LiM-YOLO: Less is More with Pyramid Level Shift for Ship Detection in Optical Remote Sensing

Daewon Chung; Hyeji Sim; Okchul Jung; Seon-Hoon Kim; Yerin Kim; Youeyun Jung

REVIEW 3 major objections 6 minor 36 references

LiM-YOLO: Less is More with Pyramid Level Shift for Ship Detection in Optical Remote Sensing

T0 review · 3 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Satellite ship detection becomes both lighter and more accurate when the deepest pyramid layer is removed and a finer stride-4 head is added, cutting parameters by 64% while beating detectors up to three times larger.

desk verdict The P2-P4 pyramid shift is a real, useful empirical recipe for ship detection, but the SOTA comparison and the Nyquist framing overreach; treat the ablation as the contribution and the headline table as unverified. read the letter →

arxiv 2512.09700 v4 pith:LAFNWRXR submitted 2025-12-10 cs.CV eess.IV

Seon-Hoon Kim , Yerin Kim , Hyeji Sim , Youeyun Jung , Okchul Jung , Daewon Chung This is my paper

classification cs.CVeess.IV

keywords shipdetectionopticalremotesensingYOLOfeaturepyramidlevelshiftsmallobjectorientedboundingboxessamplingtheorygroupnormalization

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

The paper argues that general-purpose YOLO detectors are structurally mismatched to satellite ship detection because their deepest feature level (stride 32) shrinks narrow vessels to sub-pixel representations, and that fixing this is not a matter of adding a high-resolution head but of removing the deepest layer entirely. By shifting the detection pyramid from P3-P5 to P2-P4 — adding a stride-4 head and pruning the stride-32 head and backbone — the resulting detector, LiM-YOLO, uses 64.1% fewer parameters than its YOLOv9-E baseline and yet achieves better accuracy on four ship benchmarks, including a 0.600 mAP50:95 on the integrated set, beating detectors up to three times larger. The authors support the design with a statistical analysis of ship minor-axis sizes across four datasets, an occupancy condition (ρ_minor ≥ 1) grounded in sampling theory, and an effective-receptive-field argument that P5 is redundant because P4 already spans the whole image context. They also introduce a group-normalized linear projection branch so the model trains stably at batch size 2, which memory limits force when processing 1024×1024 satellite patches. If right, this shows that for a domain with skewed, anisotropic object sizes, pruning deep layers plus adding a high-resolution head yields a better accuracy-efficiency trade-off than scaling up depth.

What carries the argument

Pyramid Level Shift Strategy: reconfiguring the detection head from P3-P5 (strides 8, 16, 32) to P2-P4 (strides 4, 8, 16), i.e., adding a high-resolution P2 head and pruning the P5 head along with the deepest backbone stage. The design is justified by the Minor Axis Occupancy Ratio ρ_minor = L_minor/S: a target must occupy at least one feature-map cell (ρ_minor ≥ 1) to avoid what the authors call sub-pixel spatial feature dilution, and only a stride-4 level satisfies this for ships whose minor axes reach the 4-pixel lower bound. A companion Group Normalized CBLinear module inserts group normalization into the auxiliary gradient branch, stabilizing training at batch size 2. The ERF analysis o

What would settle it

A controlled experiment would settle the mechanism: take YOLOv9-E with P3-P5 and reallocate the parameters that LiM-YOLO prunes (the P5 head and backbone stage) to widen the P2-P3 channels, keeping the same total parameter budget as LiM-YOLO. If this widened P3-P5 model matches LiM-YOLO's accuracy, the gain comes from added capacity at high resolution rather than from removing P5; if it falls short, the P5-removal claim is supported.

Watch

Extended reading notes

Core claim

LiM-YOLO's central claim is that the standard YOLO pyramid P3-P5 is structurally misaligned with ships in satellite imagery. Ship minor axes across four benchmarks average 17.34 px with a 95% range down to 4 px; at stride 32 the average ship occupies about half a grid cell (ρ_minor≈0.54), losing boundary and orientation cues. The paper derives a spatial representability condition, ρ_minor = L_minor/S ≥ 1, from classical sampling theory, making a stride-4 P2 head the only level that satisfies it for the narrowest ships. Effective-receptive-field analysis shows P4's ERF (~1024 px) already spans the full 1024×1024 input, so P5 is informationally redundant and adds background noise. Shifting the

Load-bearing premise

The central premise is that a network's stride behaves like a sampling interval, so a ship narrower than one grid cell suffers irretrievable aliasing; the paper asserts this from sampling theory but offers no formal theorem, and its own stride-8 baseline does detect many ships whose minor axes are below 8 pixels.

