Pith. sign in

REVIEW 4 major objections 8 minor 53 references

A fractional semantic IoU loss joins feature cues with geometry and stabilizes angled box regression.

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 · grok-4.5

2026-07-30 19:55 UTC pith:QWE3JYZG

load-bearing objection Useful empirical loss combo, but the cotθ singularity that justifies the fractional half looks mathematically wrong and the proof is missing. the 4 major comments →

arxiv 2607.23530 v1 pith:QWE3JYZG submitted 2026-07-26 cs.CV

Geometry Meets Semantics: Fractional Gradient Stabilization for Semantic-Driven Bounding Box Optimization in Visual Detection Tasks

Qi Ming , Haitian Yang , Xudong Zhao , Mingjing Zhao , Liuqian Wang , Nanqing Liu This is my paper
classification cs.CV
keywords oriented bounding boxesIoU losssemantic similarityfractional derivativegradient stabilizationobject detectionaerial imageryrotated object detection
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.

Standard IoU losses for oriented boxes ignore the content inside the box and produce wild angular gradients as the angle error shrinks. This paper builds a Semantic IoU that multiplies geometric mismatch by a feature-similarity gate, so reliable appearance cues can steer the box when geometry is weak and vice versa. It then wraps that loss in a fractional-order memory of past similarity values that damps the singular cot-theta spikes near perfect alignment. On aerial, optical-ship, and SAR benchmarks the same drop-in loss raises high-threshold accuracy across one- and two-stage detectors without changing the network. Readers who care about remote sensing, UAV, or 3D yaw boxes get a practical route to tighter orientation without hand-crafted geometric penalties.

Core claim

IoU-driven oriented-box optimization fails on two counts: it supervises only geometry, and its angular gradient blows up like cot(theta) as the angle error approaches zero. Coupling geometric IoU with an adaptive semantic similarity score (SIoU) and modulating the resulting loss by a Grünwald–Letnikov fractional derivative of historical sIoU states (FrSIoU) regularizes those gradients and yields consistent gains in high-precision metrics across bounding-box formulations and detection tasks.

What carries the argument

Fractional Semantic IoU (FrSIoU) loss: L = W_t · (1−s)(1−u), where s is a gated mix of intra-box and context feature similarity, u is geometric IoU, and W_t = exp(−λ D^v_t(sIoU)) is a weight built from the fractional history of sIoU that damps abnormal angular gradients.

Load-bearing premise

The gradient proof treats the semantic score as independent of box angle, so feature gradients never feed back into the angle update.

What would settle it

Replace RoI-warped features by angle-independent constants (or freeze the backbone) while keeping FrSIoU; if mAP75 gains on orientation-sensitive classes vanish or late-training angle-gradient traces still spike, the separation-and-stabilization claim fails.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Stricter metrics (mAP75, full mAP) rise more than loose mAP50 because late-stage angle oscillation is suppressed.
  • The same loss applies once HBB, point-set, Gaussian, polygon, or 3D-yaw boxes are mapped to a unified oriented parametrization.
  • Semantic gating alone supplies non-zero gradients for non-overlapping boxes without extra geometric penalty terms.
  • Gains transfer across optical aerial, elongated-ship, and SAR ship detectors and across one- and two-stage architectures.

Where Pith is reading between the lines

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

  • If differentiable RoI Align makes s depend on theta, a fully joint gradient could shrink residual oscillation or shift the best lambda.
  • Fractional memory of a similarity signal may stabilize other singular or periodic regression targets (periodic angle codes, extreme aspect ratios) beyond IoU.
  • Detectors that already carry strong semantic heads may amplify the gating term more than the paper’s standard backbones show.

