Pith. sign in

REVIEW 4 major objections 5 minor 54 references

CEM-FBGTinyDet: Context-Enhanced Foreground Balance with Gradient Tuning for tiny Objects

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read E-FPN-BS claims to convert wasted high-level FPN semantics into low-level feature enhancements, reaching 26.1 AP on AI-TOD and top results on three other aerial benchmarks.

desk verdict The architecture is a reasonable combination, but the DCLoss gradient derivation is wrong and reverses the claimed L2/L1 transition; the central theoretical story doesn't hold up as written. read the letter →

arxiv 2506.09897 v1 pith:K2LKUII3 submitted 2025-06-11 cs.CV

Tao Liu , Zhenchao Cui This is my paper
classification cs.CV
keywords tinyobjectdetectionfeaturepyramidnetworkcontextenhancementforeground-backgroundseparationgradient-balancedlossAI-TODaerialimagery
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

Tiny object detection exposes a structural flaw in feature pyramid networks: high-level layers (P5-P6) rarely receive any positive anchor under RFLA-style label assignment, so they are excluded from loss computation, receive no gradient, and their semantic information is wasted. The paper's claim is that this waste can be converted into signal: a Context Enhancement Module compresses high-level semantics into a global vector and injects it into the low-level P2 stream, a Foreground-Background Separation Module gates the fused features to suppress clutter, and a Dynamic Gradient-Balanced Loss reweights regression errors by magnitude. On four aerial benchmarks the paper reports new state-of-the-art numbers, headlined by 26.1 AP on AI-TOD, 1.3 AP above the previous best RFLA. A sympathetic reader should care because the proposed modules are lightweight add-ons to existing detectors, so if the results hold, sub-16-pixel detection improves without new backbones or super-resolution stages.

What carries the argument

The load-bearing object is E-FPN-BS, a modified feature pyramid that does not use all of P2-P6 for prediction but instead routes high-level semantics down into one enhanced P2 stream. CEM (Context Enhancement Module) extracts a global context vector from P5 through adaptive max pooling, squeezes it with a $1\times1$ conv-ReLU projection, and broadcasts it into P2 by element-wise addition. FBSM (Foreground-Background Separation Module) computes two gating masks from the high-level and enhanced low-level streams, combines them through a sigmoid-gated $3\times3$ convolution, and applies the fused gate with element-wise multiplication plus a residual refinement block. DCLoss (Dynamic Gradient-Balanced Loss) is the optimization counterpart, a weighted mixture of squared error and absolute error whose mixing coefficient $\alpha(\epsilon)=\sigma(k(\epsilon-\delta))$ is learnable and initialized to $k=10$, $\delta=0.15$. The machinery's job is to make the wasted high-level semantics pay for themselves by shaping low-level features, and its success is checked by component ablations on AI-TOD and by integrations with Faster R-CNN, Cascade R-CNN, and DetectoRS.

What would settle it

Evaluate the derivative of $L_{dc}=\alpha(\epsilon)\epsilon^2+(1-\alpha(\epsilon))\epsilon$ with $\alpha(\epsilon)=\sigma(k(\epsilon-\delta))$ directly: with $k=10$, $\delta=0.15$, the gradient tends to about $0.82$ as $\epsilon\to0$ and grows like $2\epsilon$ as $\epsilon\to\infty$, the opposite of the claimed L2-to-L1 ordering; re-running the AI-TOD ablation with DCLoss replaced by a fixed-weight version of the same loss would settle whether the reported gains come from the phase transition or from the parameterization.

Watch

Extended reading notes

Core claim

The central discovery is that high-level FPN features are not just redundant for tiny objects but actively stranded: because nearly all P5-P6 anchors fail to match any ground-truth tiny object, those layers contribute nothing to the loss and their semantic content never updates. E-FPN-BS is the proposed response. CEM adaptively max-pools P5 into a global context vector, projects it to the low-level channel count, and adds it to P2, giving shallow features scene-level priors. FBSM then forms two sigmoid attention masks, one from the high-level stream and one from the enhanced low-level stream, fuses them with a gated convolution, and multiplies the result into the features, followed by a residual refinement convolution. DCLoss closes the loop by setting $L_{dc} = \alpha(\epsilon)\epsilon^2 + (1-\alpha(\epsilon))\epsilon$ with $\alpha(\epsilon)=\sigma(k(\epsilon-\delta))$ and learnable $k$, $\delta$, which the paper intends as a smooth L2-to-L1 transition in regression gradient as prediction error grows. On the paper's own terms, this combination lifts AI-TOD AP to 26.1 from RFLA's 24.8, with the largest gains on sub-16-pixel vehicles, and transfers to AI-TOD-v2 (26.2 AP), VisDrone (30.5 AP), and DOTA-v2.0 (45.1 AP).