Editorial extensions

If this is right

Removing the P5 head and backbone cuts parameters by 64.1% (58.99M to 21.16M) and GFLOPs by 3.5% while improving mAP50:95 over the P3-P5 baseline on all four datasets.
On the integrated ship dataset LiM-YOLO reaches 0.600 mAP50:95, surpassing YOLOv8x (0.566), YOLOv10x (0.543), YOLO11x (0.554), YOLOv12x (0.494) and RT-DETR-X (0.545) with the fewest parameters (21.16M).
The P2 head specifically recovers small high-aspect vessels: on ShipRSImageNet, Motorboat mAP rises from 10.7% to 13.8% and Sailboat from 3.3% to 16.2%.
Pruning P5 costs some accuracy on the largest classes (Aircraft Carrier drops from 63.9% to 53.8% on ShipRSImageNet), but this is outweighed by gains on small and mid-size vessels; P4 must be kept because dropping it collapses average mAP to 32.5%.
GN-CBLinear, using group normalization, stabilizes training at micro-batch size and improves mAP50:95 consistently, e.g., from 0.428 to 0.448 on ShipRSImageNet.

Reading between the lines

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

The sampling-theoretic framing is an analogy, not a proof: the paper never shows formally that a convolutional detector cannot represent objects smaller than one grid cell, and its own stride-8 baseline detects many such objects. The gains could plausibly come from higher feature-map resolution alone, independent of aliasing theory.
The 'less is more' result likely transfers to other detection domains with extreme size skew and anisotropic targets — tiny-person detection, medical cell counting, small UAV detection — where the deepest pyramid level is rarely informative and P4 already spans the input context. A reasonable test would be applying the P2-P4 shift to other one-stage detectors on those datasets.
The design implicitly predicts that the exact choice of normalization (GN vs LayerNorm vs RMSNorm) in the auxiliary branch is secondary, since the motivation is micro-batch stability; a simpler replacement that is batch-independent should show similar gains.
The paper's own class-wise numbers imply a bounded trade-off: the largest ships lose accuracy when P5 is pruned. A testable extension would add a lightweight P5 head only when large objects are present (an adaptive depth router), which could recover Aircraft Carrier performance without much parameter cost.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

Short version: the paper has one honest, useful empirical result — shifting the YOLOv9 detection head from P3-P5 to P2-P4 (drop P5, add P2) cuts parameters by about 64% and gives modest but consistent mAP gains on four ship benchmarks under a controlled ablation protocol. That recipe is worth knowing. But the SOTA comparison table that anchors the abstract is built on an undefined "Integrated Ship Detection Dataset," and the Nyquist story is more branding than proof. Treat the headline with suspicion; the ablation is the real contribution.

What is actually new: prior specialized detectors added P2 (YOLO-Ssboat) or P6 (YOLO-RSA) while keeping P5. Removing P5 and shifting to P2-P4 is a distinct configuration, and the ablation design is fair: from-scratch training, augmentation disabled, same protocol across all model variants. The GN-CBLinear addition is minor but reasonable, and the class-wise Table IX at least shows where gains and losses land.

Soft spots, in rough order of severity. First, Table VIII: the "Integrated Ship Detection Dataset" is never defined — no composition, no class reconciliation, no protocol for how the baselines were trained or evaluated. That makes the central "surpasses models three times larger" claim unverifiable. Second, the Nyquist argument: rho_minor >= 1 is just a restatement of "object should be wider than the stride," not a sampling-theoretic derivation, and the paper provides no proof that a stride-8 detector cannot represent 4-pixel objects. The empirical gains can be explained by feature-map resolution alone. Third, the P5-redundancy story is overstated: the paper labels P5 "architecturally redundant," yet its own class-wise ablation shows the largest classes (e.g., Aircraft Carrier at 71.2% to 66.9%) degrade when P5 is removed. That is a real trade-off, not a free lunch. Fourth, the qualitative section presents a false positive as a success — a "Frigate" that is not in the ground truth is celebrated as evidence of detection power. That needs fixing. Fifth, reference hygiene: the SODA-A citation is wrong (it points to a low-power radio architecture paper), and there are repeated/misnumbered references around Group Normalization.