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

4 major / 8 minor

Summary. The manuscript proposes FrSIoU, a bounding-box regression loss for oriented object detection with two components: (i) a Semantic IoU (SIoU) term that couples geometric IoU u with a feature-similarity score s via sIoU = s + (1−s)u and L = (1−s)(1−u), with an "adaptive gradient gating" interpretation; and (ii) a fractional-order memory mechanism in which a Grünwald–Letnikov derivative of historical sIoU values produces an exponential modulation weight W_t = exp(−λD^v_t) that rescales the loss. The stated motivation is a claimed singularity in the angular gradient of the skew-IoU loss, ∂L/∂θ = IoU(1+IoU)cotθ (Eq. 2), which the fractional weight is said to regularize. Experiments on DOTA-v1.0, HRSC2016, and SSDD with Rotated RetinaNet and Oriented R-CNN show consistent gains, largest at mAP75 (e.g., +3.32 mAP75 over Oriented R-CNN on DOTA; +10.5 mAP75 for R-RetinaNet on HRSC2016).

Significance. If the empirical results hold, the paper offers a plug-and-play loss with consistent improvements across two detector families, three datasets, and two imaging modalities (optical and SAR), with the largest gains at strict IoU thresholds — the regime that matters for oriented detection. The hyperparameter study (Table 2) suggests low sensitivity to λ and α, and the evaluation protocol (mAP50/75/mAP, test-server numbers for DOTA) is more rigorous than the mAP50-only norm in this subfield. However, the theoretical contribution — the claimed cotθ singularity and its fractional-order regularization — is currently unsupported: the key proof is deferred to an appendix that is not present in the submission, and the derivation in §3.3 contains steps that appear mathematically incorrect as written. Code is promised but not yet released, and all results are single-seed without variance estimates, so the smaller deltas (e.g., +0.45 mAP50 of FrSIoU over SIoU in Table 1) cannot yet be distinguished from noise.

