RiO-DETR: DETR for Real-time Oriented Object Detection

Dongsheng Jiang; Hebei Li; Jie Chen; Peixi Wu; Wenzhang Sun; Xiangchen Yin; Xiaoyan Sun; Xinghao Wang; Yansong Peng; Yifan Zhao

REVIEW 3 major objections 7 minor 18 references

RiO-DETR: DETR for Real-time Oriented Object Detection

T0 review · 3 major / 7 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read RiO-DETR claims the first end-to-end transformer that detects rotated objects in real time, matching CNN latency while exceeding its accuracy.

desk verdict A capable real-time oriented DETR, but the speed-accuracy headline rests on a community YOLO baseline and unmatched latency protocols. read the letter →

arxiv 2603.09411 v2 pith:YETBHJ6F submitted 2026-03-10 cs.CV

Zhangchi Hu , Yifan Zhao , Yansong Peng , Wenzhang Sun , Xiangchen Yin , Jie Chen , Peixi Wu , Hebei Li

show 3 more authors

Xinghao Wang Dongsheng Jiang Xiaoyan Sun

This is my paper

classification cs.CV

keywords orientedobjectdetectionDETRreal-timeangleperiodicitycontent-drivenestimationperiodicrefinementdensesupervisionaerialimagery

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

RiO-DETR sets out to prove that transformer-based detection can be just as fast as CNN detectors for oriented (rotated) bounding boxes, not only more accurate. The authors identify three reasons oriented DETRs have been slow and inaccurate: angle is driven by appearance rather than geometry; angle periodicity breaks the usual Euclidean refinement; and the larger search space slows convergence. They redesign the decoder around these issues — decoupling angle from positional queries, using orthogonal attention heads, a bounded coarse-to-fine periodic refinement with a shortest-path loss, and rotated quadrant stitching for dense supervision. If the paper is right, it closes a long-standing gap: real-time oriented detection no longer requires NMS-based CNN pipelines, and end-to-end transformers can hold the speed–accuracy frontier on aerial, remote-sensing, and scene-text benchmarks.

What carries the argument

The load-bearing mechanism is the π-periodic quotient structure of oriented boxes, respected at every stage: positional queries encode only center and size, leaving angle to content (Geometry-Decoupled Query Encoding); attention heads split symmetrically between θ and θ+π/2 sampling (Rotation-Rectified Orthogonal Attention); decoder updates are tanh-bounded, layer-wise decayed, and wrapped modulo π, trained with the shortest-path periodic L1 loss (Decoupled Periodic Refinement); and dense supervision uses quadrant-wise 90° rotations for angular diversity (Oriented Dense O2O). All four pieces add essentially no parameters or FLOPs.

What would settle it

Re-benchmark the strongest CNN baseline and RiO-DETR-x on the same T4 with the official weights and TensorRT 10 FP16 engine; if the baseline matches or exceeds RiO-DETR-x's 81.8 AP50 at 30.5 ms, the claimed frontier is a benchmark artifact.

Watch

Extended reading notes

Core claim

Oriented DETR fails for three fixable reasons, and treating angle as a content-inferred, π-periodic quantity rather than a geometric coordinate resolves all three. Removing θ from positional queries beats every periodic positional encoding tested, showing orientation is disambiguated by appearance. Orthogonal attention heads sample the major axis and its perpendicular; a bounded, coarse-to-fine periodic update paired with a shortest-path L1 loss makes refinement seam-consistent; and quadrant-wise rotation of stitched training images adds angular diversity to dense supervision at zero cost. The sum is 78.4 AP50 at 2.7 ms and 81.8 AP50 at 29.9 ms on DOTA-1.0, better than prior real-time CNN de

Load-bearing premise

The headline speed–accuracy frontier rests on a single fair measurement setup, and one of the strongest CNN baselines is a community reimplementation rather than an official engine, so the reported margins may shift under independent measurement.

