Pith. sign in

REVIEW 4 major objections 6 minor 39 references

InterpIoU: Rethinking Bounding Box Regression with Interpolation-Based IoU Optimization

T0 review · 4 major / 6 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 Adding an interpolated-box IoU term to regression loss fixes vanishing gradients and lifts small-object detection.

desk verdict A genuinely new interpolated-box IoU loss with small but consistent gains; the simulation evidence for the strong 'penalties are unnecessary' claim does not survive contact with the actual optimizer in the main experiments. read the letter →

arxiv 2507.12420 v2 pith:B2VH7VSX submitted 2025-07-16 cs.CV

Haoyuan Liu , Hiroshi Watanabe This is my paper · ORCID
classification cs.CV
keywords boundingboxregressionIoUlossinterpolatedboxesgradientvanishingsmallobjectdetectiondynamicinterpolationenlargement
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 claims that the vanishing-gradient problem in bounding box regression--the loss stops teaching the model when the predicted box and the ground truth do not overlap--can be solved without any handcrafted geometric penalty terms. Its proposed loss, InterpIoU, adds to the ordinary IoU loss a second IoU term computed between the ground truth and a box that is a convex interpolation of the prediction and the ground truth. Because the interpolated box overlaps the ground truth even when the prediction does not, gradients keep flowing; because both terms are pure IoU, the loss stays aligned with the evaluation metric and cannot reward box enlargement. Simulations in the paper lead the authors to conclude that IoU alone is an ideal regression target, and they report consistent gains over previous IoU-based losses on COCO, VisDrone, and PASCAL VOC, most notably for small objects.

What carries the argument

The interpolated box $B_{\mathrm{int}}=(1-\alpha)B_{\mathrm{pred}}+\alpha B_{\mathrm{gt}}$ is the load-bearing object: a convex combination of prediction and ground truth that is guaranteed, for $\alpha$ above a derived lower bound, to intersect the ground truth along both axes even when the prediction is far away. The paper derives the overlap guarantee $\alpha > \max_i d_i/(d_i + l_i^g)$ along each axis, uses the empirical observation that assigned boxes satisfy $d_i < l_i^g$ to justify $\alpha > 0.5$, and sets $\alpha=0.98$. The loss is the sum of the ordinary IoU loss and the IoU loss of the interpolated box; backpropagation through the second term gives a gradient scaled by $(1-\alpha)$. D-InterpIoU replaces fixed $\alpha$ with a clamped function of $1-\mathrm{IoU}$, boosting gradients when IoU is low.

What would settle it

A controlled experiment that keeps the loss value identical but multiplies only the interpolation term's gradient by zero would settle whether the non-overlap signal is the mechanism: if AP on COCO and VisDrone is essentially unchanged, the second term is not doing the claimed work. A complementary check is to train with a small fixed $\alpha$ (e.g., 0.5), which gives a much larger interpolation gradient but less overlap margin; if regression from non-overlapping boxes still stalls, the rescue of vanishing gradients is not happening in practice.

Watch

Extended reading notes

Core claim

The central discovery is that the loss $L_{\mathrm{InterpIoU}}(B_{\mathrm{pred}},B_{\mathrm{gt}})=L_{\mathrm{IoU}}(B_{\mathrm{pred}},B_{\mathrm{gt}})+L_{\mathrm{IoU}}(B_{\mathrm{int}},B_{\mathrm{gt}})$ with $B_{\mathrm{int}}=(1-\alpha)B_{\mathrm{pred}}+\alpha B_{\mathrm{gt}}$, $\alpha=0.98$, supplies informative gradients for non-overlapping boxes while remaining a pure function of IoU. The second term is non-zero when the first term is flat, so it rescues gradient flow; its gradient is scaled by $(1-\alpha)$, which the paper argues is sufficient. Since no geometric penalty term such as center distance or aspect ratio enters the loss, the optimizer's objective never conflicts with maximizing IoU, which the paper identifies as the root cause of the bounding box enlargement phenomenon. The dynamic variant D-InterpIoU sets $\alpha_{\mathrm{dyn}}=\mathrm{clamp}(1-\mathrm{IoU}(B_{\mathrm{pred}},B_{\mathrm{gt}}),\alpha_{\mathrm{low}},\alpha_{\mathrm{high}})$, creating a stronger gradient in low-IoU regimes. The paper's empirical claim is that this loss family outperforms prior IoU-based losses across three benchmarks and multiple detectors, with the largest margins on small and crowded objects.