major comments (4)
  1. [§3.1, Eq. (2)] The claimed angular gradient ∂L_IoU/∂θ = IoU(θ)(1+IoU(θ))cotθ appears inconsistent with the actual behavior of skew IoU for center-aligned boxes. For two concentric rectangles with relative angle θ, the intersection/union areas are piecewise polynomial in θ (and in tan(θ/2)-type terms), so IoU(θ) is Lipschitz near θ=0 and 1−IoU(θ) ≈ c·θ + O(θ²); the true derivative tends to a finite constant of order 1, with at worst bounded kinks at vertex-configuration transitions. Eq. (2) instead diverges like 1/θ as θ→0. Both cannot be derivatives of the same loss. This is load-bearing: the design of W_t (Eq. 13), the stabilization story after Eq. (20), and the interpretation of the mAP75 gains as 'angular gradient stabilization' all presuppose this singularity. The proof is deferred to an Appendix that is not included in the submission, so the claim is currently unverifiable. The gradient surges in
  2. [§3.3, Eqs. (17)–(20)] The full-gradient derivation differentiates historical terms sIoU_{t−k} (k ≥ 1) with respect to the current angle θ. Historical loss values from previous optimization steps are constants with respect to the current parameters; ∂u_{t−k}/∂θ_t = 0 for k ≥ 1 unless one backpropagates through the optimization trajectory itself, which is not described and is not standard practice. Consequently, in any actual implementation W_t is a per-step scalar and the 'weighted sum of historical geometric gradients' in Eq. (20) does not exist in the gradient the network receives. Either the implementation does something different from what Eq. (20) states (in which case the equation and the surrounding mechanistic claims must be corrected), or the method backprops through training history (in which case memory/compute costs and the procedure must be specified). As written, the claimed stabilization mechani
  3. [§3.2–3.3, Eq. (9) vs. assumption after Eq. (15)] There is an internal inconsistency about ∂s/∂θ. §3.2 (Eq. 9 and the 'adaptive gradient gating' paragraph) describes the semantic term −(1−u)∂s/∂θ as the mechanism by which confident semantics guide box refinement. §3.3 then assumes ∂s/∂θ = 0 because 'gradient flow is separate from geometric regression.' Both cannot hold. Moreover, the assumption is questionable on its own terms: s is computed from RoI-Align-warped features of the predicted box, and RoI Align is differentiable with respect to box coordinates, so s genuinely depends on θ unless a stop-gradient is applied. If a stop-gradient is applied, that must be stated, and the §3.2 gating story reduces to '(1−s) reweights the geometric loss,' which is a weaker and different claim. Please clarify the exact gradient-flow design (what is detached, what is not) and make §3.2 and §3.3 consistent.
  4. [§4.2 (Table 1) and missing Appendix] Several load-bearing items are deferred to an Appendix that is absent from the submission: the proof of Eq. (2), the proof that θ never reaches 0 under skew-IoU boundary conditions, and the analysis of why s1-only supervision fails to converge. Separately, the empirical support for the fractional component specifically is thin: the SIoU→FrSIoU delta in Table 1 is +0.45 mAP50 from a single seed with no error bars, and Table 2 ablates only λ and α — the fractional order v and history length M, which define the claimed mechanism, are never ablated. The paper's central mechanism claim requires at minimum: (i) the appendix with the proofs, (ii) an ablation over v and M (including v=0 / M=0 degenerate cases that reduce FrSIoU to SIoU), and (iii) multi-seed runs with variance for the key ablation rows.
minor comments (8)
  1. [Table 1 vs. §4.2 text] Table 1 reports GIoU at 66.67 mAP50, but the text in §4.2 states 66.77%. Please reconcile.
  2. [§4.2, 'Effect of Fractional SIoU Loss'] The text says FrSIoU 'adaptively decomposes the angle term into periodic components, ensuring smooth gradient propagation near angle boundaries.' No periodic decomposition or angle-boundary handling appears anywhere in §3.3; the method described there is exponential history modulation. This sentence describes a different mechanism and should be corrected.
  3. [§3.2, naming] The name 'SIoU loss' collides with the existing SCYLLA-IoU (SIoU) loss of Gevorgyan (2022), which is not cited. Please rename or explicitly disambiguate and cite.
  4. [Figure 1] As rendered, Figure 1 appears to contain numerous extraneous panels unrelated to its caption (architecture diagrams with labels such as 'HPFormer', 'RORN', 'HCT Convolution', 'RoI Tokenizer' interleaved with panels (a)–(d)). If this is not a PDF-extraction artifact, the figure needs to be cleaned and any borrowed diagrams attributed; panel (d), which motivates the entire stabilization story, should be clearly legible with axes labeled.
  5. [Table 2] The hyperparameter table's layout is garbled (the λ/α column pairing is ambiguous). Please reformat so that each (λ, α) pair and its mAP50 is unambiguous, and state the fixed values of v, M, β, N used.
  6. [§4.1] Please specify how the N and M sampled feature locations in Eqs. (3)–(5) are chosen, the expansion factor for the context boxes b̂, and the values of v, M, β used in all experiments. These are needed for reproduction; code release upon publication would also address this.
  7. [§3.1, Eq. (2) text] The claim that the factor IoU(1+IoU) 'can increase with IoU and therefore amplify the singular behaviour' is confusingly worded: near convergence IoU→1, so the factor tends to 2 while cotθ diverges; the amplification statement adds nothing and depends on the contested Eq. (2).
  8. [General] Typos and style: 'hindering high-precision' (§1, missing object), 'Semantics IoU loss' (§5), inconsistent boldface spacing in the abstract ('Fr actional S emantic...'), 'UA V' spacing, and 'Franctional IoU' in the Figure 1 graphics. A careful proofread is recommended.

Circularity Check

1 steps flagged

Empirical loss-design paper on external benchmarks; no derivation-by-construction. Only mild self-citation for the cotθ motivation.

specific steps
  1. self citation load bearing [§3.1 Anomalous Gradient Analysis, Eq. (2); cites [4]]
    "Previous studies [4] have shown that center point convergence in OBB regression is fast and stable... we obtain the loss gradient with respect to θ as follows: ∂L_IoU/∂θ = IoU(θ)/(1+IoU(θ)) cot θ. The detailed proof can be found in the Appendix."