Editorial extensions

If this is right

Real-time oriented detection no longer requires CNN architectures with NMS; end-to-end transformers can match or beat them in speed and accuracy on DOTA-1.0, DIOR-R, and FAIR-1M-2.0.
Angle should be treated as content-inferred rather than a geometric prior: removing θ from positional queries beat every periodic positional encoding tested, suggesting a general design rule for rotated detection.
The bounded coarse-to-fine periodic refinement stabilizes angle learning across seams; layer-wise angular corrections in the trained model decay exponentially, confirming the intended schedule.
The speed–accuracy gains hold across model scales and backbones, so the recipe is portable beyond the specific configuration reported.
Oriented Dense O2O cuts the epochs to peak accuracy (60 vs 94 without augmentation on DIOR-R), reducing training cost while improving AP50.

Reading between the lines

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

If angle decoupling is right, other geometry variables that are annotation-convention-dependent (e.g., long-side vs short-side aspect ratio) might also be better left out of positional queries for ambiguous instances.
The orthogonal head-splitting could be made adaptive: for near-square objects, a 4:4 split across θ and θ+π/2 is optimal in the paper's ablations; a query-dependent split based on predicted aspect ratio might gain further.
The shortest-path periodic loss is a drop-in for any cyclic regression target in end-to-end detectors (heading, pose, phase); the paper tests it only for oriented boxes.
Quadrant-wise discrete rotation for dense supervision is a zero-cost data organization trick; it could be combined with stronger augmentation policies for rotation-sensitive tasks beyond aerial imagery.

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.

Desk Editor's Note

read the letter

RiO-DETR is the paper to read if you care about real-time oriented detection. It does three things well: it decouples angle from positional queries, splits attention heads orthogonally, and makes the refinement periodic. Each is ablated cleanly on DIOR-R, with separate tables showing each step helps. The short-path periodic loss and bounded coarse-to-fine updates are a sensible response to angle periodicity, and the ablations actually show that the periodic update alone can hurt without the matching loss. That's honest engineering.

The bigger question is the headline: 'first real-time oriented DETR' and 'new speed-accuracy frontier.' That claim depends on the YOLO26-obb comparisons, and those have two issues. First, Fig.1 marks YOLO26-obb as 'community implemented version,' but the tables cite it as if official. If those numbers come from an unofficial reimplementation, the frontier could shift. Second, the latency protocol isn't apples-to-apples: RiO-DETR is called end-to-end but excludes pre/post-processing, while YOLO uses the BatchedNMSPlugin. That's a few milliseconds difference at the nano scale, where the claim is 2.7ms. I'd want full-pipeline latency for both.

Also, the high-end margin over YOLO26x on DOTA multi-scale is 0.06 AP50 — well within run-to-run noise in oriented detection. No seeds or error bars anywhere. That doesn't kill the paper, but it tempers 'superior' to 'comparable at best until re-measured.'

The 'first real-time oriented DETR' priority claim needs a more careful prior-art check as well. It may well be true, but it's asserted rather than demonstrated.

All that said, the work is serious. The ablations are well-structured, the design decisions are motivated by the geometry of the problem, and the authors are explicit about limitations (the backbone remains open). I'd send this to a competent referee for a full review. The conditions are: release code, clarify the YOLO26-obb source, report multi-seed variance, and either measure everyone's full pipeline or clearly scope what "end-to-end" means. If those hold, this is a real contribution to real-time oriented detection.

Referee Report

3 major / 7 minor

Summary. RiO-DETR is a real-time oriented object detection transformer built on an RT-DETRv2-style baseline. It introduces three task-specific designs: Content-Driven Angle Estimation (decoupling θ from positional query encoding, plus Rotation-Rectified Orthogonal Attention), Decoupled Periodic Refinement (bounded coarse-to-fine periodic angle updates with a shortest-path periodic L1 loss), and Oriented Dense O2O (random quadrant rotations in dense supervision). Experiments on DOTA-1.0, DIOR-R, and FAIR-1M-2.0 report state-of-the-art speed-accuracy trade-offs, e.g., RiO-DETR-n at 78.4 AP50 / 2.7 ms and RiO-DETR-x at 81.8 AP50 / 29.9 ms on DOTA-1.0, outperforming YOLO26-obb variants at similar latencies. The paper claims to be the first real-time oriented DETR and attributes the gains to the three proposed components, supported by ablations on DIOR-R.

