features can be exploited architecturally, as the multi-branch head does, reducing parameters while improving accuracy.

Reading between the lines

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

  • The strict and approximate variants differ not only in equivariance but in parameters (18.1M vs 15.3M) and FLOPs (570G vs 378G), because the tuning layers add capacity; a parameter-matched strict/approximate comparison would isolate how much of the mAP gap is equivariance rather than model size.
  • If strict equivariance is the real source of the gain, it should also improve sample efficiency; training with fewer rotated augmentations or fine-tuning from less rotation-augmented data would be a direct test.
  • The same tuning-layer trick and grouped multi-branch head could be carried to other equivariant symmetries, such as reflection or scale groups, in dense prediction tasks beyond aerial detection.
  • Rotation equivariance error, reported only for backbone stages, could be logged per head branch and per orientation dimension to identify where approximate equivariance still leaks into the final detection output.
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

4 major / 5 minor

Summary. The paper proposes MessDet, a rotation-equivariant single-stage detector for aerial imagery built on an E2CNN-style reimplementation of the RTMDet backbone and neck. The central contribution is a claimed causal measurement: comparing a strictly rotation-equivariant variant (Str. MessDet) against an approximately equivariant variant (Appr. MessDet) to quantify the impact of strict rotation equivariance. The strict variant introduces a so-called tuning layer, a k=4, s=1, p=1 RE-Conv placed before each 2x downsampling to convert even-sized feature maps to odd-sized ones. The paper also contributes a rotation-equivariant channel attention mechanism and a multi-branch head that exploits the grouped structure of rotation-equivariant features. Experiments on DOTA-v1.0, DOTA-v1.5, and DIOR-R report state-of-the-art or near-state-of-the-art mAP with 15.3M–18.1M parameters, and the supplementary reports FLOPs, FPS, and training time.

Significance. If the central claim is valid, the paper would provide a useful quantitative answer to an open question in aerial detection: whether strict, not merely approximate, rotation equivariance matters for detector accuracy. The engineering contributions are also valuable: the paper ships code, includes extensive comparisons, ablates the attention and head components, and demonstrates an unusually low parameter count. The main significance is tempered, however, by two unresolved issues: the strictness of the proposed tuning layer is not proven and is doubtful as written, and the strict-versus-approximate comparison is confounded by a 2.8M parameter and 192G FLOP difference.