    The load-bearing motivation for fractional stabilization—the claimed cotθ singularity and late-stage gradient surges—is justified primarily by overlapping-author prior work [4] plus an appendix proof not present in the manuscript. This is mild: it frames interpretation of FrSIoU, but does not make the loss definition or the external mAP numbers true by construction.

full rationale

FrSIoU is a designed training objective (semantic gating plus Grünwald–Letnikov history weights), not a first-principles prediction forced by its inputs. SIoU and FrSIoU are defined by construction as loss formulae (Eqs. 7–8, 13–14); the paper then measures mAP on held-out external benchmarks (DOTA-v1.0, HRSC2016, SSDD) under standard detectors. Those numbers are independent empirical outcomes, not renormalizations of fitted parameters. Hyperparameter choices (λ, α) are ordinary validation tuning, not “predictions” of closely related quantities. The only mild circularity-adjacent element is motivational: the claimed angular singularity ∂L_IoU/∂θ ∝ cot θ (Eq. 2) and the “abnormal gradient” framing lean on prior work [4] by overlapping authors (Qi Ming et al., Gradient Calibration Loss), with the detailed proof deferred to a missing appendix. That self-citation shapes the story of why fractional weights should help, but it does not force the reported mAP gains or make FrSIoU equivalent to its inputs. Correctness doubts about Eq. (2) are separate from circularity. Score 1.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The central empirical claim rests on standard detection tooling plus several paper-specific constructions (semantic score, sIoU coupling, GL memory weight) and free knobs (λ, α, β, v, M, sample counts). Gradient-stability theory further assumes simplified center-aligned skew IoU behavior and that semantic scores do not backprop into angle—both are modeling choices, not theorems proved in the main text.

free parameters (6)
  • λ (modulation intensity) = 0.3 (selected)
    Scales the exponential weight Wt=exp(-λ Dv(sIoU)); chosen by grid search on DOTA val (Table 2).
  • α (semantic distance sensitivity) = 0.3 (selected)
    Controls s1=exp(-α dintra); grid-searched with λ in Table 2.
  • β (intra vs context semantic mix) = 0.5 (default)
    s=β s1+(1-β)s2; stated as 'generally 0.5' without full ablation.
  • v (fractional derivative order) = unspecified in main experiments
    Order of Grünwald–Letnikov operator in Dv_t(sIoU); appears in equations but is not fixed in the hyperparameter table.
  • M (historical steps) = example M=2
    Length of sIoU memory in the GL sum; illustrated with M=2, not systematically reported for final models.
  • N, M feature sample counts for s1/s2
    Number of paired RoI/context feature samples entering dintra and cosine similarity; not numerically specified.
axioms (6)
  • ad hoc to paper For center-aligned OBBs, ∂L_IoU/∂θ = IoU(θ)(1+IoU(θ)) cot θ, with singularity as θ→0 but θ never exactly 0 under skew-IoU boundaries.
    Sec. 3.1 Eq. (2); proof deferred to appendix not in the manuscript. Load-bearing for the 'unstable angular gradient' motivation.
  • ad hoc to paper Semantic score s can be treated as constant w.r.t. geometric parameters during FrSIoU gradient analysis (∂s/∂θ=0).
    Explicit 'fundamental assumption' in Sec. 3.3 used to reach Eq. (20).
  • domain assumption Grünwald–Letnikov finite-history sum is a valid practical fractional derivative of training-time sIoU for loss weighting.
    Sec. 3.3 Eqs. (10)–(12), citing fractional-calculus references [40,41]; applied to epoch/step losses without a learning-theory guarantee.
  • ad hoc to paper sIoU = s+(1-s)u and L = (1-s)(1-u) is a valid semantic-geometric coupling that preserves u=1 ⇒ sIoU=1 and supplies gradients when u=0.
    Sec. 3.2 Eqs. (7)–(8); design choice defining SIoU.
  • domain assumption Standard detection training assumptions: MMRotate pipelines, SGD, DOTA single-scale 1024 protocol, IoU-based mAP as localization quality.
    Sec. 4.1 implementation and metrics; shared with the subfield.
  • domain assumption Most box representations (HBB, points, Gaussian, polygon, 3D yaw boxes) map to a unified OBB parametrization for optimization analysis.
    Sec. 3.1 and Fig. 2; used to claim cross-formulation generality.
invented entities (3)
  • Semantic score s (intra-region MSE similarity s1 + context cosine s2) no independent evidence
    purpose: Inject appearance/context cues into box regression and gate geometric IoU.
    Defined in Sec. 3.2 Eqs. (3)–(6); no external standard definition of this exact pair.
  • SIoU / LsIoU with adaptive gradient gating no independent evidence
    purpose: Unified semantic-geometric loss claimed to fix non-overlap vanishing gradients without hand-crafted geometric penalties.
    Eqs. (7)–(9); entity is the loss itself, validated only via the paper's ablations.
  • FrSIoU weight Wt from fractional derivative of historical sIoU no independent evidence
    purpose: Modulate loss magnitude to regularize abnormal angular gradients near convergence.
    Eqs. (10)–(14); new training mechanism without independent theoretical certificate outside reported mAP gains.