Significance. If the reported comparisons hold, this is a meaningful contribution: it demonstrates that an end-to-end transformer detector can operate in the real-time regime for oriented object detection, narrowing the gap with CNN-based detectors. The three components are well-motivated, and the ablations are structured to isolate each design choice. Consistency of improvements across three large-scale benchmarks strengthens the empirical case. The authors promise code release, which is important for reproducibility. However, the headline speed-accuracy frontier rests on fair latency and baseline-parity assumptions that are not fully established in the manuscript. The internal ablations are largely coherent and support incremental gains, but the external validity of the frontier claim needs stronger evidence.

major comments (3)

[Sec. 4.2, Fig. 1, Tables 1-2] The latency-comparison protocol is not apples-to-apples. Sec. 4.2 states that 'all latencies are measured equally' but then says that for end-to-end models latency excludes pre/post-processing while for other models the BatchedNMSPlugin is employed. The headline 2.7 ms for RiO-DETR-n may exclude preprocessing steps such as letterboxing and normalization, while YOLO26-obb latencies include NMS post-processing. This asymmetry directly affects the claimed frontier, especially because the multi-scale margin over YOLO26x-obb is only 0.06 AP50 (Table 2). In addition, Fig. 1 marks YOLO26-obb as a 'community implemented version' but no source or version is specified; citing the YOLO26 paper does not establish that the OBB variant is the official one. Please measure full-pipeline latency (or explicitly quantify pre/post-processing overhead) for all models, specify the exact YOLO26-obb implementat
[Tables 1-2] All accuracy numbers appear to be from a single run with no error bars or seed variance. The multi-scale DOTA-1.0 comparison shows RiO-DETR-x at 81.76 AP50 vs YOLO26x-obb at 81.70 — a 0.06 AP50 difference that is within typical run-to-run variation for oriented detection training. Without multiple seeds or a statistical test, the claimed superiority at that operating point is not established. Please report mean ± std over at least 3 seeds for the key comparisons, or at least confirm that the reported margins are reproducible. This is particularly important for the x-scale model where the gain is small.
[Tables 5, 7, 8] The ablation stages are not consistently defined across tables. Table 5 shows Oriented Dense O2O improving from 75.46 to 75.73, while Table 7 reports Oriented Dense O2O at 73.88 AP50. Similarly, Table 8 shows that combining SP-L1 and the periodic update raises AP50 from 74.18 to 74.74, which exactly matches the Table 5 step where Rotation-Rectified Orthogonal Attention is added on top of geometry-decoupled queries (74.18 → 74.74). It appears that the ablations in Tables 7-8 are run on a different base model than the roadmap in Table 5. Please specify the exact base model for each ablation, and ideally re-run ablations on the same full model so that the incremental contribution of each component is directly attributable.

minor comments (7)