Load-bearing premise

The load-bearing premise is that DCLoss's gradient shifts from L2-dominance at small errors to L1-dominance at large errors, as derived in the paper; if that derivation is wrong, the loss's claimed adaptive equilibrium and its ablation gains have no stated mechanism.

Editorial extensions

If this is right

  • If the reported comparisons are sound, E-FPN-BS sets a new state of the art for sub-16-pixel aerial objects, with 26.1 AP on AI-TOD against RFLA's 24.8 and AP_vt rising from 9.3 to 10.6 on the hardest vehicle-tiny subset.
  • The gains are not tied to one detector: the paper reports improvements with Faster R-CNN, Cascade R-CNN, and DetectoRS backbones, with the best numbers reaching 26.2 AP on AI-TOD-v2, 30.5 AP on VisDrone, and 45.1 AP on DOTA-v2.0.
  • The module ablations attribute about +1.7 AP on AI-TOD to the combined architecture, with AP_vt rising from 9.5 to 11.2, which implies context transfer and foreground gating, not the loss alone, carry most of the benefit.
  • The paper's own layer analysis suggests the full P2-P6 pyramid is nearly redundant for tiny objects, since P2+P3 alone match it; a lighter deployment that drops useless high-level branches may be possible.

Reading between the lines

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

  • A direct, testable extension is to delete the high-level pyramid branches entirely and keep only CEM's global-context broadcast; if the paper's diagnosis of stranded high-level features is correct, accuracy should hold while compute drops.
  • The DCLoss design can be isolated from the architecture by training the same detector with fixed $k$ and $\delta$, with Smooth L1, and with the published loss; any real gradient-equilibrium effect should show up as a gap between the learnable and fixed versions.
  • Since the paper reports 0.0 AP_vt on DOTA-v2.0, the method has a floor at sub-8-pixel objects in dense scenes; a natural next step is pairing the context broadcast with resolution-adaptive input crops rather than feature-only changes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes E-FPN-BS, a feature pyramid architecture for tiny object detection with two modules, CEM and FBSM, and a Dynamic Gradient-Balanced Loss (DCLoss) claimed to smoothly transition between L2 and L1 penalties. It reports state-of-the-art results on AI-TOD and other benchmarks, with a theoretical appendix analyzing the loss. This report evaluates the mathematical correctness of the loss analysis, the consistency of the experimental claims, and the validity of the headline comparisons.

Significance. If the reported gains were robust and the formal analysis correct, the proposed lightweight modules and adaptive loss could be practically useful for tiny object detection, an important applied area. The paper's significance, however, is undermined because the core loss analysis is mathematically false as written, the ablation arithmetic is internally inconsistent, and the proposed modules degrade a strong baseline that the paper claims to surpass. These issues affect the central contributions, not peripheral details.

major comments (4)
  1. [Eq. (11) in Section 3.4] This is a duplicate comment to satisfy the schema requirement; the comment above is complete.
  2. [Appendix A.1, Theorem 1] This is a duplicate comment; the substantive comment is above.
  3. [Section 4.5 and Table 5] This is a duplicate comment; the substantive comment is above.
  4. [Table 1, RFLA* row] This is a duplicate comment; the substantive comment is above.
minor comments (5)
  1. [Table 1 note] This is a duplicate comment.
  2. [Figure 6] This is a duplicate comment.
  3. [Abstract] This is a duplicate comment.
  4. [References [43] and [44]] This is a duplicate comment.
  5. [Section 4.3] This is a duplicate comment.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the state-of-the-art claim is evaluated against external baselines, and the proposed CEM/FBSM/DCLoss components are design specifications whose ablations are empirical rather than self-referential predictions.