pith-pipeline@v1.2.0-grok45-kimik3 · 22126 in / 4505 out tokens · 99481 ms · 2026-07-30T19:55:46.456971+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Geometry Meets Semantics: Fractional Gradient Stabilization for Semantic-Driven Bounding Box Optimization in Visual Detection Tasks." pith.science (2026). https://pith.science/paper/QWE3JYZG

@misc{pith2026260723530,
  author       = {Pith},
  title        = {Pith review of: Geometry Meets Semantics: Fractional Gradient Stabilization for Semantic-Driven Bounding Box Optimization in Visual Detection Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QWE3JYZG}},
  note         = {Machine review of arXiv:2607.23530}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Bounding boxes are fundamental for object localization in visual detection tasks. Among them, oriented bounding boxes are widely used in visual detection tasks, which provide a more precise directional representation. Generally, IoU-based losses are widely adopted to optimize box regression. However, we observed that IoU-driven box optimization suffers from two key issues: (1) it relies solely on geometric properties while ignoring semantic cues; (2) orientation optimization suffers from unstable gradients, causing oscillations in orientation convergence. In this paper, we propose a Fractional Semantic IoU loss to achieve unified semantic-geometric learning with gradient stabilization. First, we design a semantic similarity metric to guide IoU optimization, building a Semantic IoU loss (SIoU loss) with an adaptive gradient gating mechanism. Then, we revisit the gradient instability issue in oriented box optimization and extend the SIoU loss to a fractional-order formulation to build the \textbf{Fr}actional \textbf{S}emantic \textbf{IoU} \textbf{loss} (FrSIoU loss). The FrSIoU loss accumulates historical IoU states to regularize abnormal gradients during bounding box optimization process. Extensive experiments demonstrate that our approach achieves stable performance gains across different bounding box formulations and diverse visual detection tasks. The code will be available on GitHub.

Figures

Figures reproduced from arXiv: 2607.23530 by Haitian Yang, Liuqian Wang, Mingjing Zhao, Nanqing Liu, Qi Ming, Xudong Zhao.

Figure 2
Figure 2. Figure 2: Illustration of different types of bounding box representations. OBB shows strong generality and can be [PITH_FULL_IMAGE:figures/full_fig_p003_2.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

53 extracted references · 2 linked inside Pith

  1. [1]

    Rotation-robust intersection over union for 3D object detection

    Yu Zheng, Danyang Zhang, Sinan Xie, Jiwen Lu, and Jie Zhou. Rotation-robust intersection over union for 3D object detection. InEur . Conf. Comput. Vis., 2020

    2020

  2. [2]

    Deep dive into gradients: Better optimization for 3d object detection with gradient-corrected IoU supervision

    Qi Ming, Lingjuan Miao, Zhe Ma, Lin Zhao, Zhiqiang Zhou, Xuhui Huang, Yuanpei Chen, and Yufei Guo. Deep dive into gradients: Better optimization for 3d object detection with gradient-corrected IoU supervision. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2023

    2023

  3. [3]

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

    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.Proc. Adv. Neural Inf. Process. Syst., 2021

    2021

  4. [4]

    Gradient calibration loss for fast and accurate oriented bounding box regression.IEEE Trans

    Qi Ming, Lingjuan Miao, Zhiqiang Zhou, Junjie Song, and Aleksandra Pizurica. Gradient calibration loss for fast and accurate oriented bounding box regression.IEEE Trans. Geosci. Remote Sens., 62, 2024