major comments (4)
  1. [Sec. 4.1, Fig. 4b] The tuning layer that defines the 'strict' variant is a k=4, p=1, s=1 RE-Conv. On the standard integer lattice this layer has effective kernel offsets in {-1,0,1,2} in each spatial axis, and under a 90-degree rotation this support maps to {-2,-1,0,1} in one axis, which is not the same set. Therefore the convolution is not exactly equivariant under the grid rotation used in Eq. (1) and Fig. 2 unless an additional interpolation or symmetry convention is specified. The paper cites the condition (i-k) mod s = 0 from [16] only for the downsampling layer, not for the tuning layer itself. Since the tuning layer is the only architectural difference between Str. MessDet and Appr. MessDet, the manuscript does not currently establish that Str. MessDet is strictly rotation-equivariant. The authors should either prove exact equivariance of the full network including tuning layers, or measure the equivariance error of the complete Str. MessDet and show that it is exactly zero.
  2. [Sec. 5.3, Tables 1, 3, 6; Table 9] The central comparison is confounded by capacity. Str. MessDet uses 18.1M parameters and 570G FLOPs, while Appr. MessDet uses 15.3M parameters and 378G FLOPs, with the difference coming from the inserted tuning layers. The reported improvements of 0.67 mAP on DOTA-v1.0, 0.76 mAP on DOTA-v1.5, and 0.77 mAP on DIOR-R are attributed entirely to strict equivariance, but no parameter-matched or FLOP-matched baseline is provided. A control that adds the same number of non-equivariant convolutional parameters to Appr. MessDet, or one that removes layers from Str. MessDet to match capacity, is needed to separate the effect of equivariance from the effect of additional model capacity.
  3. [Sec. 5.2, Sec. 5.3] The magnitude of the central effect is small relative to typical training noise in aerial detection. The differences between Str. and Appr. MessDet are about 0.7 to 0.8 mAP on the three datasets, and the paper reports a single run for each configuration. No standard deviation, seed variation, or statistical significance test is reported. Given that the entire conclusion rests on this difference, the authors should report results over at least three seeds with mean and standard deviation, and ideally a paired comparison.
  4. [Sec. 5.4, Fig. 5] Figure 5 is presented as evidence that Str. MessDet has lower equivariance error, but the paper does not state whether the Str. curves are exactly zero or merely smaller than the Appr. curves. The y-axis resolution makes this ambiguous. The authors should report numerical values of the equivariance error for each stage of both variants and, crucially, isolate the contribution of the tuning layers by measuring the error with and without them. Without this, Figure 5 does not resolve the strictness question raised by the k=4 tuning layer.
minor comments (5)
  1. [Sec. 4.1] The choice of k=4 for the tuning layer is motivated only by output parity; the manuscript should explain why an odd-sized symmetric kernel with appropriate padding or cropping cannot serve the same purpose, since odd-sized kernels are the natural choice for exact discrete rotation equivariance.
  2. [Eq. (5)] The notation 1/(ijk) in the equivariance error definition is unclear; please define the index ranges and the normalization explicitly.
  3. [Sec. 5.2, Table 5] The number of convolutional modules in the multi-branch head is selected based on the best mAP in the ablation table; the paper should clarify whether this selection was made on the test set and discuss the potential for selection bias.
  4. [Sec. 9, Table 9] The FLOPs, FPS, and training-time differences between Str. and Appr. MessDet are useful, but the cost-benefit trade-off of strict equivariance should be discussed in the main text, especially since the parameter and FLOP increases are non-negligible.
  5. [Sec. 1, Fig. 1] The caption states that Str. MessDet 'eliminates such errors entirely,' which is stronger than anything demonstrated in the paper; please align the wording with the actual measurements.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the impact claim rests on held-out empirical comparisons, not on fitted inputs or self-citations.

full rationale

The paper's central claim is that strict rotation equivariance improves aerial detection, supported by mAP differences between Str. MessDet and Appr. MessDet on DOTA-v1.0, DOTA-v1.5, and DIOR-R. These numbers come from held-out test sets after standard training, so no test-set quantity is fitted or renamed as a prediction. The strict-versus-approximate distinction is an architectural control (presence or absence of the tuning layer before 2x downsampling), not a parameter fitted to the final mAP, and the equivariance error in Eq. (5) is defined independently of the detector loss and then measured. The paper's justification for strictness relies on an external condition (i-k) mod s = 0 attributed to [16], not on a self-citation chain, and the authors do not invoke their own prior work to forbid alternatives. The reviewer's concern that the k=4 tuning layer may not itself be rotation-equivariant on a discrete grid is a technical validity or correctness issue, not a circularity: the claim does not reduce by construction to its inputs, and the mAP advantage is not logically entailed by the architecture definition. The parameter/FLOPs difference between the two variants is an experimental confound, but confounding is not circular reasoning. Overall, no step in the derivation chain is equivalent to its own premise.