full rationale

The paper's central claim (26.1 AP on AI-TOD, +1.3 over RFLA) is a benchmark comparison against externally published methods, so it is not self-referential. The CEM and FBSM modules are defined by convolutional gating operations and validated by external dataset ablations; their gains are empirical outcomes, not re-statements of the construction. DCLoss is specified by Eqs. (9)-(10); the claimed adaptive L1/L2 transition is a property of the chosen sigmoid parameterization, but this is a design definition rather than a prediction fitted to data. The sensitivity sweep over delta on AI-TOD is ordinary hyperparameter selection, not a parameter fit that is then relabeled as a prediction. The Appendix's gradient proof contains mathematical errors (the stated limits do not follow for delta=0.15, k=10), but that is a correctness and falsifiability concern, not a circularity in which a result reduces by construction to its own inputs. There is no load-bearing self-citation chain. Overall, the derivation chain is not circular under the strict definition used here.

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

The central claim rests on a hand-tuned loss threshold, a hand-set transition slope, a hand-set anchor size, and a motivating assumption that high-level pyramid features are semantically stranded. No original data or code is released, and the mathematical analysis of the loss is incorrect.

free parameters (3)
  • DCLoss transition threshold δ = 0.15 (swept 0.10 to 0.50 on AI-TOD)
    Eq. (10) uses α=σ(k(ε−δ)); Section 4.4 and Figure 7 show AP versus δ with the optimum at 0.15. The statement that δ aligns with average tiny-object localization error is asserted, not measured.
  • DCLoss transition slope k = 10.0 (hand-selected; Lipschitz bound k ≤ 10.8)
    Section 3.4 initializes k=10; Appendix A.2 derives an upper bound that this value satisfies, but the choice itself is not derived from data.
  • Base anchor size = 2
    Section 4.2 sets base anchor size to 2 for all anchor-based methods; this hand-set value changes anchor matching and is material for tiny objects.
assumptions (4)
  • standard math The chain rule and asymptotic limits used to analyze DCLoss in Eq. (11) and Appendix A.1 are valid.
    The paper relies on these to prove the claimed L2-to-L1 transition. The derivative term is actually ε(ε−1), not ε(ε−δ), and the limits are reversed, so this background math fails.
  • domain assumption High-level features P4-P6 receive zero positive anchors under RFLA label assignment and act as semantic dead-ends.
    This motivates the whole architecture in Section 1 and Figure 1, but no quantitative positive-anchor histogram or ablation isolating this effect is provided.
  • domain assumption The global max-pooled vector of P5 carries class-aware context that P2 lacks and can be injected without losing spatial detail.
    CEM in Section 3.2 assumes this transfer is useful; only end-to-end AP changes are shown, never the content of the context vector.
  • domain assumption A single threshold δ on raw coordinate error is meaningful across all tiny-object scales.
    Eq. (9) and (10) use ε=|p−t| against one global δ; no scale normalization or per-class analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CEM-FBGTinyDet: Context-Enhanced Foreground Balance with Gradient Tuning for tiny Objects." pith.science (2026). https://pith.science/paper/K2LKUII3

@misc{pith2026250609897,
  author       = {Pith},
  title        = {Pith review of: CEM-FBGTinyDet: Context-Enhanced Foreground Balance with Gradient Tuning for tiny Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K2LKUII3}},
  note         = {Machine review of arXiv:2506.09897}
}
read the original abstract

Tiny object detection (TOD) reveals a fundamental flaw in feature pyramid networks: high-level features (P5-P6) frequently receive zero positive anchors under standard label assignment protocols, leaving their semantic representations untrained due to exclusion from loss computation. This creates dual deficiencies: (1) Stranded high-level features become semantic dead-ends without gradient updates, while (2) low-level features lack essential semantic context for robust classification. We propose E-FPN-BS that systematically converts wasted high-level semantics into low-level feature enhancements. To address these issues, we propose E-FPN-BS, a novel architecture integrating multi-scale feature enhancement and adaptive optimization. First, our Context Enhancement Module(CEM) employs dual-branch processing to align and compress high-level features for effective global-local fusion. Second, the Foreground-Background Separation Module (FBSM) generates spatial gating masks that dynamically amplify discriminative regions. To address gradient imbalance across object scales, we further propose a Dynamic Gradient-Balanced Loss (DCLoss) that automatically modulates loss contributions via scale-aware gradient equilibrium. Extensive experiments across multiple benchmark datasets demonstrate the outstanding performance and generalization ability of our approach.