Load-bearing premise

The load-bearing premise is that the interpolation term's gradient, scaled by $1-\alpha=0.02$, is large enough to actually drive training for non-overlapping boxes; the paper supports this with a toy simulation (Adam, learning rate 0.01) rather than a convergence proof or an isolated full-training experiment.

Editorial extensions

If this is right

  • For a prediction with zero overlap, the second term yields a nonzero gradient toward the ground truth, so non-overlap no longer means no learning signal.
  • Because both terms of InterpIoU are IoU-based, the optimizer has no incentive to inflate the predicted box to satisfy a geometric penalty, which the paper argues removes the bounding box enlargement problem.
  • The loss depends on no shape-, size-, or distribution-specific handcrafted parameters, which the paper identifies as the reason it stays stable for small and densely packed objects.
  • D-InterpIoU strengthens gradients in the low-IoU regime, acting as an implicit curriculum that helps the model escape flat loss regions.
  • On COCO, VisDrone, and PASCAL VOC, the paper reports that InterpIoU and D-InterpIoU improve localization accuracy over previous IoU-based losses across several detectors, with the largest gains on small objects.

Reading between the lines

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

  • Editorial inference: the same interpolation construction could transfer to other overlap metrics with flat regions, such as rotated-box IoU or segmentation IoU, where vanishing gradients are also a known obstacle.
  • Editorial inference: because the interpolation gradient is scaled by $(1-\alpha)=0.02$ at the chosen $\alpha=0.98$, an untested extension is to anneal $\alpha$ during training to expose a stronger non-overlap signal early, which would isolate whether the magnitude or the mere presence of that signal drives the gains.
  • Editorial inference: if pure-IoU interpolation is as sufficient as the paper claims, the loss family should show its largest relative gains on datasets with extreme aspect-ratio or density distributions, where geometric penalties are most misaligned; this prediction is not tested in the paper.
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 / 6 minor

Summary. The paper proposes InterpIoU, a bounding-box regression loss that adds to the standard IoU loss a second IoU term computed between the ground-truth box and an interpolated box B_int = (1-alpha)B_pred + alpha B_gt. With alpha close to 1, the interpolated box overlaps the target even when the prediction is disjoint, providing a nonzero gradient signal without explicit geometric penalty terms. The paper also proposes Dynamic InterpIoU (D-InterpIoU), which sets alpha = clamp(1 - IoU(B_pred,B_gt), alpha_low, alpha_high) to strengthen gradients in low-IoU regimes. The central claims are that the interpolation term alone supplies meaningful non-overlap gradients, that handcrafted geometric penalties are unnecessary and even harmful, and that the proposed losses consistently improve small-object detection on VOC, VisDrone, and COCO with SSD, YOLOv8, and DINO. The paper supports these claims with a gradient derivation, toy regression simulations, an analysis of the bounding-box enlargement problem, and benchmark experiments.

Significance. If the central claims hold, the paper offers a conceptually simple alternative to the increasingly complex geometric penalties in IoU-based losses, and the proposed loss would be attractive for small-object and dense-scene detection where penalty terms are known to be unstable. The core construction is easy to implement and the gradient analysis in Eqs. (23)-(26) is correct and clearly presented. The paper also makes a falsifiable claim: interpolation-based IoU supervision should suffice without center-distance, aspect-ratio, or angle terms. However, the current evidence for that claim is not fully load-bearing: the supporting simulation uses Adam, which masks the small (1-alpha) gradient scale that matters for the SGD-based main experiments, and the D-InterpIoU gradient is not derived for the case where alpha depends on the prediction. The benchmark gains are small and are reported without variance, and the dynamic variant has dataset-specific clamp ranges. With additional experiments and clarifications the contribution could be solid, but in its present form the strong conclusions outrun the evidence.

