YOLO-SPCI: Enhancing Remote Sensing Object Detection via Selective-Perspective-Class Integration

KinTak U; Lian Peng; Xiangcheng Li; Xinyuan Wang; Yilin He

REVIEW 4 major objections 4 minor 43 references

YOLO-SPCI: Enhancing Remote Sensing Object Detection via Selective-Perspective-Class Integration

T0 review · 4 major / 4 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 YOLO-SPCI inserts a three-part attention module into YOLOv8's backbone and reports 92.0% mAP50 on NWPU VHR-10, up from 88.9%.

desk verdict A plausible lightweight attention add-on for YOLOv8 with a sensible architecture and decent ablations, but the headline +3.1 mAP gain rests on single-run, split-unspecified experiments, so treat it as promising but unproven. read the letter →

arxiv 2505.21370 v1 pith:IM3WEBK2 submitted 2025-05-27 cs.CV

Xinyuan Wang , Lian Peng , Xiangcheng Li , Yilin He , KinTak U This is my paper

classification cs.CV

keywords remotesensingobjectdetectionYOLOv8attentionmechanismmulti-scalefeaturefusionSelective-Perspective-ClassIntegrationNWPUVHR-10DIORlightweightdetector

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

YOLO-SPCI is an object detector for remote sensing imagery built by inserting a lightweight attention module, SPCI, into two stages of the YOLOv8 backbone while leaving the neck and head unchanged. The paper's central claim is that this insertion improves multi-scale feature representation enough to raise mean average precision at IoU 0.5 on NWPU VHR-10 from 88.9% to 92.0%, and on DIOR from 80.5% to 81.2%. A sympathetic reader would care because the gain comes from a small module adding modest parameters and compute rather than from a larger detector, suggesting that cheap architectural additions can help on dense, cluttered aerial imagery. The claim rests on ablations attributing the gain to three submodules: global stream gating, spatial context fusion, and class-aware reweighting, with the spatial fusion module the largest single contributor.

What carries the argument

The carrying mechanism is the Selective-Perspective-Class Integration module, a residual-style attention block made of three submodules. The Selective Stream Gate pools the whole feature map and uses two 1x1 convolutions to produce channel weights; the Perspective Fusion Module concatenates global average and max pooling and passes them through a 7x7 convolution to produce spatial context weights; and the Class Discrimination Module applies 1x1 and 3x3 convolutions with batch normalization to generate class-aware channel weights. The three outputs are added element-wise and passed through dropout before the block output, preserving a direct residual path. The argument depends on the claim that these three complementary refinements, inserted after the P3 and P5 backbone stages, improve both small-object localization and large-object recognition without altering the neck or head.

What would settle it

Re-running the same training protocol several times with different random seeds on the same data splits and checking whether YOLO-SPCI consistently beats YOLOv8n by the reported margins would settle the claim; if run-to-run variation of a few mAP points overlaps the baseline, the central improvement claim fails.

Watch

Extended reading notes

Core claim

The discovery the paper asserts is that a single attention unit combining global channel gating, average- and max-pool spatial context fusion, and class-aware channel reweighting can improve a YOLOv8 backbone's remote sensing detection when placed at both a shallow high-resolution stage (P3) and a deep semantic stage (P5). Specifically, YOLO-SPCI reaches 92.0% mAP50 on NWPU VHR-10, beating the unmodified YOLOv8n baseline's 88.9% and also the listed YOLOv5x and PR-Deformable DETR results, while staying lightweight at 3.1M parameters and 8.3 GFLOPs. On DIOR it reports 81.2% mAP50 and 60.9% mAP50-95, above the baseline's 80.5% and 59.9%. The authors further claim that per-class gains concentrate in small, dense, and structurally regular categories such as storage tanks, baseball diamonds, basketball courts, and bridges, and that dual placement at P3 and P5 outperforms either placement alone.

Load-bearing premise

The load-bearing premise is that the reported mAP gaps of 3.1 points on NWPU VHR-10 and 0.7 points on DIOR are real and reproducible, even though each configuration was trained once with no repeated seeds or error bars and the train/validation splits and augmentation details are not given.