    2024

  5. [5]

    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. InProc. Int. Conf. Mach. Learn.PMLR, 2021

    2021

  6. [6]

    Grasp pose detection from a single rgb image

    Hu Cheng, Yingying Wang, and Max Q-H Meng. Grasp pose detection from a single rgb image. InProc. IEEE/RSJ Int. Conf. Intell. Robots Syst.IEEE, 2021

    2021

  7. [7]

    Point2rbox-v2: Rethinking point-supervised oriented object detection with spatial layout among instances

    Yi Yu, Botao Ren, Peiyuan Zhang, Mingxin Liu, Junwei Luo, Shaofeng Zhang, Feipeng Da, Junchi Yan, and Xue Yang. Point2rbox-v2: Rethinking point-supervised oriented object detection with spatial layout among instances. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2025

    2025

  8. [8]

    Towards accurate medical image segmentation with gradient-optimized dice loss.IEEE Signal Process

    Qi Ming and Xiaowu Xiao. Towards accurate medical image segmentation with gradient-optimized dice loss.IEEE Signal Process. Lett., 31, 2024

    2024

  9. [9]

    Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding.Int

    Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding.Int. J. Comput. Vis., 133(2), 2025

    2025

  10. [10]

    Faster R-CNN: Towards real-time object detection with region proposal networks.IEEE Trans

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks.IEEE Trans. Pattern Anal. Mach. Intell., 39(6), 2016

    2016

  11. [11]

    SSD: Single shot multibox detector

    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. SSD: Single shot multibox detector. InProc. Eur . Conf. Comput. Vis.Springer, 2016

    2016

  12. [12]

    J. Han, J. Ding, J. Li, and G. S. Xia. Align deep features for oriented object detection.IEEE Trans. Geosci. Remote Sens., 2021

    2021

  13. [13]

    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. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2019

    2019

  14. [14]

    Polygon-to-polygon distance loss for rotated object detection

    Yang Yang, Jifeng Chen, Xiaopin Zhong, and Yuanlong Deng. Polygon-to-polygon distance loss for rotated object detection. Proc. AAAI Conf. Artif. Intell., 36(3), 2022

    2022

  15. [15]

    Iou loss for 2d/3d object detection

    Dingfu Zhou, Jin Fang, Xibin Song, Chenye Guan, Junbo Yin, Yuchao Dai, and Ruigang Yang. Iou loss for 2d/3d object detection. InProc. Int. Conf. 3D Vis.IEEE, 2019

    2019

  16. [16]

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

    Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2019

    2019

  17. [17]

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

    Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-IoU loss: Faster and better learning for bounding box regression. InProc. AAAI Conf. Artif. Intell., 2020

    2020

  18. [18]

    A systematic iou-related method: Beyond simplified regression for better localization.IEEE Trans

    Hanyang Peng and Shiqi Yu. A systematic iou-related method: Beyond simplified regression for better localization.IEEE Trans. Image Process., 30, 2021

    2021

  19. [19]

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

    Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. DOTA: A large-scale dataset for object detection in aerial images. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2018

    2018

  20. [20]

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

    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 J. Photogramm. Remote Sens., 159, 2020

    2020

  21. [21]

    Developing a more reliable framework for extracting traffic data from a uav video.IEEE Trans

    Xiaohe Li and Jianping Wu. Developing a more reliable framework for extracting traffic data from a uav video.IEEE Trans. Intell. Transp. Syst., 24(11), 2023

    2023

  22. [22]

    Weakly misalignment-free adaptive feature alignment for uavs-based multimodal object detection

    Chen Chen, Jiahao Qi, Xingyue Liu, Kangcheng Bin, Ruigang Fu, Xikun Hu, and Ping Zhong. Weakly misalignment-free adaptive feature alignment for uavs-based multimodal object detection. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2024

    2024

  23. [23]

    Rethinking iou-based optimization for single-stage 3d object detection

    Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, Min-Jian Zhao, and Gim Hee Lee. Rethinking iou-based optimization for single-stage 3d object detection. InProc. Eur . Conf. Comput. Vis.Springer, 2022