major comments (4)
  1. [Sec. 3.3, Eq. (19), Sec. 4.1] The simulation that supports the claim that the interpolation term provides sufficient non-overlap supervision and that geometric penalties are unnecessary is run with Adam at learning rate 0.01 (Sec. 3.3). Because Adam normalizes each coordinate update by its running gradient magnitude, the reported curves are nearly invariant to the (1-alpha)=0.02 factor in Eq. (19), whereas the YOLOv8 and SSD experiments in Sec. 4.1 use SGD-style training under the Ultralytics framework with no reported optimizer override. The simulation also reports average loss/penalty values rather than final localization error. Please add SGD versions of the simulation, report final center and size errors, and include a real-training ablation that removes or rescales the interpolation term to isolate its effective contribution; otherwise the claimed advantage over O(1) geometric penalties in the benchmark settings is not established.
  2. [Sec. 3.2, Eqs. (18)-(20)] For D-InterpIoU, alpha_dyn in Eq. (20) is a function of IoU(B_pred, B_gt), so B_int depends on B_pred through both (1-alpha) and alpha itself. The gradient decomposition in Eqs. (18)-(19), which treats alpha as constant, is therefore incomplete if alpha is not detached during backpropagation; the missing term is (B_gt - B_pred) * d(alpha)/d(B_pred). The paper should state explicitly whether alpha is detached and, if not, give the D-InterpIoU gradient. This matters because the 'gradient boost zone' discussion and Fig. 4 treat alpha as an externally varied parameter, which does not reveal the behavior of the differentiable loss; if alpha is detached, the dynamic mechanism is purely a loss-landscape change and the claim of an adaptive gradient boost should be qualified.
  3. [Sec. 4.2, Table 4] D-InterpIoU introduces two free parameters per dataset: the clamps in Sec. 4.2 are [0.95,0.99] on VOC, [0.60,0.99] on VisDrone, and [0.90,0.99] on COCO, while the ablation in Table 4 uses [0.00,0.99]. These ranges are not justified by a principled selection rule, and the statement in Sec. 1 that D-InterpIoU 'does not introduce complexity or handcrafted components' is therefore too strong. Please provide a validation-based selection rule for the clamps, or report sensitivity of the final metrics to the clamp values, and temper the handcrafted-component claim accordingly.
  4. [Tables 1-3] The reported gains over the best baseline are small (e.g., Table 3: YOLOv8 mAP 49.4 vs 49.3 for SIoU; DINO 49.7 vs 49.6 for PIoU), and in the DINO/COCO row D-InterpIoU is below PIoU (49.4 vs 49.6). No standard deviations, multiple seeds, or significance tests are provided. Since the paper claims consistent superiority over state-of-the-art losses, please add repeated-run statistics or explicitly restrict the claim to the settings where the difference is above noise.
minor comments (6)
  1. [Sec. 3.4.2, Eqs. (25)-(26)] The sentence 'this term pushes w and h to decrease' is imprecise: the second term in Eqs. (25)-(26) is positive when w or h increases, so it counteracts the first term but does not by itself produce a decrease unless the intersection term is appropriately signed. Please rephrase.
  2. [Fig. 5] The axes and curves in Fig. 5 are not fully labeled; please specify which subplot shows which metric and which curve corresponds to which loss, and state whether the plotted values are loss values, penalty values, or final errors.
  3. [Sec. 3.3] The simulation setup does not state the coordinate normalization or the box parameterization used in Eq. (12); please clarify whether the interpolation is applied to absolute [x,y,w,h] coordinates or to network-encoded offsets, since the equivalence matters for the real detectors.
  4. [Sec. 4.1] For SSD, YOLOv8, and DINO, please specify whether the loss is computed on decoded bounding boxes or on the network's raw regression targets; linear interpolation in Eq. (12) may not commute with the anchor or query encoding used by these detectors.
  5. [Sec. 3.1, Eq. (17)] The lower bound in Eq. (17) should explicitly note that the max is taken only over axes with d_i > 0; as written, the role of the max is not fully explained for overlapping axes.
  6. [Sec. 4.2] The paper states that alpha = 0.98 is used unless otherwise specified, but D-InterpIoU uses different alpha ranges on every dataset; please make the relationship between the static alpha and the dynamic clamp ranges explicit in the experimental setup.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the loss definition and overlap guarantee are derived self-containedly, and benchmark-driven coefficient choices are hyperparameter tuning rather than a circular reduction.