The internal ablation is credible and likely reproducible; the SOTA table and the theory are not. This paper deserves peer review because the architectural recipe is useful and the ablation is competently run — but the reviewers should demand the integrated dataset be defined, baselines be re-run under a stated protocol, and the Nyquist/P5-redundancy claims be toned down.

Regards.

Referee Report

3 major / 6 minor

Summary. The paper proposes LiM-YOLO, a ship-detection detector built on YOLOv9-E. Based on a statistical analysis of ship sizes across SODA-A, DOTA-v1.5, FAIR1M-v2.0, and ShipRSImageNet-V1, the authors shift the detection pyramid from P3-P5 (strides 8,16,32) to P2-P4 (strides 4,8,16), pruning P5 and adding a high-resolution P2 head. They justify this shift with a Nyquist-sampling argument and introduce a Group-Normalized CBLinear (GN-CBLinear) auxiliary branch to stabilize micro-batch training. The central empirical claim is that LiM-YOLO, with 21.16M parameters, achieves mAP50:95 of 0.600 on an 'Integrated Ship Detection Dataset,' surpassing larger SOTA detectors (e.g., YOLOv8x with 69.47M parameters). Internal ablations (Tables IV-VII) show consistent improvements from the P-level shift and GN-CBLinear under a fixed training protocol.

Significance. If verified, the P2-P4 shift is a practically valuable architectural insight for small-object detection in remote sensing: it cuts parameters by 64% while improving accuracy over the YOLOv9-E baseline in controlled ablations. The ablation design (from-scratch training, augmentation disabled, same protocol across variants) is a strength and makes the internal comparisons credible. The paper also ships code, which aids reproducibility. However, the headline SOTA claim is currently not verifiable because the 'Integrated Ship Detection Dataset' is undefined and the baseline training protocol is ambiguous. The Nyquist-Shannon justification is an analogy, not a derivation, and the 'structural necessity' of stride 4 is overstated. These issues limit the current contribution to a well-executed empirical study rather than a rigorously established general principle.

major comments (3)

[Section VII-B, Table VIII] The 'Integrated Ship Detection Dataset' for the headline SOTA comparison is never defined. Section VI-A describes four datasets with different class schemes (1, 9, 24 classes) and GSDs; the composition of the integrated set, how OBB labels are reconciled, and how mAP@50:95 is computed across heterogeneous class sets are omitted. Without this, the claim of surpassing detectors up to three times larger (0.600 vs. 0.566 for YOLOv8x) cannot be verified or reproduced. Specify the dataset composition, cropping, label mapping, and evaluation code.
[Section VI-B / Table VIII] It is ambiguous whether the 'stringent experimental protocol' (from-scratch training, augmentation disabled, batch size 2) applies to the SOTA baselines in Table VIII. If it does, the comparison is unfair because public baselines typically rely on pretrained backbones and augmentation; if it does not, the training setup for the baselines is missing. State the protocol for every model in Table VIII and justify comparability with from-scratch, augmentation-free training.
[Section III-B / IV-A, Eq. (1)] The Nyquist-Shannon argument is an unproven analogy. The condition ρ_minor ≥ 1 is a restatement that the object minor axis must be at least one feature-map stride, and the 4-pixel lower bound is derived from the same data statistics used to motivate stride 4, making the 'structural necessity' of P2 circular. No formal argument shows that a CNN cannot represent objects smaller than one grid cell; the paper's own baseline detects many such objects. Please replace the sampling-theoretic derivation with a rigorous statement or explicitly label it as an empirical design heuristic.

minor comments (6)