[Sec. 3.2] The text states that Rotation-Rectified Orthogonal Attention introduces 'no additional parameters or GFlops,' but Table 5 shows FLOPs increasing from 97.01 to 97.18 G when this component is added. The overhead is small, but the wording is inaccurate. Suggest qualifying it as 'negligible overhead' and reporting the exact change.
[Eq. (7)] The periodic normalization in Eq. (7) is confusing: it checks whether (θ_raw mod π) < 0 and adds π, but the standard modulo operation with positive divisor is already nonnegative. Please define wrapπ explicitly as in Appendix C and use it consistently.
[Appendix D, Table 3] The row labeled 'YOLO26m-obb' in Appendix D Table 3 lists the backbone as YOLO26l (25.6M params, 230G FLOPs). This is likely a typo; the main text Table 2 also lists YOLO26m-obb. Please correct the labeling.
[References] Reference [55] has a malformed author entry ('Chen, W., , G., Song, Y.'). Please fix the author list. Also, several references are incomplete (e.g., missing page numbers for some arXiv papers), though this is a minor formatting issue.
[Appendix B] The final sentence says 'Our claims above are further proved by Sec. 4.6 in the main text, Appendix E.3 and F.' Visualizations and ablations do not constitute a proof; please rephrase to 'supported by' or 'corroborated by.'
[Fig. 1 caption] The caption says '* denotes a community implemented version' but does not say whose implementation or how to obtain it. Since YOLO26-obb is a key baseline, please provide a link or specification of the implementation used.
[Sec. 1] The paper repeatedly uses the term 'real-time' without a quantitative definition. On T4, 29.9 ms corresponds to ~33 FPS, which may or may not be considered real-time depending on the application. Please state the latency threshold or FPS criterion used to define 'real-time.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical evaluations of architecture components on public benchmarks, and none of the reported predictions reduce by construction to fitted parameters or self-citations.

full rationale

RiO-DETR does not claim a first-principles derivation chain. Its central claim is a speed-accuracy balance (Sec. 1) supported by measured AP50 and latency on DOTA-1.0, DIOR-R, and FAIR-1M-2.0 (Tables 1-4). The three proposed components are design choices with explicit ablations (Tables 5-9 and Appendix Tables 5-8). The shortest-path periodic L1 loss (Eq. 8) is a standard circular-distance loss in oriented detection; the paper cites prior periodic-angle literature ([5,51-54]) and does not present the loss itself as a derived prediction. The bounded periodic update (Eqs. 5-7) is an architectural mechanism, and its decay factor alpha0 is selected by ablation (Appendix Table 5) on DIOR-R and then used on other test splits, which is normal hyperparameter selection rather than fitting a parameter to the headline result. Self-citations to D-FINE and DOME-DETR appear as building blocks or related-work context, but D-FINE is a published, externally evaluated method used to construct the baseline, and the baseline is explicitly shown to remain 1.18 AP50 below state of the art (Table 6), so the central improvement is not carried by the self-citation. The benchmark-parity concerns about YOLO26-obb being a community implementation and about latency excluding pre/post-processing for end-to-end models are important external-validity questions, but they are not circularity: they concern whether comparisons are fair, not whether the paper's derivations reduce to their inputs. No circular step can be exhibited from the paper's equations or citations.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's gains are empirical and rest on a handful of hyperparameters (α0, head-split ratio, loss weights) chosen by ablation on DIOR-R. There are no new physical entities. The main domain assumption — that angle should be inferred from content rather than positional geometry — is supported by ablations but remains a modeling hypothesis.

free parameters (3)

α0 (decay base for Decoupled Periodic Refinement) = 1.5
Selected by validation on DIOR-R (Appendix F, Table 5); controls the coarse-to-fine angle step schedule and directly affects convergence/accuracy.
Attention head split ratio (θ : θ+π/2) = 4:4
Ablated in Appendix F, Table 7; symmetric orthogonal split chosen as best. This is a free design choice tuned on DIOR-R.
Loss weights λ_L1, λ_KLD, λ_angle = 5, 5, 1 (default)
Fixed across all models following RT-DETRv2/RHINO practice (Appendix A.2); not tuned in this paper but still free parameters the result depends on.