full rationale

The paper's derivation chain is self-contained. L_InterpIoU is explicitly defined as LIoU(Bpred,Bgt) + LIoU(Bint,Bgt) with Bint=(1-α)Bpred+αBgt, and the claim that Bint overlaps Bgt for sufficiently large α follows from elementary interval geometry (Eqs. 14-17), independent of the loss itself. The gradient expression in Eq. 19 openly includes the (1-α) factor, so the scale of the interpolation term is disclosed rather than hidden. The simulation experiments compare the proposed loss against baseline losses under Adam; using the proposed method as one of the compared losses is the standard way to test a regression loss and does not make the comparison circular. The choice α=0.98 and the per-dataset clamp ranges for D-InterpIoU are selected through ablations and experimental tuning, which is a benchmark-driven hyperparameter concern and a potential generalization risk, but no reported result is forced by construction from those fitted values. There are also no load-bearing self-citations: the reference list contains no prior work by the present authors, and no uniqueness theorem or prior result by the authors is invoked to exclude alternative losses. Concerns about the small (1-α)=0.02 gradient scale under SGD, and about the toy simulation using Adam whose normalization can mask constant-factor scaling, are empirical and correctness-oriented rather than circularity in the derivation. Overall, the central formulation and its supporting analysis do not reduce to their own inputs.

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

Most of the paper's machinery is standard loss engineering. The main free parameters are the interpolation coefficient α and the dynamic clamp bounds, which are tuned per dataset. The interpolation strategy relies on the empirical claim that assigned boxes are near the ground truth (ATSS/TAL), and on the unproven assumption that the (1-α)-scaled gradient is strong enough to guide non-overlapping boxes.

free parameters (3)
  • static interpolation coefficient α = 0.98
    Set to 0.98 for all main experiments based on ablation on VisDrone and VOC (Table 4). The loss and its behavior depend strongly on this value.
  • D-InterpIoU lower clamp α_low = 0.95 (VOC), 0.60 (VisDrone), 0.90 (COCO)
    Per-dataset tuned lower bound for the dynamic interpolation coefficient. Different values are used on each benchmark, indicating sensitivity to data distribution.
  • D-InterpIoU upper clamp α_high = 0.99 (all datasets)
    Upper bound for the dynamic coefficient, chosen to maintain minimum overlap. Also a tuned constant.
assumptions (3)
  • domain assumption Assigned positive boxes under ATSS/TAL satisfy the empirical condition d_i < l^g_i along each axis.
    Used in Sec 3.1 to justify that α > 0.5 guarantees non-zero overlap between the interpolated and ground-truth boxes. Not proven for all assignments.
  • standard math Linearly interpolating box coordinates in [x,y,w,h] space yields a valid bounding box for any α in [0,1].
    Convex combination preserves the box parameterization, used in Eq. 12.
  • domain assumption The gradient of the IoU loss with respect to the interpolated box is well-defined and informative even when the interpolated box is very close to the ground truth.
    The paper relies on this to claim that the (1-α)-scaled gradient moves non-overlapping predictions. Only demonstrated in simulation, not proved.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InterpIoU: Rethinking Bounding Box Regression with Interpolation-Based IoU Optimization." pith.science (2026). https://pith.science/paper/B2VH7VSX