[Section V-B, Eq. (6)] The GN-CBLinear equation uses γ and β without defining them; specify that these are the learnable affine parameters of GroupNorm. Also, the discussion of batch-size-dependent normalization cites [36] (the ERF paper) when it should cite [18] (Wu and He, Group Normalization).
[Section III-A.1] Reference [19] for SODA-A is incorrect: the cited paper is about a low-power architecture for software radio, not the SODA small-object detection benchmark. Please provide the correct citation (e.g., the SODA dataset paper by Cheng et al.).
[References] Reference [17] (Shannon) is dated 2006; the original is from 1949 (with later reprints). Update to the correct original or a canonical reprint.
[Section IV-B / Table II] The ERF values (e.g., 667.7, 860.5) are presented without describing how they were computed. Add a sentence on the gradient-based measurement method and the network weights used, so the numbers can be reproduced.
[Throughout] The text contains typos and grammatical issues, e.g., 'p roposed' (Section II-C), 'rigoroulsy validatae' (Section VI-A), 'Recptive Field' (Fig. 4), and 'coasta' (Section II-B). A thorough language edit is needed.
[Section I / Tables IV-VII] The claim that P5 introduces 'excessive background context' is causal; the ablations show consistent but modest gains from pruning P5 (e.g., SODA-A mAP50-95 0.656→0.660, ShipRSImageNet 0.415→0.428). Consider adding a direct analysis of feature-map content or ERF to support the mechanism.

Circularity Check

1 steps flagged · score 4.0 of 10

The 'Nyquist necessity' of P2 is the paper's own rho_minor>=1 occupancy condition restated, not an independent derivation; the empirical ablations remain independent.

self definitional [Section III-B (Minor Axis Distribution) and Eq. (1); also Section IV-A]
"Translating this principle from the time domain to the spatial grid of CNNs, the network's stride acts as the sampling interval. Consequently, to resolve ship features approaching the 4-pixel lower bound, the sampling grid must be sufficiently dense to capture their spatial variations. A stride of 8 violates this Nyquist criterion [17] for such small targets. Therefore, adopting a stride of 4 (P2) is not merely beneficial but architecturally essential to preserve the spatial integrity of narrow vessel features against aliasing."

The claimed derivation is the paper's own definition: rho_minor = L_minor/S, with the imposed condition rho_minor >= 1. Since Table I gives the 95% lower bound of the minor axis as 4 pixels, rho_minor >= 1 is algebraically equivalent to S <= 4. Thus 'stride 4 is architecturally essential' is just a restatement of the occupancy inequality the authors introduced ('a target must occupy at least one grid cell'), relabeled as a Nyquist/Shannon necessity. The same data statistics (the 4-pixel lower bound) are used both to select the stride and to justify it, so the first-principles claim reduces to the definition. It is not a fitted accuracy parameter and does not affect the independent value of the controlled ablations.

full rationale

The central empirical contribution is not circular: the stepwise ablations (Tables IV-VII) compare YOLOv9-E with P3-P5, the +P2 expansion, the P2-P4 shift, and the GN-CBLinear module under a stated protocol, and the reported gains are internally controlled. The P5-pruning argument based on measured ERF values is empirical and falsifiable rather than definitional. I found no load-bearing self-citation: the references used for theoretical support (Shannon, Group Normalization, ERF) are external and not authored by the present authors. The main circularity is confined to the rhetorical 'Nyquist' justification of the P2 head, where the conclusion follows from the paper's own occupancy ratio definition plus the data-derived 4-pixel lower bound. Separately, the SOTA comparison in Table VIII rests on an 'Integrated Ship Detection Dataset' that is never defined, and the baseline training protocol is ambiguous; these are reproducibility/verifiability concerns, not circularity, so they do not increase the circularity score. Overall, the derivation of 'P2 is structurally necessary' is partly self-definitional, but because the paper's actual performance claims are supported by independent ablations, the score is moderate rather than high.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities, but its design rests on several domain assumptions, most notably the Nyquist-to-stride analogy and the sufficiency of P4's ERF. The free parameters are architectural choices fitted to dataset statistics rather than learned weights.

free parameters (4)

Pyramid stride configuration (P2-P4) = S=4, 8, 16
Chosen from the 95% lower bound of the minor-axis distribution (4 px) to satisfy rho_minor>=1; a data-fitted design choice rather than a learned parameter.
95% quantile threshold for design = 95% range [4, 64] px (minor axis)
The decision to design for the 95% range rather than the full min-max range is arbitrary and affects the architectural conclusion.
Group count G in GN-CBLinear = 32
Set to 32 following common GroupNorm practice; no sweep or justification is provided.
Input resolution = 1024x1024
Standard for remote sensing detection but chosen by hand; affects ERF estimates and stride effects.