    2022

  24. [24]

    Dynamic anchor learning for arbitrary-oriented object detection

    Qi Ming, Zhiqiang Zhou, Lingjuan Miao, Hongwei Zhang, and Linhao Li. Dynamic anchor learning for arbitrary-oriented object detection. InProc. AAAI Conf. Artif. Intell., 2021

    2021

  25. [25]

    PIoU Loss: Towards accurate oriented object detection in complex environments

    Zhiming Chen, Kean Chen, Weiyao Lin, John See, Hui Yu, Yan Ke, and Cong Yang. PIoU Loss: Towards accurate oriented object detection in complex environments. InProc. Eur . Conf. Comput. Vis., 2020. 10 APREPRINT- JULY28, 2026

    2020

  26. [26]

    Polygon pixel iou: Similarity metric between polygons with different number of vertices for arbitrary-shaped text spotting

    Rei Endo, Taro Miyazaki, Takahiro Mochizuki, and Yoshihiko Kawai. Polygon pixel iou: Similarity metric between polygons with different number of vertices for arbitrary-shaped text spotting. InProc. IEEE Int. Conf. Acoust. Speech Signal Process. IEEE, 2025

    2025

  27. [27]

    Wise-iou: bounding box regression loss with dynamic focusing mechanism.arXiv preprint arXiv:2301.10051, 2023

    Zanjia Tong, Yuhang Chen, Zewei Xu, and Rong Yu. Wise-iou: bounding box regression loss with dynamic focusing mechanism.arXiv preprint arXiv:2301.10051, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    Real-time rotation-invariant face detection with progressive calibration networks

    Xuepeng Shi, Shiguang Shan, Meina Kan, Shuzhe Wu, and Xilin Chen. Real-time rotation-invariant face detection with progressive calibration networks. InIEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2018

    2018

  29. [29]

    Five-degree-of-freedom strawberry pose estimation using key points and oriented bounding box detection.Computers and Electronics in Agriculture, 239, 2025

    Tianxiao Zhu, Zhonghua Miao, Chunjiang Zhao, Liping Chen, and Ya Xiong. Five-degree-of-freedom strawberry pose estimation using key points and oriented bounding box detection.Computers and Electronics in Agriculture, 239, 2025

    2025

  30. [30]

    Rotation-sensitive regression for oriented scene text detection

    Minghui Liao, Zhen Zhu, Baoguang Shi, Gui-song Xia, and Xiang Bai. Rotation-sensitive regression for oriented scene text detection. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2018

    2018

  31. [31]

    Oriented reppoints for aerial object detection

    Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reppoints for aerial object detection. InIEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022

    2022

  32. [32]

    Optimized point set representation for oriented object detection in remote-sensing images.IEEE Trans

    Junjie Song, Lingjuan Miao, Zhiqiang Zhou, Qi Ming, and Yunpeng Dong. Optimized point set representation for oriented object detection in remote-sensing images.IEEE Trans. Geosci. Remote Sens., 20, 2023

    2023

  33. [33]

    Oriented object detection in aerial images with box boundary-aware vectors

    Jingru Yi, Pengxiang Wu, Bo Liu, Qiaoying Huang, Hui Qu, and Dimitris Metaxas. Oriented object detection in aerial images with box boundary-aware vectors. InIEEE/CVF Winter Conf. Appl. Comput. Vis., 2021

    2021

  34. [34]

    The kfiou loss for rotated object detection.arXiv preprint arXiv:2201.12558, 2022

    Xue Yang, Yue Zhou, Gefan Zhang, Jirui Yang, Wentao Wang, Junchi Yan, Xiaopeng Zhang, and Qi Tian. The kfiou loss for rotated object detection.arXiv preprint arXiv:2201.12558, 2022

    Pith/arXiv arXiv 2022

  35. [35]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEur . Conf. Comput. Vis., 2014

    2014

  36. [36]

    The pascal visual object classes (voc) challenge.Int

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge.Int. J. Comput. Vis., 88(2), 2010

    2010

  37. [37]