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

No new physical entities; the free parameters are architectural hyperparameters; the main axioms are the borrowed equivariance condition from [16] and the assumption that custom layers preserve it.

free parameters (3)
  • Number of orientation dimensions N = 8
    Set following ReDet [20] and FRED [26]; controls the cyclic group C8 and the grouping factor in RE-CA and multi-branch head, and was not swept in this paper.
  • Tuning layer kernel size k = 4
    Chosen as the smallest even kernel that maps even input 2n to odd output 2n-1 with p=1, s=1; this choice adds parameters compared to the approximate variant.
  • Multi-branch head number of conv modules = 3
    Selected by ablation in Table 5 as the best balance between mAP and parameters (1.5M head params); tuned on DOTA-v1.0.
assumptions (3)
  • standard math For a stride-s downsampling layer, strict rotation equivariance is maintained iff (i - k) mod s = 0 for input size i and kernel size k (Edixhoven et al. [16]).
    Used in Sec. 4.1 to justify the tuning layer design and cited from [16].
  • domain assumption The E2CNN group-convolution machinery, including a newly introduced k=4 tuning layer and the RE-CA repeated-weight scheme, exactly implements C8-equivariant maps without breaking equivariance.
    No formal proof or software-level equivariance test is provided for the custom layers; the paper reasons by parity condition and design.
  • domain assumption Random rotation augmentation during training does not eliminate the benefit of strict equivariance, so the observed mAP gap can be attributed to equivariance rather than augmentation.
    Both Appr. and Str. models use the same augmentations, but the interaction is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Measuring the Impact of Rotation Equivariance on Aerial Object Detection." pith.science (2026). https://pith.science/paper/RH2AOOAD

@misc{pith2026250709896,
  author       = {Pith},
  title        = {Pith review of: Measuring the Impact of Rotation Equivariance on Aerial Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RH2AOOAD}},
  note         = {Machine review of arXiv:2507.09896}
}
read the original abstract

Due to the arbitrary orientation of objects in aerial images, rotation equivariance is a critical property for aerial object detectors. However, recent studies on rotation-equivariant aerial object detection remain scarce. Most detectors rely on data augmentation to enable models to learn approximately rotation-equivariant features. A few detectors have constructed rotation-equivariant networks, but due to the breaking of strict rotation equivariance by typical downsampling processes, these networks only achieve approximately rotation-equivariant backbones. Whether strict rotation equivariance is necessary for aerial image object detection remains an open question. In this paper, we implement a strictly rotation-equivariant backbone and neck network with a more advanced network structure and compare it with approximately rotation-equivariant networks to quantitatively measure the impact of rotation equivariance on the performance of aerial image detectors. Additionally, leveraging the inherently grouped nature of rotation-equivariant features, we propose a multi-branch head network that reduces the parameter count while improving detection accuracy. Based on the aforementioned improvements, this study proposes the Multi-branch head rotation-equivariant single-stage Detector (MessDet), which achieves state-of-the-art performance on the challenging aerial image datasets DOTA-v1.0, DOTA-v1.5 and DIOR-R with an exceptionally low parameter count.

Figures

Figures reproduced from arXiv: 2507.09896 by the authors.

Figure 1
Figure 1. Visualization of features extracted by different meth [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Rotation-equivariant feature. The upper half illustrates [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. MessDet architecture. “Str.” and “Appr.” refer to strict [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Comparison of different downsampling processes, where [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: The impact of various rotation angles on models perfor [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Breaking and maintaining strict rotation equivariance. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Rotation-Equivariant Channel Attention Mechanism. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Illustration of Feature Rearrangement. In the head network of MessDet, features with inher￾ent group properties from different orientations are fed to different branches. To achieve this, the features are rear￾ranged along the channel dimension. The rearrangement proce…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 64 canonical work pages

  1. [16]

    van Gemert

    Tom Edixhoven, Attila Lengyel, and Jan C. van Gemert. Us- ing and abusing equivariance. In ICCVW, pages 119–128,

    work page

  2. [1]

    Poly kernel inception network for remote sensing detection

    Xinhao Cai, Qiuxia Lai, Yuwei Wang, Wenguan Wang, Zeren Sun, and Yazhou Yao. Poly kernel inception network for remote sensing detection. In CVPR, pages 27706–27716,

    work page

  3. [2]