assumptions (4)

domain assumption OBB orientation is a content/semantics-disambiguated quantity, so angle should be learned from content features rather than positional queries (Sec. 3.2, Appendix B).
Load-bearing design premise; if orientation is actually well predicted from geometry, the main angle-decoupling gain would not hold. Supported only by ablation, not a theorem.
domain assumption Angle space is π-periodic with long-side canonicalization θ∈[0,π), and the shortest-path periodic distance is the correct training metric (Sec. 3.3, Appendix C).
Assumes the benchmark annotation convention and that geodesic loss is optimal; consistent with prior oriented-detection literature but still a modeling assumption.
domain assumption Stitching four rotated copies of an image for dense supervision does not introduce harmful domain shift or boundary artifacts (Sec. 3.4).
The augmentation's benefit is shown empirically; the assumption that it always helps across datasets/backbones is not proven.
domain assumption The benchmarks (DOTA-1.0, DIOR-R, FAIR-1M-2.0) are labeled consistently and the official evaluation servers are ground truth (Sec. 4.1).
Standard for the field; the reported AP50 numbers depend on this.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RiO-DETR: DETR for Real-time Oriented Object Detection." pith.science (2026). https://pith.science/paper/YETBHJ6F

@misc{pith2026260309411,
  author       = {Pith},
  title        = {Pith review of: RiO-DETR: DETR for Real-time Oriented Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YETBHJ6F}},
  note         = {Machine review of arXiv:2603.09411}
}

read the original abstract

We present RiO-DETR: DETR for Real-time Oriented Object Detection, the first real-time oriented detection transformer to the best of our knowledge. Adapting DETR to oriented bounding boxes (OBBs) poses three challenges: semantics-dependent orientation, angle periodicity that breaks standard Euclidean refinement, and an enlarged search space that slows convergence. RiO-DETR resolves these issues with task-native designs while preserving real-time efficiency. First, we propose Content-Driven Angle Estimation by decoupling angle from positional queries, together with Rotation-Rectified Orthogonal Attention to capture complementary cues for reliable orientation. Second, Decoupled Periodic Refinement combines bounded coarse-to-fine updates with a Shortest-Path Periodic Loss for stable learning across angular seams. Third, Oriented Dense O2O injects angular diversity into dense supervision to speed up angle convergence at no extra cost. Extensive experiments on DOTA-1.0, DIOR-R, and FAIR-1M-2.0 demonstrate RiO-DETR establishes a new speed--accuracy trade-off for real-time oriented detection. GitHub Repository: https://github.com/RicePasteM/RiO-DETR.

Figures

Figures reproduced from arXiv: 2603.09411 by the authors.

**Figure 1.** Comparisons with other detectors in terms of model size (left), latency (mid), and computational cost (right) on DOTA-1.0 under single-scale training and testing protocol. * denotes a community implemented version. Abstract. We present RiO-DETR : DETR for Real-t ime Oriented Object Detection, the first real-time oriented detection transformer to the best of our knowledge. Adapting DETR to oriented bounding boxes (OB… view at source ↗

**Figure 2.** The main architecture of our proposed RiO-DETR. The framework highlights three key components: (A) Content-Driven Angle Estimation, (B) Rotation-Rectified Orthogonal Attention, and (C) Decoupled Periodic Refinement. 3.2 Content-Driven Angle Estimation Existing DETR-based oriented object detectors treat the orientation θ as a geometric component symmetric to box coordinates (cx, cy, w, h). Consequently, these method… view at source ↗

**Figure 3.** An intuitive illustration of (D) Decoupled Periodic Refinement and (E) Oriented Dense O2O. S(\mathbf {p}_q, \Delta \mathbf {p}_{qk}) = \mathbf {p}_{q} + \mathbf {R}(\theta ^{(h)}) (\Delta \mathbf {p}_{qk} \odot \mathbf {s}_q) (3) where ⊙ denotes the element-wise multiplication, and R(θ (h) ) is the rotation matrix derived from the head-specific angle: \mathbf {R}(\theta ^{(h)}) = \begin {bmatrix} \cos (\theta ^{(h)}… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Visualization of deformable attention sampling points on DOTA. Decoder Layer 1 Output Decoder Layer 2 Output Decoder Layer 3 Output Decoder Layer 4 Output [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