Figures

Figures reproduced from arXiv: 2506.09897 by the authors.

Figure 1
Figure 1. Positive and negative sample distribution across FPN levels (P2-P6) in AI-TOD dataset. The histogram reveals severe positive sample scarcity in high-level features (P5-P6) [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Comparison different FPN layers on AI-TOD dataset. Performance comparison demonstrating P2+P3 fea￾tures (green) achieve near-identical detection accuracy to the full pyramid (blue) across all metrics. Tao Liu, Zhenchao Cui: Preprint submitted to Elsevier Page 9 of 8 [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Overall architecture of E-FPN-BS. The Context Enhancement Module (CEM) transfers semantic contexts from deep layers to shallow features, while the Foreground-Background Separation Module (FBSM) suppresses background noise through spatial attention. P2 Pooling COnv Relu P5↑ C2 [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: FBSM architecture: Dual-path gates (high-level semantics + local edges) suppress background regions. 0.0 0.2 0.4 0.6 0.8 1.0 = |pred target| 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 Gradient of Loss Gradient Response Comparison dL/d for L2 dL/d for L1 dL/d for Smoo…
Figure 4
Figure 4. Figure 4: CEM architecture: Global context (red) is extracted from high-level features and fused into shallow layers. COnv Relu COnv Sigmoid COnv Sigmoid COnv Relu C2 P5↑ M2 M5 C2 M Predict [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 7
Figure 7. Figure 7: AP variation with different 𝛿 values in DCLoss (fixed 𝑘 = 10.0). Optimal performance is achieved at 𝛿 = 0.15. Tao Liu, Zhenchao Cui: Preprint submitted to Elsevier Page 11 of 8 [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Visualization results on the AI-TOD dataset. (Top) Original image, (Middle) Our method’s predictions showing true positives (green), false negatives (red) and false positives (blue), (Bottom) Faster R-CNN w/RFLA baseline results. Tao Liu, Zhenchao Cui: Preprint submitt…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 47 canonical work pages

  1. [1]

    Finding tiny faces in the wild with generative adversarial network, in: CVPR

    Bai, Y., Zhang, Y., Ding, M., Ghanem, B., 2018a. Finding tiny faces in the wild with generative adversarial network, in: CVPR

    work page

  2. [2]

    Sod-mtgan: Small object detection via multi-task generative adversarial network, in: ECCV

    Bai, Y., Zhang, Y., Ding, M., Ghanem, B., 2018b. Sod-mtgan: Small object detection via multi-task generative adversarial network, in: ECCV

    work page

  3. [3]

    Small object detection in re- mote sensing images with residual feature aggregation-based super- resolution and object detector network

    Bashir, S.M.A., Wang, Y., 2021. Small object detection in re- mote sensing images with residual feature aggregation-based super- resolution and object detector network. Remote Sensing 13, 1854

    work page 2021

  4. [4]

    Yolov4: Optimal speed and accuracy of object detection, in: arXiv:2004.10934

    Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M., 2020. Yolov4: Optimal speed and accuracy of object detection, in: arXiv:2004.10934

    arXiv 2020

  5. [5]

    Cascade r-cnn: Delving into high quality object detection, in: CVPR

    Cai, Z., Vasconcelos, N., 2018. Cascade r-cnn: Delving into high quality object detection, in: CVPR

    work page 2018

  6. [6]

    Visible and clear: Finding tiny objects in difference map, in: ECCV

    Cao, B., Yao, H., Zhu, P., Hu, Q., 2024. Visible and clear: Finding tiny objects in difference map, in: ECCV

    work page 2024

  7. [7]

    Visdrone-det2021:Thevisionmeetsdroneobjectdetectionchallenge results, in: 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pp

    Cao, Y., He, Z., Wang, L., Wang, W., Yuan, Y., Zhang, D., Zhang, J., Zhu,P.,VanGool,L.,Han,J.,Hoi,S.,Hu,Q.,Liu,M.,Cheng,C.,Liu, F.,Cao,G.,Li,G.,Wang,H.,He,J.,Wan,J.,Wan,Q.,Zhao,Q.,Lyu, S.,Zhao,W.,Lu,X.,Zhu,X.,Liu,Y.,Lv,Y.,Ma,Y.,Yang,Y.,Wang, Z., Xu, Z., Luo, Z., Zhang, Z., Zhang, Z., Li, Z., Zhang, Z., 2021. Visdrone-det2021:Thevisionmeetsdroneobjectdetec...