Editorial extensions

If this is right

If the central claim holds, inserting SPCI at P3 and P5 gives a cheap accuracy lift over the YOLOv8n baseline without touching the neck or head.
The ablation implies that the Perspective Fusion Module is the critical subcomponent: removing it drops mAP50 back to the baseline level of 88.9%, so spatial context fusion carries most of the gain.
Removing the Selective Stream Gate drops mAP50 to 90.4% and removing the Class Discrimination Module drops it to 91.2%, indicating all three components contribute but not equally.
The reported per-class results predict particular benefit for compact, densely packed, and geometrically structured classes rather than a uniform gain across all ten categories.
Because the module keeps total parameters at 3.1M and compute at 8.3 GFLOPs, the same architecture could be used in settings where computation is limited.

Reading between the lines

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

The paper reports single training runs without error bars, so a natural extension would be to re-evaluate YOLO-SPCI against YOLOv8n with multiple random seeds and report mean and standard deviation, especially since the DIOR gap is only 0.7 points.
The module's components resemble existing channel and spatial attention mechanisms, so an informative next test would be an apples-to-apples comparison under identical training settings against SE, CBAM, and other attention blocks inserted at the same positions.
If the gain transfers, the design could be dropped into other YOLO versions or transformer detectors, though the paper only tests P3 and P5 insertion in YOLOv8, leaving that generalization unverified.
The claimed class-discrimination benefit could be probed directly by measuring per-class feature separability or confusion-matrix errors before and after the Class Discrimination Module, rather than relying on overall mAP alone.

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

Quick take: this is a competent, incremental attention-module paper. The SPCI block is a combination of existing ideas — SSG is basically SE channel gating, PFM is CBAM-style spatial attention, CDM is a small conv-based channel reweighting. The new bit is the specific combination plus insertion at P3 and P5 in a YOLOv8 backbone, and the authors show in ablation that each submodule contributes something and the dual placement beats either placement alone. The writing is clear, the equations match the described architecture, and the cost is attractive (3.1M params, 8.3 GFLOPs). Credit where due: they compare against a decent set of recent baselines, and the per-class numbers in Table 1 are consistent with a real improvement rather than a typo.

The soft spots are real, though. The main one is experimental hygiene. The +3.1 mAP50 gain over YOLOv8n on NWPU VHR-10, and the smaller DIOR gain, come from single training runs with no seeds, no error bars, and no stated train/validation split or augmentation policy. NWPU VHR-10 has only 650 positive images; published numbers on this benchmark move by several points with split choice, and run-to-run variance for a YOLOv8n-sized model is usually around 0.5–1.0 mAP. So the headline claim is not yet supported. The DIOR number (0.7 points) is inside the noise band. This is fixable — repeated seeds on a fixed split, plus code release — but it's the load-bearing issue.

Second, there's a contradiction in Section 4.1: DIOR is described as a transfer test 'without any retraining or parameter adjustment,' but Section 4.2 says all experiments use the same training configuration, and the reported DIOR numbers clearly require training on DIOR. That needs a correction; it's secondary but sloppy.

Third, the 'class discrimination' claim for CDM is overstated. No class labels enter the module; it's a channel reweighting block with a 3x3 conv. Calling it class-aware is a stretch. That doesn't kill the paper, but the framing should be toned down.

The central architecture is coherent, and the ablations are informative. If the results hold up under repeated trials, this is a useful small improvement for remote-sensing detection. As it stands, I'd want code and seed-resolved numbers before trusting the specific gain. The citation pattern is fine — they cite SE, CBAM, and the YOLO line — and there's no self-citation inflation.

Bottom line: worth a serious referee, but only with the expectation of a revision that adds seeds, split details, and a code link. I wouldn't base any of my own conclusions on the +3.1 yet.

Referee Report

4 major / 4 minor

Summary. This manuscript proposes YOLO-SPCI, an attention-based enhancement to the YOLOv8 backbone for remote sensing object detection. The SPCI module combines a Selective Stream Gate (SSG), a Perspective Fusion Module (PFM), and a Class Discrimination Module (CDM), inserted after the P3 and P5 stages. The authors report mAP50 of 92.0% on NWPU VHR-10, beating YOLOv8n (88.9%) by 3.1 points and several prior detectors, plus mAP50 of 81.2% on DIOR (baseline 80.5%). Ablations in Table 3 attribute improvements to all three components and to dual placement.