assumptions (5)

ad hoc to paper A CNN's stride acts as a Nyquist sampling interval, and a target with minor axis smaller than the stride is 'irrevocably lost' (Section IV-A).
This equivalence between Shannon sampling and CNN feature grids is asserted, not derived. CNNs can detect objects smaller than the stride using context, as the paper's own baseline demonstrates.
domain assumption The Effective Receptive Field measured by gradient-magnitude sensitivity (Eq. 2) faithfully represents the information region actually used by the network (Section IV-B).
The paper relies on ERF estimates to argue that P4 already covers global context and P5 adds nothing, but ERF is a heuristic measure, not a proven bound on usable information.
domain assumption The P4 ERF of ~1024 px suffices for all ships, including large ones (Section IV-B).
Used to justify pruning P5; yet the paper's own Table IX shows Aircraft Carrier mAP drops from 71.2 to 66.9 when P5 is removed, partially contradicting this assumption.
domain assumption Training from scratch without augmentation for 100 epochs is a fair and sufficient protocol for comparing detectors (Section VI-B).
The authors deliberately disable augmentation to isolate architectural effects, but this makes the reported mAP values incomparable to published results and may disadvantage baselines that benefit more from augmentation.
domain assumption The 95% range of ship sizes in the four datasets captures the design-relevant distribution (Section III-B).
The choice of 95% rather than another quantile is not justified, and the min values (as low as 1 px) are ignored even though stride 4 also violates rho_minor>=1 for those instances.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LiM-YOLO: Less is More with Pyramid Level Shift for Ship Detection in Optical Remote Sensing." pith.science (2026). https://pith.science/paper/LAFNWRXR

@misc{pith2026251209700,
  author       = {Pith},
  title        = {Pith review of: LiM-YOLO: Less is More with Pyramid Level Shift for Ship Detection in Optical Remote Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LAFNWRXR}},
  note         = {Machine review of arXiv:2512.09700}
}

read the original abstract

General-purpose object detectors face fundamental structural limitations when applied to ship detection in satellite imagery, where the ship scale distribution is concentrated at small sizes and high aspect ratios. In conventional You Only Look Once architectures, the deepest feature pyramid level (stride 32) compresses narrow vessels into sub-pixel representations, causing severe spatial feature dilution and compromising accurate ship boundary regression. We propose Less is More YOLO, a streamlined detector built upon the extra-large variant of YOLOv9, to address these domain-specific structural conflicts. From a statistical analysis of ship scale distributions across four major benchmarks (SODA-A, DOTA-v1.5, FAIR1M-v2.0, and ShipRSImageNet), we introduce a Pyramid Level Shift Strategy that shifts the detection head from strides 8, 16, and 32 to strides 4, 8, and 16. This shift satisfies a spatial representability condition derived from the Nyquist-Shannon principle for the narrowest targets, while eliminating the computational redundancy of the deepest pyramid level. To further stabilize training on high-resolution satellite inputs, we incorporate a group-normalized composite-backbone projection module, mitigating gradient instability in memory-constrained micro-batch regimes. Validated on these four datasets, our detector attains an mAP50:95 of 0.600 with only 21.16 million parameters, a 64.1% reduction from the extra-large YOLOv9 baseline (58.99 million). Despite this compact size, our model surpasses state-of-the-art detectors up to three times larger, validating that a well-targeted pyramid level shift achieves a "Less is More" balance between accuracy and efficiency. The code is available at https://github.com/egshkim/LiM-YOLO.

Figures

Figures reproduced from arXiv: 2512.09700 by the authors.