    Hybrid task cascade for instance seg- mentation

    Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaox- iao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance seg- mentation. In CVPR, pages 4974–4983, 2019. 7

    work page 2019

  4. [3]

    A survey on object detec- tion in optical remote sensing images

    Gong Cheng and Junwei Han. A survey on object detec- tion in optical remote sensing images. ISPRS journal of pho- togrammetry and remote sensing, 117:11–28, 2016. 1

    2016

  5. [4]

    Learning rotation-invariant convolutional neural networks for object detection in vhr optical remote sensing images

    Gong Cheng, Peicheng Zhou, and Junwei Han. Learning rotation-invariant convolutional neural networks for object detection in vhr optical remote sensing images. IEEE TGRS, 54(12):7405–7415, 2016. 2, 3

    work page 2016

  6. [5]

    Anchor-free oriented proposal generator for object detection

    Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE TGRS, 60:1– 11, 2022. 1, 2

    work page 2022

  7. [6]

    Dual- aligned oriented detector

    Gong Cheng, Yanqing Yao, Shengyang Li, Ke Li, Xingx- ing Xie, Jiabao Wang, Xiwen Yao, and Junwei Han. Dual- aligned oriented detector. IEEE TGRS , 60:1–11, 2022. 1, 2

    work page 2022

  8. [7]

    Towards large-scale small object detection: Survey and benchmarks

    Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE TPAMI, 45(11):13467–13488, 2023. 1

    work page 2023

Show all 69 references
  1. [8]

    Group equivariant convolu- tional networks

    Taco Cohen and Max Welling. Group equivariant convolu- tional networks. In ICML, pages 2990–2999, 2016. 2, 3, 4

    2016

  2. [9]

    Steerable cnns

    Taco Cohen and Max Welling. Steerable cnns. In ICLR,

  3. [10]

    Spherical cnns

    Taco Cohen, Mario Geiger, Jonas K ¨ohler, and Max Welling. Spherical cnns. In ICLR, 2018. 3

    2018

  4. [11]

    MMYOLO: OpenMMLab YOLO series toolbox and benchmark

    MMYOLO Contributors. MMYOLO: OpenMMLab YOLO series toolbox and benchmark. https://github.com/ open-mmlab/mmyolo, 2022. 6, 2

    2022

  5. [12]

    Ao2-detr: Arbitrary-oriented object detection trans- former

    Linhui Dai, Hong Liu, Hao Tang, Zhiwei Wu, and Pinhao Song. Ao2-detr: Arbitrary-oriented object detection trans- former. IEEE TCSVT, 33(5):2342–2356, 2023. 1

    2023

  6. [13]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6, 7

    2009

  7. [14]

    Learning roi transformer for oriented object detection in aerial images

    Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In CVPR, pages 2849–2858, 2019. 1, 3, 6, 7

    2019

  8. [15]

    Object de- tection in aerial images: A large-scale benchmark and chal- lenges

    Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Object de- tection in aerial images: A large-scale benchmark and chal- lenges. IEEE TPAMI, 44(11):7778–7796, 2022. 2, 5, 7

    2022

  9. [17]

    Point- based estimator for arbitrary-oriented object detection in aerial images

    Kun Fu, Zhonghan Chang, Yue Zhang, and Xian Sun. Point- based estimator for arbitrary-oriented object detection in aerial images. IEEE TGRS, 59(5):4370–4387, 2021. 6

    2021

  10. [18]

    Rich feature hierarchies for accurate object detection and semantic segmentation

    Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, pages 580–587, 2014. 1

    2014

  11. [19]

    Beyond bounding-box: Convex- hull feature adaptation for oriented and densely packed ob- ject detection

    Zonghao Guo, Chang Liu, Xiaosong Zhang, Jianbin Jiao, Xi- angyang Ji, and Qixiang Ye. Beyond bounding-box: Convex- hull feature adaptation for oriented and densely packed ob- ject detection. In CVPR, pages 8792–8801, 2021. 1, 3, 2