Significance. If the reported gains are reproducible, the contribution is a lightweight and architecturally compatible attention block with measurable improvements on a challenging remote sensing benchmark. Strengths include a clear module formulation (Eqs. 1-8), a consistent architecture description, and a fairly extensive comparison table and ablation. However, the absence of multiple seeds, error bars, and split details means the headline claims are not yet statistically supported. The paper would be strengthened by releasing code and by reporting means and variances over repeated runs.

major comments (4)

[§4.2, §4.3, Tables 1–3] All reported results come from a single training run per configuration. No random seed, no error bar, and no train/validation split is specified for either dataset. For NWPU VHR-10 (650 positive images), published mAP values depend strongly on the split, and run-to-run variance of YOLOv8-sized detectors is typically on the order of 0.5–1.0 mAP. Thus the 3.1-point advantage over the baseline in Table 1 and the 0.7-point advantage in Table 2 could be within noise or an artifact of an unstated split. The paper should fix the split, provide standard augmentation details, and report mean ± std over at least 3–5 seeds for the baseline, the full model, and the key ablations.
[§4.1 and §4.3] These two sections contradict each other. Section 4.1 states that DIOR is used 'without any retraining or parameter adjustment,' i.e., as a transfer test, but Section 4.3 describes the DIOR results as outperforming models such as YOLOv8n and variants, which are trained on DIOR. Please clarify whether the DIOR numbers in Table 2 come from a model trained on DIOR or from the NWPU-trained model transferred to DIOR; if the latter, the comparison in Table 2 is not apples-to-apples and should be redone against transfer baselines.
[§3.2.3, Eqs. (5)–(8), Table 3] The CDM is called 'class-aware' and is claimed to enhance inter-class separability, but it uses no class labels or class-conditional information; it is a small convolutional block applied uniformly to the feature map. The ablation (B4 vs B7 in Table 3) only reports mAP, which does not demonstrate improved class discrimination. Please either provide direct evidence of class-level separation (e.g., per-class precision changes, confusion matrices, or feature statistics) or rephrase the claims to say that CDM performs local nonlinear refinement rather than class-aware attention.
[§4.4, Fig. 2] The placement of the SPCI module at P3 and P5 is justified post hoc by the same ablation that evaluates it (B5–B7), and only two single-stage placements plus the combined placement are considered. This is a mild selection-on-validation issue, and with single runs there is no statistical basis for claiming that P3+P5 is 'optimal.' In addition, the caption of Fig. 2 refers to variants B1–B8 while Table 3 defines only B1–B7. Please reconcile the numbering and, if possible, add a random-placement or different-stage control to support the placement claim.

minor comments (4)

[Table 1] Some rows have per-class values concatenated without delimiters (e.g., '100.090.1' for PR-Deformable DETR and '99.190.7' for YOLOv8n), making the table hard to read.
[Eqs. (1) and (3)] The notation Conv1 is used in both equations, but in Eq. (1) it denotes a 1×1 convolution and in Eq. (3) a 7×7 convolution; this overloading should be resolved with distinct subscripts.
[§4.2] The sentence 'All subsequent experiments, including those on the DIOR dataset, are conducted using this model as the baseline, with no further changes to the architecture, training configuration, or evaluation settings' is ambiguous about whether DIOR results come from retraining on DIOR or from the NWPU-trained model; this ambiguity should be removed.
[§4.4, Table 3] For B3 (Disable PFM), mAP50 is identical to the baseline (88.9%) while mAP50-95 drops to 55.5%. The text says PFM removal causes 'the most significant performance degradation,' but does not discuss why the primary metric is unchanged; this asymmetry should be explained.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the headline mAP result is an empirical measurement, and the SPCI equations define a learned transformation rather than a fitted restatement of the outcome.

full rationale

YOLO-SPCI is an empirical architecture paper. The SPCI module (SSG, PFM, CDM) is defined by standard attention equations (Eqs. 1-8), and the headline claim (92.0 vs 88.9 mAP50 on NWPU VHR-10, Table 1) is a measured training/evaluation outcome, not a quantity derived from a fitted parameter or from an assumed result. No parameter is fitted to a subset and then "predicted" on a closely related quantity. The module placement (P3/P5) is validated post hoc by ablation (B5-B7), which is standard empirical model selection rather than a self-justifying loop. The paper invokes no uniqueness theorem from the authors' prior work, and the cited related work is not load-bearing for the claimed contribution. The DIOR transferability statement in Section 4.1 ("without any retraining or parameter adjustment") is inconsistent with reporting DIOR scores in Table 2, and the experiments are single-run with no seed, split, or augmentation details; these are correctness and reproducibility risks, but they are not circularity. Accordingly, no step in the paper's derivation or evaluation chain reduces to its own inputs by construction.