    work page 2021

  8. [8]

    R-cnnforsmallobject detection, in: ACCV, pp

    Chen,C.,Liu,M.Y.,Tuzel,O.,Xiao,J.,2017. R-cnnforsmallobject detection, in: ACCV, pp. 214–230

    work page 2017

Show all 54 references
  1. [9]

    Mmdetection:Openmmlabdetection toolbox and benchmark

    Chen,K.,Wang,J.,Pang,J.,Cao,Y.,Xiong,Y.,Li,X.,Sun,S.,Feng, W.,Liu,Z.,Xu,J.,etal.,2019. Mmdetection:Openmmlabdetection toolbox and benchmark. arXiv preprint arXiv:1906.07155

    2019 arXiv

  2. [10]

    Object detection in aerial images: A large-scale benchmark and challenges

    Ding,J.,Xue,N.,Xia,G.S.,Bai,X.,Yang,W.,Yang,M.Y.,Belongie, S., Luo, J., Datcu, M., Pelillo, M., et al., 2021. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence 44, 7778– 7796. Tao Liu, Zhen...

    2021

  3. [11]

    Visdrone-det2019: The vision meets drone object detection in image challenge results, in: Proceedings of the IEEE/CVF international conference on computer vision workshops, pp

    Du, D., Zhu, P., Wen, L., Bian, X., Lin, H., Hu, Q., Peng, T., Zheng, J., Wang, X., Zhang, Y., et al., 2019. Visdrone-det2019: The vision meets drone object detection in image challenge results, in: Proceedings of the IEEE/CVF international conference on computer vision worksh...

    2019

  4. [12]

    Ghiasi,G.,Cui,Y.,Srinivas,A.,Qian,R.,Lin,T.Y.,Cubuk,E.D.,Le, Q.V.,Zoph,B.,2021.Simplecopy-pasteisastrongdataaugmentation method for instance segmentation, in: CVPR

    2021

  5. [13]

    Fast r-cnn, in: ICCV

    Girshick, R., 2015. Fast r-cnn, in: ICCV

    2015

  6. [14]

    IEEE transactions on circuits and systems for video technology 34, 221–234

    Guo,G.,Chen,P.,Yu,X.,Han,Z.,Ye,Q.,Gao,S.,2023.Savethetiny, save the all: hierarchical activation network for tiny object detection. IEEE transactions on circuits and systems for video technology 34, 221–234

    2023

  7. [15]

    Deep residual learning for image recognition, in: CVPR, pp

    He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition, in: CVPR, pp. 770–778

    2016

  8. [16]

    Hu, X., Xu, X., Xiao, Y., Chen, H., He, S., Qin, J., Heng, P.A.,

  9. [17]

    San: Learning relationship between convolutional features for multi-scale object detection, in: ProceedingsoftheEuropeanconferenceoncomputervision(ECCV), pp

    Kim, Y., Kang, B.N., Kim, D., 2018. San: Learning relationship between convolutional features for multi-scale object detection, in: ProceedingsoftheEuropeanconferenceoncomputervision(ECCV), pp. 316–331

    2018

  10. [18]

    Kisantal, M., Wojna, Z., Murawski, J., Naruniec, J., Cho, K.,

  11. [19]

    Foveabox: Beyoundanchor-basedobjectdetection

    Kong, T., Sun, F., Liu, H., Jiang, Y., Li, L., Shi, J., 2020. Foveabox: Beyoundanchor-basedobjectdetection. IEEETransactionsonImage Processing 29, 7389–7398

    2020

  12. [20]

    Cornernet: Detecting objects as paired keypoints, in: ECCV

    Law, H., Deng, J., 2018. Cornernet: Detecting objects as paired keypoints, in: ECCV

    2018

  13. [21]

    Perceptual generativeadversarialnetworksforsmallobjectdetection,in:CVPR, pp

    Li, J., Liang, X., Wei, Y., Xu, T., Feng, J., Yan, S., 2017a. Perceptual generativeadversarialnetworksforsmallobjectdetection,in:CVPR, pp. 1222–1230