**Figure 1.** Heatmap of Ship Major Axis Distribution risk of information loss during down-sampling. Although the original dataset includes multiple categories such as airplanes and vehicles, we extracted only ship instances for this study. The processed subset consists of 1,030 training images and 323 validation images, containing 37,971 and 21,908 OBBannotated ship objects, respectively. 2) DOTA-v1.5: DOTA-v1.5 [20] is an adva… view at source ↗

**Figure 3.** The overall architecture of the YOLOv9-E baseline. The detection head follows the conventional multi-scale configuration at pyramid levels P3, P4, and P5 (Strides 8, 16, 32) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Effective Recptive Field of YOLOv9e when the head level is P2/P3/P4/P5, with approximated diameter of 667.7, 860.5, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

Figures from the paper (4 more)

**Figure 5.** Figure 5: The overall architecture of the proposed LiM-YOLO. To address the scale mismatch in satellite imagery, we shift the detection pyramid levels from the conventional P3-P5 to P2-P4. The High-Resolution P2 Head (Stride 4) is introduced to recover fine-grained spatial detai…

**Figure 6.** Figure 6: Performance difference (∆) heatmaps comparing Model I–IV with different architectural configurations against the Baseline (YOLOv9-E). The panels correspond to: (Top-Left) Impact of adding the high-resolution P2 head (+P2H). (TopRight) The proposed LiM strategy pruning…

**Figure 7.** Figure 7: Class-wise performance comparison (∆) between the Baseline and the proposed LiM-YOLO (Model V). The heatmap displays differential values for Precision, Recall, F1- score, and mAP@50-95 across all ship categories sorted by size (small to large). Dominating red cells (∆ …

**Figure 8.** Figure 8: Qualitative comparison of detection results across four different remote sensing datasets. The rows correspond to SODA-A, DOTA-v1.5, FAIR1M-v2.0, and ShipRSImageNet datasets, from top to bottom. The columns represent the inference results of the Baseline (YOLOv9-E), Li…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 7 linked inside Pith

[1]

Advanced ship detection and ocean monitoring with satel- lite imagery and deep learning for marine science applications,

M. Bakirci, “Advanced ship detection and ocean monitoring with satel- lite imagery and deep learning for marine science applications,”Regional Studies in Marine Science, vol. 81, p. 103975, 2025

2025
[2]

Vessel detection leveraging satellite imagery and yolo in maritime surveillance,

R. Magalh ˜aes, A. P. Falc˜ao, and A. Barbosa, “Vessel detection leveraging satellite imagery and yolo in maritime surveillance,”Remote Sensing Applications: Society and Environment, p. 101730, 2025

2025
[3]

Vessel detection in satellite images using deep learning,

D. Sankhe and S. Bhosale, “Vessel detection in satellite images using deep learning,”Engineering, Technology & Applied Science Research, vol. 14, no. 6, pp. 18 357–18 362, 2024

2024
[4]

Ship detection with deep learning in optical remote-sensing images: A survey of challenges and advances,

T. Zhao, Y . Wang, Z. Li, Y . Gao, C. Chen, H. Feng, and Z. Zhao, “Ship detection with deep learning in optical remote-sensing images: A survey of challenges and advances,”Remote Sensing, vol. 16, no. 7, p. 1145, 2024

2024
[5]

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

J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779– 788

2016
[6]

Object detection in 20 years: A survey,

Z. Zou, K. Chen, Z. Shi, Y . Guo, and J. Ye, “Object detection in 20 years: A survey,”Proceedings of the IEEE, vol. 111, no. 3, pp. 257–276, 2023

2023
[7]

Yolov3: An incremental improvement,

J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” arXiv preprint arXiv:1804.02767, 2018

arXiv 2018
[8]

Yolov4: Op- timal speed and accuracy of object detection,

A. Bochkovskiy, C.-Y . Wang, and H.-Y . M. Liao, “Yolov4: Op- timal speed and accuracy of object detection,”arXiv preprint arXiv:2004.10934, 2020

arXiv 2004

Show all 36 references

[9]

Ultralytics yolov5,

G. Jocher, “Ultralytics yolov5,” 2020. [Online]. Available: https: //github.com/ultralytics/yolov5

2020
[10]

Yolov6: A single-stage object detection framework for industrial applications,

C. Li, L. Li, H. Jiang, K. Weng, Y . Geng, L. Li, Z. Ke, Q. Li, M. Cheng, W. Nieet al., “Yolov6: A single-stage object detection framework for industrial applications,”arXiv preprint arXiv:2209.02976, 2022

2022 arXiv
[11]

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

C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao, “Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 7464–7475