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

No new physical entities are introduced. The paper proposes a software module and makes an unsupported functional claim about class discrimination, which is captured in the axioms and red flags.

free parameters (5)

SSG bottleneck ratio = c_in/16
The hidden width of the channel gate is set to 1/16 of input channels without a sensitivity study.
CDM hidden channel scale = c_out/4
The intermediate width for the class discrimination block is chosen as 1/4 of output channels, with no ablation on this choice.
PFM kernel size = 7x7
Taken from the CBAM spatial-attention design, not tuned for remote sensing inputs.
SPCI insertion stages = P3 and P5
Ablation (B5-B7) shows that both are better than either alone, but no search over other stage combinations is reported.
Dropout rate = not stated
The dropout layer shown in Figure 1 is never assigned a probability in the text.

assumptions (4)

standard math Convolution, pooling, and batch normalization behave as implemented in PyTorch.
The modules rely on standard deep learning primitives.
domain assumption NWPU VHR-10 and DIOR results generalize to real remote sensing deployment.
The paper argues for practical applicability but only tests on two public benchmarks.
domain assumption P3 and P5 in YOLOv8 correspond to small-object and large-object feature levels.
Section 3.1 motivates the placement using this widely used but heuristic claim.
ad hoc to paper A class-aware attention weight can be learned from the feature map alone without class labels.
Section 3.2.3 defines CDM as class-aware but the module never sees object class labels or a classification supervision targeting feature distinctiveness.

how reviews work

0 comments

Cite this review

Pith. "Pith review of YOLO-SPCI: Enhancing Remote Sensing Object Detection via Selective-Perspective-Class Integration." pith.science (2026). https://pith.science/paper/IM3WEBK2

@misc{pith2026250521370,
  author       = {Pith},
  title        = {Pith review of: YOLO-SPCI: Enhancing Remote Sensing Object Detection via Selective-Perspective-Class Integration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IM3WEBK2}},
  note         = {Machine review of arXiv:2505.21370}
}

read the original abstract

Object detection in remote sensing imagery remains a challenging task due to extreme scale variation, dense object distributions, and cluttered backgrounds. While recent detectors such as YOLOv8 have shown promising results, their backbone architectures lack explicit mechanisms to guide multi-scale feature refinement, limiting performance on high-resolution aerial data. In this work, we propose YOLO-SPCI, an attention-enhanced detection framework that introduces a lightweight Selective-Perspective-Class Integration (SPCI) module to improve feature representation. The SPCI module integrates three components: a Selective Stream Gate (SSG) for adaptive regulation of global feature flow, a Perspective Fusion Module (PFM) for context-aware multi-scale integration, and a Class Discrimination Module (CDM) to enhance inter-class separability. We embed two SPCI blocks into the P3 and P5 stages of the YOLOv8 backbone, enabling effective refinement while preserving compatibility with the original neck and head. Experiments on the NWPU VHR-10 dataset demonstrate that YOLO-SPCI achieves superior performance compared to state-of-the-art detectors.

Figures

Figures reproduced from arXiv: 2505.21370 by the authors.

**Figure 2.** Quantitative comparison of mAP50 and mAP50-95 across all ablation variants [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗

**Figure 3.** Comparison of detection results between the baseline model and the proposed [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 33 canonical work pages

[1]

Military target detection method based on efficientdet and transfer learning,

L. Li et al. , “Military target detection method based on efficientdet and transfer learning,” Engineering Applications of Artificial Intelligence , vol. 126, p. 107413, 2024

work page 2024
[2]

An online continual object detector on vhr remote sensing images via replay-based approach,

X. Zhang et al. , “An online continual object detector on vhr remote sensing images via replay-based approach,” Engineering Applications of Artificial Intelligence , vol. 114, p. 105091, 2022

work page 2022
[3]

Enhancing aerial object detection with se- lective frequency interaction network,

W. Weng, M. Wei, J. Ren, and F. Shen, “Enhancing aerial object detection with se- lective frequency interaction network,” IEEE Transactions on Artificial Intelligence , vol. 1, no. 01, pp. 1–12, 2024

work page 2024
[4]