    2021

  12. [20]

    Redet: A rotation-equivariant detector for aerial object detection

    Jiaming Han, Jian Ding, Nan Xue, and Gui-Song Xia. Redet: A rotation-equivariant detector for aerial object detection. In CVPR, pages 2786–2795, 2021. 2, 3, 6, 7

    2021

  13. [21]

    Align deep features for oriented object detection

    Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE TGRS, 60: 1–11, 2022. 1, 3, 6, 2

    2022

  14. [22]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Gir- shick. Mask r-cnn. In ICCV, pages 2961–2969, 2017. 3, 7

    2017

  15. [23]

    Shape-adaptive selection and measurement for oriented object detection

    Liping Hou, Ke Lu, Jian Xue, and Yuqiu Li. Shape-adaptive selection and measurement for oriented object detection. In AAAI, pages 923–932, 2022. 6, 2

    2022

  16. [24]

    G- rep: Gaussian representation for arbitrary-oriented object de- tection

    Liping Hou, Ke Lu, Xue Yang, Yuqiu Li, and Jian Xue. G- rep: Gaussian representation for arbitrary-oriented object de- tection. Remote Sensing, 15(3):757, 2023. 1, 6

    2023

  17. [25]

    Squeeze-and-excitation net- works

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation net- works. In CVPR, pages 7132–7141, 2018. 2, 5

    2018

  18. [26]

    Fred: Towards a full rotation-equivariance in aerial image object detection

    Chanho Lee, Jinsu Son, Hyounguk Shon, Yunho Jeon, and Junmo Kim. Fred: Towards a full rotation-equivariance in aerial image object detection. In AAAI, pages 2883–2891,

  19. [27]

    Object detection in optical remote sensing images: A survey and a new benchmark

    Ke Li, Gang Wan, Gong Cheng, Liqiu Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS Journal of Photogram- metry and Remote Sensing , 159:296–307, 2020. 1, 2, 5, 7, 8

    2020

  20. [28]

    Ori- ented reppoints for aerial object detection

    Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Ori- ented reppoints for aerial object detection. In CVPR, pages 1829–1838, 2022. 1, 3, 6, 7

    2022

  21. [29]

    Large selective kernel network for remote sensing object detection

    Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. In ICCV, pages 16794– 16805, 2023. 3, 6, 7, 2

    2023

  22. [30]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117–2125,

  23. [31]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2980–2988, 2017. 7

    2017

  24. [32]

    Path aggregation network for instance segmentation

    Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, pages 8759–8768, 2018. 4

    2018

  25. [33]

    A high resolution optical satellite image dataset for ship recognition and some new baselines

    Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In ICPRAM, pages 324–331, 2017. 1

    2017

  26. [34]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6, 2

    2017 arXiv

  27. [35]

    Rtmdet: An empirical study of designing real-time object detectors

    Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 2, 4, 5, 6, 7, 8

    2022 arXiv

  28. [36]

    A data and compute efficient design for limited- resources deep learning

    Mirgahney Mohamed, Gabriele Cesa, Taco Cohen, and Max Welling. A data and compute efficient design for limited- resources deep learning. arXiv preprint arXiv:2004.09691 ,

    2004 arXiv

  29. [37]

    Dynamic refinement network for oriented and densely packed object detection

    Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, and Chang- sheng Xu. Dynamic refinement network for oriented and densely packed object detection. In CVPR, pages 11207– 11216, 2020. 1, 2, 6

    2020

  30. [38]

    Adaptive rotated convolution for rotated object de- tection

    Yifan Pu, Yiru Wang, Zhuofan Xia, Yizeng Han, Yulin Wang, Weihao Gan, Zidong Wang, Shiji Song, and Gao Huang. Adaptive rotated convolution for rotated object de- tection. In ICCV, pages 6589–6600, 2023. 3, 6

    2023

  31. [39]