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

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2016

    2016

  38. [38]

    Yale University, 1978

    Michael Ian Shamos.Computational geometry. Yale University, 1978

    1978

  39. [39]

    Mask R-CNN

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. InProc. IEEE/CVF Int. Conf. Comput. Vis., 2017

    2017

  40. [40]

    A new definition of fractional derivative.Journal of computational and applied mathematics, 264, 2014

    Roshdi Khalil, Mohammed Al Horani, Abdelrahman Yousef, and Mohammad Sababheh. A new definition of fractional derivative.Journal of computational and applied mathematics, 264, 2014

    2014

  41. [41]

    The grünwald–letnikov method for fractional differential equations.Computers & Mathematics with Applications, 62(3), 2011

    Rudolf Scherer, Shyam L Kalla, Yifa Tang, and Jianfei Huang. The grünwald–letnikov method for fractional differential equations.Computers & Mathematics with Applications, 62(3), 2011

    2011

  42. [42]

    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. InProc. Int. Conf. Pattern Recognit. Appl. Methods., 2017

    2017

  43. [43]

    Sar ship detection dataset (ssdd): Official release and comprehensive data analysis.Remote Sensing, 13(18), 2021

    Tianwen Zhang, Xiaoling Zhang, Jianwei Li, Xiaowo Xu, Baoyou Wang, Xu Zhan, Yanqin Xu, Xiao Ke, Tianjiao Zeng, Hao Su, et al. Sar ship detection dataset (ssdd): Official release and comprehensive data analysis.Remote Sensing, 13(18), 2021

    2021

  44. [44]

    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, Wenwei Zhang, and Kai Chen. Mmrotate: A rotated object detection benchmark using pytorch. InProc. ACM Int. Conf. Multimedia, 2022

    2022

  45. [45]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. InIEEE/CVF Int. Conf. Comput. Vis., 2017

    2017

  46. [46]

    Arbitrary-oriented object detection with circular smooth label

    Xue Yang and Junchi Yan. Arbitrary-oriented object detection with circular smooth label. InProc. Eur . Conf. Comput. Vis. Springer, 2020

    2020

  47. [47]

    R3Det: Refined single-stage detector with feature refinement for rotating object.Proc

    Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3Det: Refined single-stage detector with feature refinement for rotating object.Proc. AAAI Conf. Artif. Intell., 35(4), May 2021

    2021

  48. [48]

    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 Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., June 2021

    2021

  49. [49]

    Orientedformer: An end-to-end transformer-based oriented object detector in remote sensing images.IEEE Trans

    Jiaqi Zhao, Zeyu Ding, Yong Zhou, Hancheng Zhu, Wen-Liang Du, Rui Yao, and Abdulmotaleb El Saddik. Orientedformer: An end-to-end transformer-based oriented object detector in remote sensing images.IEEE Trans. Geosci. Remote Sens., 2024

    2024

  50. [50]

    Rethinking boundary discontinuity problem for oriented object detection

    Hang Xu, Xinyuan Liu, Haonan Xu, Yike Ma, Zunjie Zhu, Chenggang Yan, and Feng Dai. Rethinking boundary discontinuity problem for oriented object detection. InConf. Comput. Vis. Pattern Recognit., 2024

    2024

  51. [51]

    Gsdet: Gaussian splatting for oriented object detection

    Zeyu Ding, Jiaqi Zhao, Yong Zhou, Wen-liang Du, Hancheng Zhu, and Rui Yao. Gsdet: Gaussian splatting for oriented object detection. InProc. Int. Jt. Conf. Artif. Intell., 2025

    2025

  52. [52]

    Gaucho: Gaussian distributions with cholesky decomposition for oriented object detection

    José Henrique Lima Marques, Jeffri Murrugarra-Llerena, and Claudio R Jung. Gaucho: Gaussian distributions with cholesky decomposition for oriented object detection. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2025. 11 APREPRINT- JULY28, 2026

    2025

  53. [53]

    Oriented R-CNN for object detection

    Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented R-CNN for object detection. InIEEE/CVF Int. Conf. Comput. Vis., 2021. 12

    2021