2023
[12]

Ultralytics yolov8,

G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics yolov8,” 2023. [Online]. Available: https://github.com/ultralytics/ultralytics

2023
[13]

Yolov9: Learning what you want to learn using programmable gradient information,

C.-Y . Wang, I.-H. Yeh, and H.-Y . Mark Liao, “Yolov9: Learning what you want to learn using programmable gradient information,” inEuro- pean conference on computer vision. Springer, 2024, pp. 1–21

2024
[14]

Yolov10: Real-time end-to-end object detection,

A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Hanet al., “Yolov10: Real-time end-to-end object detection,”Advances in Neural Information Processing Systems, vol. 37, pp. 107 984–108 011, 2024. JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 16

2024
[15]

Yolov11: An overview of the key architectural enhancements,

R. Khanam and M. Hussain, “Yolov11: An overview of the key architectural enhancements,”arXiv preprint arXiv:2410.17725, 2024

2024 arXiv
[16]

Yolov12: Attention-centric real-time object detectors,

Y . Tian, Q. Ye, and D. Doermann, “Yolov12: Attention-centric real-time object detectors,”arXiv preprint arXiv:2502.12524, 2025

2025 arXiv
[17]

Communication in the presence of noise,

C. E. Shannon, “Communication in the presence of noise,”Proceedings of the IRE, vol. 37, no. 1, pp. 10–21, 2006

2006
[18]

Group normalization,

Y . Wu and K. He, “Group normalization,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 3–19

2018
[19]

Soda: A low-power architecture for software radio,

Y . Lin, H. Lee, M. Woh, Y . Harel, S. Mahlke, T. Mudge, C. Chakrabarti, and K. Flautner, “Soda: A low-power architecture for software radio,” ACM SIGARCH Computer Architecture News, vol. 34, no. 2, pp. 89–101, 2006

2006
[20]

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

G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang, “Dota: A large-scale dataset for object detection in aerial images,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3974–3983

2018
[21]

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

X. Sun, P. Wang, Z. Yan, F. Xu, R. Wang, W. Diao, J. Chen, J. Li, Y . Feng, T. Xuet al., “Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 184, pp. 116– 130, 2022

2022
[22]

Shiprsimagenet: A large-scale fine-grained dataset for ship detection in high-resolution optical remote sensing images,

Z. Zhang, L. Zhang, Y . Wang, P. Feng, and R. He, “Shiprsimagenet: A large-scale fine-grained dataset for ship detection in high-resolution optical remote sensing images,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 14, pp. 8458– 8472, 2021

2021
[23]

Detrs beat yolos on real-time object detection,

Y . Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y . Liu, and J. Chen, “Detrs beat yolos on real-time object detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 16 965–16 974

2024
[24]

Yolox: Exceeding yolo series in 2021,

Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “Yolox: Exceeding yolo series in 2021,”arXiv preprint arXiv:2107.08430, 2021

2021 arXiv
[25]

Ultralytics yolo11,

G. Jocher and J. Qiu, “Ultralytics yolo11,” 2024. [Online]. Available: https://github.com/ultralytics/ultralytics

2024
[26]

Lmo-yolo: A ship detection model for low- resolution optical satellite imagery,

Q. Xu, Y . Li, and Z. Shi, “Lmo-yolo: A ship detection model for low- resolution optical satellite imagery,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 15, pp. 4117– 4131, 2022

2022
[27]

Yolov7-ship: A lightweight algorithm for ship object detection in complex marine environments,

Z. Jiang, L. Su, and Y . Sun, “Yolov7-ship: A lightweight algorithm for ship object detection in complex marine environments,”Journal of Marine Science and Engineering, vol. 12, no. 1, p. 190, 2024

2024
[28]

Cm-yolo: Typical object detection method in remote sensing cloud and mist scene images,

J. Hu, Y . Wei, W. Chen, X. Zhi, and W. Zhang, “Cm-yolo: Typical object detection method in remote sensing cloud and mist scene images,” Remote Sensing, vol. 17, no. 1, p. 125, 2025

2025
[29]

Shadowfpn-yolo: A real- time nms-free detector for remote sensing ship detection,