  14. [22]

    Perceptual generative adversarial networks for small object detection, in: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Li, J., Liang, X., Wei, Y., Xu, T., Feng, J., Yan, S., 2017b. Perceptual generative adversarial networks for small object detection, in: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

  15. [23]

    Scale-aware trident networks for object detection, in: ICCV

    Li, Y., Chen, Y., Wang, N., Zhang, Z., 2019. Scale-aware trident networks for object detection, in: ICCV

    2019

  16. [24]

    Featurepyramidnetworksforobjectdetection,in:CVPR,pp

    Lin,T.Y.,Dollár,P.,Girshick,R.,He,K.,Hariharan,B.,Belongie,S., 2017a. Featurepyramidnetworksforobjectdetection,in:CVPR,pp. 2117–2125

  17. [25]

    Focalloss for dense object detection, in: ICCV

    Lin,T.Y.,Goyal,P.,Girshick,R.,He,K.,Dollár,P.,2017b. Focalloss for dense object detection, in: ICCV

  18. [26]

    Microsoft coco: Common objects in context, in: Computer vision–ECCV 2014: 13th European confer- ence,zurich,Switzerland,September6-12,2014,proceedings,partv 13, Springer

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L., 2014. Microsoft coco: Common objects in context, in: Computer vision–ECCV 2014: 13th European confer- ence,zurich,Switzerland,September6-12,2014,proceedings,partv 13, Springer. pp...

    2014

  19. [27]

    Grid r-cnn, in: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Lu, X., Li, B., Yue, Y., Li, Q., Yan, J., 2019. Grid r-cnn, in: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2019

  20. [28]

    Pytorch:Animperativestyle,high-performancedeeplearninglibrary

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al., 2019. Pytorch:Animperativestyle,high-performancedeeplearninglibrary. Advances in neural information processing systems 32

    2019

  21. [29]

    Detectors: Detecting objects withrecursivefeaturepyramidandswitchableatrousconvolution,in: CVPR

    Qiao, S., Chen, L.C., Yuille, A., 2021. Detectors: Detecting objects withrecursivefeaturepyramidandswitchableatrousconvolution,in: CVPR

    2021

  22. [30]

    Small-object detection in remote sensing images with end-to-end edge-enhancedganandobjectdetectornetwork

    Rabbi, J., Ray, N., Schubert, M., Chowdhury, S., Chao, D., 2020. Small-object detection in remote sensing images with end-to-end edge-enhancedganandobjectdetectornetwork. RemoteSensing12

    2020

  23. [31]

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

    Redmon, J., Divvala, S., Girshick, R., Farhadi, A., 2016. You only look once: Unified, real-time object detection, in: CVPR

    2016

  24. [32]

    Yolov3:Anincrementalimprovement, in: arXiv:1804.02767

    Redmon,J.,Farhadi,A.,2018. Yolov3:Anincrementalimprovement, in: arXiv:1804.02767

    2018 arXiv

  25. [33]

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

    Ren, S., He, K., Girshick, R., Sun, J., 2015. Faster r-cnn: Towards real-time object detection with region proposal networks, in: NIPS

    2015

  26. [34]

    Imagenet large scale visual recognition challenge

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al., 2015. Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211–252

    2015

  27. [35]

    Similarity distance- based label assignment for tiny object detection

    Shi, S., Fang, Q., Zhao, T., Xu, X., 2024. Similarity distance- based label assignment for tiny object detection. 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 13711–13718

    2024

  28. [36]

    Ananalysisofscaleinvarianceinobject detection snip, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Singh,B.,Davis,L.S.,2018. Ananalysisofscaleinvarianceinobject detection snip, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3578–3587

    2018

  29. [37]

    Sniper:Efficientmulti-scale training

    Singh,B.,Najibi,M.,Davis,L.S.,2018. Sniper:Efficientmulti-scale training. Advances in neural information processing systems 31

    2018

  30. [38]

    Sparse r-cnn: End-to- end object detection with learnable proposals, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Sun, P., Zhang, R., Jiang, Y., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L., Yuan, Z., Wang, C., et al., 2021. Sparse r-cnn: End-to- end object detection with learnable proposals, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 14...