**Figure 6.** Figure 6: t-SNE visualization of angular features, demonstrating GeometryDecoupled Query Encoding’s clustering. geometric prior, which is consistent with our design choice to avoid injecting θ as positional geometry. Further t-SNE analyses are provided in Appendix E. 5 Conclus…

**Figure 1.** Figure 1: Visualizations of layer-wise angular refinement for square-like instances. 0 10 20 30 40 50 60 70 Epoch 0.2 0.3 0.4 0.5 0.6 0.7 AP50 RiO-DETR-m w/o Dense O2O w/o Oriented Dense O2O [PITH_FULL_IMAGE:figures/full_fig_p026_1.png]

**Figure 3.** Figure 3: Per-category t-SNE visualizations with orientation coloring on FAIR-1M-2.0. lanes), we occasionally observe mild angle-correlated gradients or sub-structures, which is expected given dataset bias and scene geometry. Overall, these visualizations qualitatively support …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 5 linked inside Pith

[1]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Cai, X., Lai, Q., Wang, Y., Wang, W., Sun, Z., Yao, Y.: Poly kernel inception network for remote sensing detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 27706–27716 (2024)

2024
[2]

IEEE transactions on circuits and systems for video tech- nology33(5), 2342–2356 (2022)

Dai, L., Liu, H., Tang, H., Wu, Z., Song, P.: Ao2-detr: Arbitrary-oriented object detection transformer. IEEE transactions on circuits and systems for video tech- nology33(5), 2342–2356 (2022)

2022
[3]

In: 2009 IEEE conference on computer vision and pattern recognition

Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248–255. Ieee (2009)

2009
[4]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Ding, J., Xue, N., Long, Y., Xia, G.S., Lu, Q.: Learning roi transformer for oriented object detection in aerial images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2849–2858 (2019)

2019
[5]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Han, J., Ding, J., Xue, N., Xia, G.S.: Redet: A rotation-equivariant detector for aerial object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2786–2795 (2021)

2021
[6]

In: Proceedings of the computer vision and pattern recognition conference

Huang, S., Lu, Z., Cun, X., Yu, Y., Zhou, X., Shen, X.: Deim: Detr with improved matching for fast convergence. In: Proceedings of the computer vision and pattern recognition conference. pp. 15162–15171 (2025)

2025
[7]

In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

Lee, H., Song, M., Koo, J., Seo, J.: Hausdorff distance matching with adaptive query denoising for rotated detection transformer. In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1872–1882. IEEE (2025)

2025
[8]

li et al

Li, Y., Li, X., Dai, Y., Hou, Q., Liu, L., Liu, Y., Cheng, M.M., Yang, J.: Lsknet: A foundation lightweight backbone for remote sensing: Y. li et al. International Journal of Computer Vision133(3), 1410–1431 (2025)

2025

Show all 18 references

[9]

arXiv preprint arXiv:2407.17140 (2024)

Lv, W., Zhao, Y., Chang, Q., Huang, K., Wang, G., Liu, Y.: Rt-detrv2: Improved baseline with bag-of-freebies for real-time detection transformer. arXiv preprint arXiv:2407.17140 (2024)

2024 arXiv
[10]

arXiv preprint arXiv:2212.07784 (2022)

Lyu, C., Zhang, W., Huang, H., Zhou, Y., Wang, Y., Liu, Y., Zhang, S., Chen, K.: Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784 (2022)

2022 arXiv
[11]

arXiv preprint arXiv:2410.13842 (2024)

Peng, Y., Li, H., Wu, P., Zhang, Y., Sun, X., Wu, F.: D-fine: Redefine re- gression task in detrs as fine-grained distribution refinement. arXiv preprint arXiv:2410.13842 (2024)

2024 arXiv
[12]

arXiv preprint arXiv:2509.25164 (2025)

Sapkota, R., Cheppally, R.H., Sharda, A., Karkee, M.: Yolo26: key architectural enhancements and performance benchmarking for real-time object detection. arXiv preprint arXiv:2509.25164 (2025)