X. Yang, A. S. A. Mohamed, and C. Wang, “Shadowfpn-yolo: A real- time nms-free detector for remote sensing ship detection,”IEEE Access, 2025

2025
[30]

Omni-dimensional dynamic convolution,

C. Li, A. Zhou, and A. Yao, “Omni-dimensional dynamic convolution,” arXiv preprint arXiv:2209.07947, 2022

2022 arXiv
[31]

Carafe: Content-aware reassembly of features,

J. Wang, K. Chen, R. Xu, Z. Liu, C. C. Loy, and D. Lin, “Carafe: Content-aware reassembly of features,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 3007–3016

2019
[32]

Yolo-rsa: a multiscale ship detection algorithm based on optical remote sensing image,

Z. Fang, X. Wang, L. Zhang, and B. Jiang, “Yolo-rsa: a multiscale ship detection algorithm based on optical remote sensing image,”Journal of Marine Science and Engineering, vol. 12, no. 4, p. 603, 2024

2024
[33]

Yolo-ssboat: Super-small ship detection network for large-scale aerial and remote sensing scenes,

Y . Zeng, X. Wang, J. Zou, and H. Wu, “Yolo-ssboat: Super-small ship detection network for large-scale aerial and remote sensing scenes,” Remote Sensing, vol. 17, no. 11, p. 1948, 2025

1948
[34]

Internimage: Exploring large-scale vision foundation models with deformable convolutions,

W. Wang, J. Dai, Z. Chen, Z. Huang, Z. Li, X. Zhu, X. Hu, T. Lu, L. Lu, H. Liet al., “Internimage: Exploring large-scale vision foundation models with deformable convolutions,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 14 408–14 419

2023
[35]

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

Z. Liu, L. Yuan, L. Weng, and Y . Yang, “A high resolution optical satellite image dataset for ship recognition and some new baselines,” inInternational conference on pattern recognition applications and methods, vol. 2. SciTePress, 2017, pp. 324–331

2017
[36]

Understanding the effective receptive field in deep convolutional neural networks,

W. Luo, Y . Li, R. Urtasun, and R. Zemel, “Understanding the effective receptive field in deep convolutional neural networks,”Advances in neural information processing systems, vol. 29, 2016

2016

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Advanced ship detection and ocean monitoring with satel- lite imagery and deep learning for marine science applications,

M. Bakirci, “Advanced ship detection and ocean monitoring with satel- lite imagery and deep learning for marine science applications,”Regional Studies in Marine Science, vol. 81, p. 103975, 2025

2025

[2] [2]

Vessel detection leveraging satellite imagery and yolo in maritime surveillance,

R. Magalh ˜aes, A. P. Falc˜ao, and A. Barbosa, “Vessel detection leveraging satellite imagery and yolo in maritime surveillance,”Remote Sensing Applications: Society and Environment, p. 101730, 2025

2025

[3] [3]

Vessel detection in satellite images using deep learning,

D. Sankhe and S. Bhosale, “Vessel detection in satellite images using deep learning,”Engineering, Technology & Applied Science Research, vol. 14, no. 6, pp. 18 357–18 362, 2024

2024

[4] [4]

Ship detection with deep learning in optical remote-sensing images: A survey of challenges and advances,

T. Zhao, Y . Wang, Z. Li, Y . Gao, C. Chen, H. Feng, and Z. Zhao, “Ship detection with deep learning in optical remote-sensing images: A survey of challenges and advances,”Remote Sensing, vol. 16, no. 7, p. 1145, 2024

2024

[5] [5]

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

J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779– 788

2016

[6] [6]

Object detection in 20 years: A survey,

Z. Zou, K. Chen, Z. Shi, Y . Guo, and J. Ye, “Object detection in 20 years: A survey,”Proceedings of the IEEE, vol. 111, no. 3, pp. 257–276, 2023

2023

[7] [7]

Yolov3: An incremental improvement,

J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” arXiv preprint arXiv:1804.02767, 2018

arXiv 2018

[8] [8]

Yolov4: Op- timal speed and accuracy of object detection,

A. Bochkovskiy, C.-Y . Wang, and H.-Y . M. Liao, “Yolov4: Op- timal speed and accuracy of object detection,”arXiv preprint arXiv:2004.10934, 2020

arXiv 2004