@misc{pith2026250712420,
  author       = {Pith},
  title        = {Pith review of: InterpIoU: Rethinking Bounding Box Regression with Interpolation-Based IoU Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2VH7VSX}},
  note         = {Machine review of arXiv:2507.12420}
}
read the original abstract

Bounding box regression (BBR) is fundamental to object detection, where the regression loss is crucial for accurate localization. Existing IoU-based losses often incorporate handcrafted geometric penalties to address IoU's non-differentiability in non-overlapping cases and enhance BBR performance. However, these penalties are sensitive to box shape, size, and distribution, often leading to suboptimal optimization for small objects and undesired behaviors such as bounding box enlargement due to misalignment with the IoU objective. To address these limitations, we propose InterpIoU, a novel loss function that replaces handcrafted geometric penalties with a term based on the IoU between interpolated boxes and the target. By using interpolated boxes to bridge the gap between predictions and ground truth, InterpIoU provides meaningful gradients in non-overlapping cases and inherently avoids the box enlargement issue caused by misaligned penalties. Simulation results further show that IoU itself serves as an ideal regression target, while existing geometric penalties are both unnecessary and suboptimal. Building on InterpIoU, we introduce Dynamic InterpIoU, which dynamically adjusts interpolation coefficients based on IoU values, enhancing adaptability to scenarios with diverse object distributions. Experiments on COCO, VisDrone, and PASCAL VOC show that our methods consistently outperform state-of-the-art IoU-based losses across various detection frameworks, with particularly notable improvements in small object detection, confirming their effectiveness.

Figures

Figures reproduced from arXiv: 2507.12420 by the authors.

Figure 1
Figure 1. The geometry factors that are considered in previous [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the interpolation strategy in InterpIoU. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the relationship between the predicted [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The plot visualizes the loss value as a prediction box [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of the optimization behaviors of various IoU-based losses during regression. Each subfigure reports the average [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visualization of the simulated prediction distribution [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visualization of BBR using various IoU-based losses. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Comparison of relative Bbox area distributions across [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Visualization results on the VisDrone dataset. The first column shows the Ground Truth (GT), the second column shows the [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 36 canonical work pages

  1. [1]

    Yolov4: Optimal speed and accuracy of object detection, 2020

    Alexey Bochkovskiy, Chien-Yao Wang, and Hong- Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection, 2020. 1

    work page 2020

  2. [2]

    A method of object detection with attention mechanism and c2f dcnv2 for complex traffic scenes

    Fei Cai, Zhong Qu, Shufang Xia, and Shiyan Wang. A method of object detection with attention mechanism and c2f dcnv2 for complex traffic scenes. Expert Systems with Applications, 267:126141, 2025. 1

    work page 2025

  3. [3]

    Visdrone-det2021: The vision meets drone object de- tection challenge results

    Yaru Cao, Zhijian He, Lujia Wang, Wenguan Wang, Yix- uan Yuan, Dingwen Zhang, Jinglin Zhang, Pengfei Zhu, Luc Van Gool, Junwei Han, Steven Hoi, Qinghua Hu, and Ming Liu. Visdrone-det2021: The vision meets drone object de- tection challenge results. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Work- shops, pages 2847–28...

    work page 2021

  4. [4]

    Dynamic correlation learn- ing and regularization for multi-label confidence calibration

    Tianshui Chen, Weihang Wang, Tao Pu, Jinghui Qin, Zhijing Yang, Jie Liu, and Liang Lin. Dynamic correlation learn- ing and regularization for multi-label confidence calibration. IEEE Transactions on Image Processing, 2024. 1

    work page 2024

  5. [5]

    Object detection for autonomous vehicles under adverse weather conditions

    Zhige Chen, Zhigang Zhang, Qizheng Su, Kai Yang, Yan- dong Wu, Lei He, and Xiaolin Tang. Object detection for autonomous vehicles under adverse weather conditions. Ex- pert Systems with Applications, 296:128994, 2026. 1

    work page 2026

  6. [6]

    A virtually as- sisted digital twin enabled object detection in smart indus- trial manufacturing