A novel multi- frequency coordinated module for sar ship detection,

C. Qiao, F. Shen, X. Wang, R. Wang, F. Cao, S. Zhao, and C. Li, “A novel multi- frequency coordinated module for sar ship detection,” in 2022 IEEE 34th Interna- tional Conference on Tools with Artificial Intelligence (ICTAI) . IEEE, 2022, pp. 804–811

work page 2022
[5]

Enhanced few-shot object detection for remote sensing images via pro- gressive integration learning,

P. Li et al. , “Enhanced few-shot object detection for remote sensing images via pro- gressive integration learning,” Engineering Applications of Artificial Intelligence , vol. 126, p. 107435, 2024

work page 2024
[6]

High-resolution network for static infrared weak and small targets detection,

H. Yue et al. , “High-resolution network for static infrared weak and small targets detection,” Engineering Applications of Artificial Intelligence , vol. 113, p. 105040, 2024

work page 2024
[7]

Unmanned aerial vehicles advances in object detection and tracking for environmental monitoring applications,

M. A. Al-Garadi et al. , “Unmanned aerial vehicles advances in object detection and tracking for environmental monitoring applications,” Engineering Applications of Ar- tificial Intelligence , vol. 126, p. 107385, 2024

work page 2024
[8]

Scfnet: Semantic correction and focus network for remote sensing object detection,

J. Li, W. Wang, T. Zhang, and S. Yang, “Scfnet: Semantic correction and focus network for remote sensing object detection,” Engineering Applications of Artificial Intelligence, vol. 120, p. 105970, 2023

work page 2023

Show all 43 references

[9]

Yolov5-ship: Improved yolov5 for ship detection in remote sensing images,

M. Liu, J. Li, K. Wang et al. , “Yolov5-ship: Improved yolov5 for ship detection in remote sensing images,” Engineering Applications of Artificial Intelligence , vol. 126, p. 107423, 2024

2024
[10]

Ultralytics YOLOv8: Cutting- Edge Object Detection Models,

G. Jocher, A. Chaurasia, J. Qiu, and T. Stoken, “Ultralytics YOLOv8: Cutting- Edge Object Detection Models,” https://github.com/ultralytics/ultralytics, 2023, ac- cessed: 2025-03-29. 18

2023
[11]

Deformable DETR: Deformable transformers for end-to-end object detection,

X. Zhu, W. Su, L. Lu, B. Xu, X. Li, X. Wang, and J. Dai, “Deformable DETR: Deformable transformers for end-to-end object detection,” arXiv preprint arXiv:2010.04159, 2020

2010 arXiv
[12]

Pr-deformable detr: Detr for remote sensing object detection,

Y. Chen, B. Liu, and L. Yuan, “Pr-deformable detr: Detr for remote sensing object detection,” IEEE Geoscience and Remote Sensing Letters , vol. 21, pp. 1–5, 2024

2024
[13]

Squeeze-and-excitation networks,

J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) , 2018

2018
[14]

Cbam: Convolutional block attention module,

S. Woo, J. Park, J.-Y. Lee, and I. S. Kweon, “Cbam: Convolutional block attention module,” Proceedings of the European Conference on Computer Vision (ECCV) , pp. 3–19, 2018

2018
[15]

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

S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” Advances in Neural Information Processing Systems (NIPS) , 2015

2015
[16]

Ssd: Single shot multibox detector,

W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C. Berg, “Ssd: Single shot multibox detector,” in European Conference on Computer Vision (ECCV), 2016

2016
[17]

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,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

2016
[18]

Yolov3: An incremental improvement,

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

2018 arXiv
[19]

Yolov4: Optimal speed and accu- racy of object detection,

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

2004 arXiv
[20]

Cspnet: A new backbone that can enhance learning capability of cnn,

C.-Y. Wang, H.-Y. M. Liao, Y.-H. Wu, P.-Y. Chen, J.-W. Hsieh, and I.-H. Yeh, “Cspnet: A new backbone that can enhance learning capability of cnn,”Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPR W), 2020

2020
[21]

Yolov5 github repository,

G. Jocher, A. Chaurasia, J. Qiu, and T. Stoken, “Yolov5 github repository,” 2020. [Online]. Available: https://github.com/ultralytics/yolov5