    You only look once: Unified, real-time object de- tection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object de- tection. In CVPR, pages 779–788, 2016. 1

    2016

  32. [40]

    Faster r-cnn: Towards real-time object detection with region proposal networks

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137–1149, 2017. 7

    2017

  33. [41]

    Attentive group equivariant convolutional networks

    David Romero, Erik Bekkers, Jakub Tomczak, and Mark Hoogendoorn. Attentive group equivariant convolutional networks. In ICML, pages 8188–8199, 2020. 4

    2020

  34. [42]

    Fair1m: A benchmark dataset for fine- grained object recognition in high-resolution remote sens- ing imagery

    Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine- grained object recognition in high-resolution remote sens- ing imagery. ISPRS Journal of Photogrammetry and Remote Se...

    2022

  35. [43]

    Cspnet: A new backbone that can enhance learning capability of cnn

    Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh, and I-Hau Yeh. Cspnet: A new backbone that can enhance learning capability of cnn. In CVPRW, pages 390–391, 2020. 4

    2020

  36. [44]

    Advancing plain vision transformer toward remote sensing foundation model

    Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE TGRS, 61:1–15, 2023. 1

    2023

  37. [45]

    Learning center probability map for de- tecting objects in aerial images

    Jinwang Wang, Wen Yang, Heng-Chao Li, Haijian Zhang, and Gui-Song Xia. Learning center probability map for de- tecting objects in aerial images. IEEE TGRS , 59(5):4307– 4323, 2021. 1, 3, 6

    2021

  38. [46]

    General e(2)-equivariant steerable cnns

    Maurice Weiler and Gabriele Cesa. General e(2)-equivariant steerable cnns. In NeurIPS, page 14334–14345, 2019. 2, 3, 4

    2019

  39. [47]

    Hamprecht, and Martin Storath

    Maurice Weiler, Fred A. Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In CVPR, pages 849–858, 2018. 3

    2018

  40. [48]

    Cbam: Convolutional block attention module

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 2, 5

    2018

  41. [49]

    Worrall, Stephan J

    Daniel E. Worrall, Stephan J. Garbin, Daniyar Turmukham- betov, and Gabriel J. Brostow. Harmonic networks: Deep translation and rotation equivariance. In CVPR, pages 5028– 5037, 2017. 3

    2017

  42. [50]

    Dota: A large-scale dataset for object detection in aerial images

    Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Be- longie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liang- pei Zhang. Dota: A large-scale dataset for object detection in aerial images. In CVPR, pages 3974–3983, 2018. 1, 5, 6

    2018

  43. [51]

    Oriented r-cnn for object detection

    Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In ICCV, pages 3520–3529, 2021. 1, 3, 6, 2

    2021

  44. [52]

    Dynamic coarse-to-fine learning for oriented tiny object detection

    Chang Xu, Jian Ding, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Dynamic coarse-to-fine learning for oriented tiny object detection. In CVPR, pages 7318– 7328, 2023. 7

    2023

  45. [53]

    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. 2, 5

    2022 arXiv

  46. [54]

    Gliding vertex on the horizontal bounding box for multi-oriented object detec- tion

    Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on the horizontal bounding box for multi-oriented object detec- tion. IEEE TPAMI, 43(4):1452–1459, 2021. 1, 3, 6

    2021

  47. [55]

    Arbitrary-oriented object detec- tion with circular smooth label

    Xue Yang and Junchi Yan. Arbitrary-oriented object detec- tion with circular smooth label. In ECCV, pages 677–694,

  48. [56]

    Scrdet: Towards more robust detection for small, cluttered and rotated objects

    Xue Yang, Jirui Yang, Junchi Yan, Yue Zhang, Tengfei Zhang, Zhi Guo, Xian Sun, and Kun Fu. Scrdet: Towards more robust detection for small, cluttered and rotated objects. In ICCV, pages 8232–8241, 2019. 1, 3, 6, 2