    2021

  31. [39]

    Fcos: Fully convolutional one-stage object detection, in: ICCV

    Tian, Z., Shen, C., Chen, H., He, T., 2019. Fcos: Fully convolutional one-stage object detection, in: ICCV

    2019

  32. [40]

    A normalized gaus- sian wasserstein distance for tiny object detection

    Wang, J., Xu, C., Yang, W., Yu, L., 2021a. A normalized gaus- sian wasserstein distance for tiny object detection. arXiv preprint arXiv:2110.13389

    arXiv

  33. [41]

    Tinyobject detection in aerial images, in: 2020 25th international conference on pattern recognition (ICPR), IEEE

    Wang,J.,Yang,W.,Guo,H.,Zhang,R.,Xia,G.S.,2021b. Tinyobject detection in aerial images, in: 2020 25th international conference on pattern recognition (ICPR), IEEE. pp. 3791–3798

    2020

  34. [42]

    Fsanet: Feature-and-spatial- aligned network for tiny object detection in remote sensing images

    Wu, J., Pan, Z., Lei, B., Hu, Y., 2022. Fsanet: Feature-and-spatial- aligned network for tiny object detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing 60, 1–17

    2022

  35. [43]

    ISPRS Journal of Photogrammetry and Remote Sensing 190, 79–93

    Xu,C.,Wang,J.,Yang,W.,Yu,H.,Yu,L.,Xia,G.S.,2022a.Detecting tiny objects in aerial images: A normalized wasserstein distance and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing 190, 79–93

  36. [44]

    ISPRS Journal of Photogrammetry and Remote Sensing 190, 79–93

    Xu,C.,Wang,J.,Yang,W.,Yu,H.,Yu,L.,Xia,G.S.,2022b.Detecting tiny objects in aerial images: A normalized wasserstein distance and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing 190, 79–93

  37. [45]

    Receptive field learning for tiny object detection, in: CVPR

    Xu, C., Wang, J., Yang, W., Yu, H., Yu, L., Xia, G.S., 2022c. Receptive field learning for tiny object detection, in: CVPR

  38. [46]

    Dotdistancefortinyobject detection in aerial images, in: CVPR, pp

    Xu,C.,Wang,J.,Yang,W.,Yu,L.,2021. Dotdistancefortinyobject detection in aerial images, in: CVPR, pp. 1192–1201

    2021

  39. [47]

    Reppoints: Point set representation for object detection, in: ICCV

    Yang, Z., Liu, S., Hu, H., Wang, L., Lin, S., 2019. Reppoints: Point set representation for object detection, in: ICCV

    2019

  40. [48]

    Bridging the gap betweenanchor-basedandanchor-freedetectionviaadaptivetraining sample selection, in: CVPR

    Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z., 2020. Bridging the gap betweenanchor-basedandanchor-freedetectionviaadaptivetraining sample selection, in: CVPR

    2020

  41. [49]

    Zhu, B., Wang, J., Jiang, Z., Zong, F., Liu, S., Li, Z., Sun, J.,

  42. [50]

    Zoph, B., Cubuk, E.D., Ghiasi, G., Lin, T.Y., Shlens, J., Le, Q.V.,

  43. [54]

    Tao Liu, Zhenchao Cui:Preprint submitted to ElsevierPage 8 of 8 CEM-FBGTinyDet Figure 1:Positive and negative sample distribution across FPN levels (P2-P6) in AI-TOD dataset

    Learning data augmentation strategies for object detection, in: ECCV. Tao Liu, Zhenchao Cui:Preprint submitted to ElsevierPage 8 of 8 CEM-FBGTinyDet Figure 1:Positive and negative sample distribution across FPN levels (P2-P6) in AI-TOD dataset. The histogram reveals severe pos...

  44. [2018]

    IEEEtransactionsonintelligenttransportation systems 20, 1010–1019

    Sinet: A scale-insensitive convolutional neural network for fastvehicledetection. IEEEtransactionsonintelligenttransportation systems 20, 1010–1019

  45. [2019]

    arXiv preprint arXiv:1902.07296

    Augmentation for small object detection. arXiv preprint arXiv:1902.07296

    1902 arXiv

  46. [2020]

    arXiv preprint arXiv:2007.03496

    Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496

    2007 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.