    Veeresh Dachepalli and Sreelatha Gavini. A virtually as- sisted digital twin enabled object detection in smart indus- trial manufacturing. Expert Systems with Applications, 294: 128574, 2025. 1 13

    work page 2025

  7. [7]

    Dynamic fea- ture and context enhancement network for faster detection of small objects

    Shijiao Ding, Maomao Xiong, Xinghua Wang, Zekai Zhang, Qinghui Chen, Jinglin Zhang, Xing Wang, Zaigui Zhang, Dagang Li, Sen Xu, and Weiping Ding. Dynamic fea- ture and context enhancement network for faster detection of small objects. Expert Systems with Applications , 265: 125732, 2025. 2

    work page 2025

  8. [8]

    Adaptive feature fusion and task-dynamic alignment for real-time object detection on edge devices.Ex- pert Systems with Applications, 290:128341, 2025

    Yan Ding, Wenjie Wu, Yunan Zhai, Liyuan Zhang, Lin Pan, and Jia Zhao. Adaptive feature fusion and task-dynamic alignment for real-time object detection on edge devices.Ex- pert Systems with Applications, 290:128341, 2025. 1

    work page 2025

Show all 39 references
  1. [9]

    Everingham, S

    M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual ob- ject classes challenge: A retrospective.International Journal of Computer Vision, 111(1):98–136, 2015. 10

    2015

  2. [10]

    Lud-yolo: A novel lightweight object detection network for unmanned aerial vehicle

    Qingsong Fan, Yiting Li, Muhammet Deveci, Kaiyang Zhong, and Seifedine Kadry. Lud-yolo: A novel lightweight object detection network for unmanned aerial vehicle. Infor- mation Sciences, 686:121366, 2025. 1

    2025

  3. [11]

    Scott, and Weilin Huang

    Chengjian Feng, Yujie Zhong, Yu Gao, Matthew R. Scott, and Weilin Huang. Tood: Task-aligned one-stage object de- tection. In ICCV, pages 3490–3499. IEEE, 2021. 6

    2021

  4. [12]

    Rc-detr: Improving detrs in crowded pedestrian detection via rank- based contrastive learning

    Feng Gao, Jiaxu Leng, Ji Gan, and Xinbo Gao. Rc-detr: Improving detrs in crowded pedestrian detection via rank- based contrastive learning. Neural Networks, 182:106911,

  5. [13]

    Yolox: Exceeding yolo series in 2021, 2021

    Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021, 2021. 1, 3

    2021

  6. [14]

    Siou loss: More powerful learning for bounding box regression, 2022

    Zhora Gevorgyan. Siou loss: More powerful learning for bounding box regression, 2022. 1, 4, 8

    2022

  7. [15]

    Girshick, J

    R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich fea- ture hierarchies for accurate object detection and semantic segmentation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 580–587, 2014. 1

    2014

  8. [16]

    Girshick

    Ross B. Girshick. Fast r-cnn. In ICCV, pages 1440–1448. IEEE Computer Society, 2015. 1

    2015

  9. [17]

    Mfel-yolo for small object detection in uav aerial images

    Ting Hou, Chengcai Leng, Jiaxin Wang, Zhao Pei, Jinye Peng, Irene Cheng, and Anup Basu. Mfel-yolo for small object detection in uav aerial images. Expert Systems with Applications, 291:128459, 2025. 2

    2025

  10. [18]

    YOLOv5 by Ultralytics, 2020

    Glenn Jocher. YOLOv5 by Ultralytics, 2020. 1

    2020

  11. [19]

    Ultralytics yolov8, 2023

    Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8, 2023. 1, 10

    2023

  12. [20]

    Belongie, Lubomir D

    Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll’a r, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312,

    arXiv

  13. [21]

    Powerful-iou: More straightforward and faster bounding box regression loss with a nonmonotonic focusing mechanism

    Can Liu, Kaige Wang, Qing Li, Fazhan Zhao, Kun Zhao, and Hongtu Ma. Powerful-iou: More straightforward and faster bounding box regression loss with a nonmonotonic focusing mechanism. Neural Networks, 170:276–284, 2024. 1, 4, 7, 8