2025
[13]

In: Proceedings of the IEEE/CVF international conference on computer vision

Xie, X., Cheng, G., Wang, J., Yao, X., Han, J.: Oriented r-cnn for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3520–3529 (2021)

2021
[14]

arXiv preprint arXiv:2203.16250 (2022)

Xu, S., Wang, X., Lv, W., Chang, Q., Cui, C., Deng, K., Wang, G., Dang, Q., Wei, S., Du, Y., et al.: Pp-yoloe: An evolved version of yolo. arXiv preprint arXiv:2203.16250 (2022)

2022 arXiv
[15]

arXiv preprint arXiv:2501.03775 (2025)

Yuan, X., Zheng, Z., Li, Y., Liu, X., Liu, L., Li, X., Hou, Q., Cheng, M.M.: Strip r-cnn: Large strip convolution for remote sensing object detection. arXiv preprint arXiv:2501.03775 (2025)

2025 arXiv
[16]

IEEE transactions on geoscience and remote sensing62, 1–15 (2024) 12

Zeng, Y., Chen, Y., Yang, X., Li, Q., Yan, J.: Ars-detr: Aspect ratio-sensitive detection transformer for aerial oriented object detection. IEEE transactions on geoscience and remote sensing62, 1–15 (2024) 12

2024
[17]

In: European conference on computer vision

Zhao, Z., Xue, Q., He, Y., Bai, Y., Wei, X., Gong, Y.: Projecting points to axes: Oriented object detection via point-axis representation. In: European conference on computer vision. pp. 161–179. Springer (2024)

2024
[18]

In: Proceedings of the 30th ACM international conference on multimedia

Zhou, Y., Yang, X., Zhang, G., Wang, J., Liu, Y., Hou, L., Jiang, X., Liu, X., Yan, J., Lyu, C., et al.: Mmrotate: A rotated object detection benchmark using pytorch. In: Proceedings of the 30th ACM international conference on multimedia. pp. 7331–7334 (2022)

2022

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Cai, X., Lai, Q., Wang, Y., Wang, W., Sun, Z., Yao, Y.: Poly kernel inception network for remote sensing detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 27706–27716 (2024)

2024

[2] [2]

IEEE transactions on circuits and systems for video tech- nology33(5), 2342–2356 (2022)

Dai, L., Liu, H., Tang, H., Wu, Z., Song, P.: Ao2-detr: Arbitrary-oriented object detection transformer. IEEE transactions on circuits and systems for video tech- nology33(5), 2342–2356 (2022)

2022

[3] [3]

In: 2009 IEEE conference on computer vision and pattern recognition

Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248–255. Ieee (2009)

2009

[4] [4]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Ding, J., Xue, N., Long, Y., Xia, G.S., Lu, Q.: Learning roi transformer for oriented object detection in aerial images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2849–2858 (2019)

2019

[5] [5]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Han, J., Ding, J., Xue, N., Xia, G.S.: Redet: A rotation-equivariant detector for aerial object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2786–2795 (2021)

2021

[6] [6]

In: Proceedings of the computer vision and pattern recognition conference

Huang, S., Lu, Z., Cun, X., Yu, Y., Zhou, X., Shen, X.: Deim: Detr with improved matching for fast convergence. In: Proceedings of the computer vision and pattern recognition conference. pp. 15162–15171 (2025)

2025

[7] [7]

In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

Lee, H., Song, M., Koo, J., Seo, J.: Hausdorff distance matching with adaptive query denoising for rotated detection transformer. In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1872–1882. IEEE (2025)

2025

[8] [8]

li et al

Li, Y., Li, X., Dai, Y., Hou, Q., Liu, L., Liu, Y., Cheng, M.M., Yang, J.: Lsknet: A foundation lightweight backbone for remote sensing: Y. li et al. International Journal of Computer Vision133(3), 1410–1431 (2025)

2025