    2019

  49. [57]

    R3det: Refined single-stage detector with feature refinement for ro- tating object

    Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for ro- tating object. In AAAI, pages 3163–3171, 2021. 1, 2, 6

    2021

  50. [58]

    Rethinking rotated object detection with gaussian wasserstein distance loss

    Xue Yang, Junchi Yan, Qi Ming, Wentao Wang, Xiaopeng Zhang, and Qi Tian. Rethinking rotated object detection with gaussian wasserstein distance loss. In ICML, pages 11830– 11841, 2021. 2, 3, 6

    2021

  51. [59]

    Learning high-precision bounding box for rotated object detection via kullback- leibler divergence

    Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, and Junchi Yan. Learning high-precision bounding box for rotated object detection via kullback- leibler divergence. In NeurIPS, pages 18381–18394, 2021. 2, 3, 6

    2021

  52. [60]

    Scrdet++: Detecting small, cluttered and rotated objects via instance-level feature denoising and rota- tion loss smoothing

    Xue Yang, Junchi Yan, Wenlong Liao, Xiaokang Yang, Jin Tang, and Tao He. Scrdet++: Detecting small, cluttered and rotated objects via instance-level feature denoising and rota- tion loss smoothing. IEEE TPAMI, 45(2):2384–2399, 2023. 3

    2023

  53. [61]

    The kfiou loss for rotated object detection

    Xue Yang, Yue Zhou, Gefan Zhang, Jirui Yang, Wentao Wang, Junchi Yan, Xiaopeng Zhang, and Qi Tian. The kfiou loss for rotated object detection. In ICLR, 2023. 2

    2023

  54. [62]

    Reppoints: Point set representation for object detection

    Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In ICCV, pages 9657–9666, 2019. 2

    2019

  55. [63]

    Strip r-cnn: Large strip convolution for remote sensing object detection

    Xinbin Yuan, ZhaoHui Zheng, Yuxuan Li, Xialei Liu, Li Liu, Xiang Li, Qibin Hou, and Ming-Ming Cheng. Strip r-cnn: Large strip convolution for remote sensing object detection. arXiv preprint arXiv:2501.03775, 2025. 3

    2025 arXiv

  56. [64]

    Mmrotate: A rotated object detection benchmark using pytorch

    Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. Mmrotate: A rotated object detection benchmark using pytorch. In ACM MM, page 7331–7334, 2022. 6, 2

    2022

  57. [65]

    Deep learning in remote sensing: A comprehensive review and list of resources

    Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep learning in remote sensing: A comprehensive review and list of resources. IEEE Geoscience and Remote Sensing Maga- zine, 5(4):8–36, 2017. 1

    2017

  58. [66]

    Object detection in 20 years: A survey

    Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proc. IEEE, 111(3):257–276, 2023. 1 Measuring the Impact of Rotation Equivariance on Aerial Object Detection Supplementary Material

    2023

  59. [67]

    An Example Where Downsampling Breaks Rotation Equivariance In Section 3.2, we describe how conventional downsam- pling layers can break strict rotation equivariance. This occurs because the center sampling points of the convolu- tion kernels do not match before and after rotat...

  60. [68]

    The mathematical formulation of RE-CA is provided in Section 4.2, and its schematic diagram is shown in Fig

    The Further Details of MessDet This paper introduces the rotation-equivariant channel at- tention (RE-CA), enabling rotation-equivariant networks to be implemented with more advanced network structures. The mathematical formulation of RE-CA is provided in Section 4.2, and its ...

  61. [69]

    Experiments Details and The Further Ex- periments Our model is implemented using the MMYOLO [11] and MMRotate [64] frameworks and trained for 36 epochs on DOTA-v1.0, DOTA-v1.5 and DIOR-R. During training, we followed most mainstream methods [1, 20, 29, 51] by em- MessDet(witho...

Pith tools

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