    2024

  14. [22]

    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. In Computer Vision – ECCV 2016 , pages 21–37, Cham, 2016. Springer International Publishing. 1, 10

    2016

  15. [23]

    Yolov3: An incremental improvement, 2018

    Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018. 1

    2018

  16. [24]

    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 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779–788, 2016. 1, 3

    2016

  17. [25]

    Girshick, and Jian Sun

    Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NIPS, pages 91–99, 2015. 1

    2015

  18. [26]

    Generalized in- tersection 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 in- tersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658–666,

  19. [27]

    Diff-mosaic: Augmenting realistic representations in infrared small target detection via diffu- sion prior

    Yukai Shi, Yupei Lin, Pengxu Wei, Xiaoyu Xian, Tianshui Chen, and Liang Lin. Diff-mosaic: Augmenting realistic representations in infrared small target detection via diffu- sion prior. IEEE Transactions on Geoscience and Remote Sensing, 62:1–11, 2024. 2

    2024

  20. [28]

    Multi- object garbage image detection algorithm based on sp-ssd

    Ling Tan, Hui Wu, Zifeng Xu, and Jingming Xia. Multi- object garbage image detection algorithm based on sp-ssd. Expert Systems with Applications, 263:125773, 2025. 1

    2025

  21. [29]

    Wise- iou: bounding box regression loss with dynamic focusing mechanism

    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. 4, 6, 7, 8

    2023 arXiv

  22. [30]

    Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors, 2022

    Chien-Yao Wang, Alexey Bochkovskiy, and Hong- Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors, 2022. 1

    2022

  23. [31]

    Density-guided two-stage small object detection in uav images

    Bojun Xie, Yanjie Wang, Meihui Han, Yidan Wang, and Jun- fen Chen. Density-guided two-stage small object detection in uav images. Expert Systems with Applications, page 129346,

  24. [32]

    Unitbox: An advanced object detection net- work

    Jiahui Yu, Yuning Jiang, Zhangyang Wang, Zhimin Cao, and Thomas Huang. Unitbox: An advanced object detection net- work. In Proceedings of the 24th ACM International Con- ference on Multimedia, page 516–520, New York, NY , USA,

  25. [33]

    Esod-yolov8: Small object detection enhanced with auto- disturbance rejection convolution

    Zhenhua Yu, Huize Liang, Ou Ye, and Yun Zhang. Esod-yolov8: Small object detection enhanced with auto- disturbance rejection convolution. Expert Systems with Ap- plications, 296:129046, 2026. 2

    2026

  26. [34]

    Ni, and Heung-Yeung Shum

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR. OpenReview.net, 2023. 1, 10

    2023

  27. [35]

    Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z. Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, pages 9756–9765. Computer Vision Foundation / IEEE, 2020. 6

    2020

  28. [36]

    Focal and efficient iou loss for accu- rate bounding box regression

    Yi-Fan Zhang, Weiqiang Ren, Zhang Zhang, Zhen Jia, Liang Wang, and Tieniu Tan. Focal and efficient iou loss for accu- rate bounding box regression. Neurocomputing, 506:146– 157, 2022. 4, 8 14

    2022

  29. [37]

    Distance-iou loss: Faster and bet- ter learning for bounding box regression

    Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and bet- ter learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, pages 12993– 13000, 2020. 1, 4, 8

    2020

  30. [38]

    Enhancing ge- ometric factors in model learning and inference for object detection and instance segmentation

    Zhaohui Zheng, Ping Wang, Dongwei Ren, Wei Liu, Rong- guang Ye, Qinghua Hu, and Wangmeng Zuo. Enhancing ge- ometric factors in model learning and inference for object detection and instance segmentation. IEEE transactions on cybernetics, 52(8):8574–8586, 2021. 1, 4, 7, 8 15

    2021

  31. [2016]

    Association for Computing Machinery. 1, 3

Pith tools

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