2020
[22]

Imaggarment-1: Fine-grained garment generation for controllable fashion design,

F. Shen, J. Yu, C. Wang, X. Jiang, X. Du, and J. Tang, “Imaggarment-1: Fine-grained garment generation for controllable fashion design,”arXiv preprint arXiv:2504.13176, 2025

2025 arXiv
[23]

Long-term talkingface generation via motion-prior conditional diffusion model,

F. Shen, C. Wang, J. Gao, Q. Guo, J. Dang, J. Tang, and T.-S. Chua, “Long-term talkingface generation via motion-prior conditional diffusion model,” arXiv preprint arXiv:2502.09533, 2025. 19

2025 arXiv
[24]

Feature pyramid networks for object detection,

T.-Y. Lin, P. Doll´ ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” Proceedings of the IEEE conference on com- puter vision and pattern recognition , pp. 2117–2125, 2017

2017
[25]

Deep high-resolution representation learning for visual recog- nition,

J. Wang, K. Sun, T. Cheng, B. Jiang, C. Deng, Y. Zhao, D. Liu, Y. Mu, M. Tan, X. Wang, and W. Liu, “Deep high-resolution representation learning for visual recog- nition,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 43, no. 10, pp. 3349–3364, 2020

2020
[26]

Mobilenets: Efficient convolutional neural networks for mo- bile vision applications,

A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. An- dreetto, and H. Adam, “Mobilenets: Efficient convolutional neural networks for mo- bile vision applications,” arXiv preprint arXiv:1704.04861 , 2017

2017 arXiv
[27]

Imagpose: A unified conditional framework for pose-guided person generation,

F. Shen and J. Tang, “Imagpose: A unified conditional framework for pose-guided person generation,” Advances in neural information processing systems , vol. 37, pp. 6246–6266, 2024

2024
[28]

Imagdressing- v1: Customizable virtual dressing,

F. Shen, X. Jiang, X. He, H. Ye, C. Wang, X. Du, Z. Li, and J. Tang, “Imagdressing- v1: Customizable virtual dressing,” in Proceedings of the AAAI Conference on Arti- ficial Intelligence , vol. 39, no. 7, 2025, pp. 6795–6804

2025
[29]

End- to-end object detection with transformers,

N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End- to-end object detection with transformers,” Proceedings of the European Conference on Computer Vision (ECCV) , pp. 213–229, 2020

2020
[30]

Multi-class geospatial object detection and geographic image classification based on collection of part detectors,

G. Cheng, J. Han, P. Zhou, and L. Guo, “Multi-class geospatial object detection and geographic image classification based on collection of part detectors,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 98, pp. 119–132, 2014

2014
[31]

Object detection in optical remote sensing images: A survey and a new benchmark,

K. Li, X. Wu, G.-S. Peng, X. Chen, and K. Zhao, “Object detection in optical remote sensing images: A survey and a new benchmark,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 149, pp. 16–34, 2019

2019
[32]

Dynamic pseudo-label generation for weakly supervised object detection in remote sensing images,

H. Wang, H. Li, W. Qian, W. Diao, L. Zhao, J. Zhang, and D. Zhang, “Dynamic pseudo-label generation for weakly supervised object detection in remote sensing images,” Remote Sensing, vol. 13, no. 8, p. 1461, 2021

2021
[33]

Deformable convnet with aspect ratio constrained nms for object detection in remote sensing imagery,

Z. Xu, X. Xu, L. Wang, R. Yang, and F. Pu, “Deformable convnet with aspect ratio constrained nms for object detection in remote sensing imagery,” Remote Sens., vol. 9, no. 12, 2017

2017
[34]

Deformable detr: Deformable transformers for end-to-end object detection,

X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to-end object detection,” in Proc. Int. Conf. Learn. Represent., 2021

2021
[35]

Rotated feature network for multiorientation object detection of remote-sensing images,

K. Zhou, Z. Zhang, C. Gao, and J. Liu, “Rotated feature network for multiorientation object detection of remote-sensing images,” in IEEE Geosci. Remote Sens. Lett. , vol. 18, no. 1, 2021, pp. 33–37. 20

2021
[36]

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

S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 39, no. 6, pp. 1137–1149, 2017

2017
[37]

Sig-nms-based faster r-cnn combining transfer learning for small target detection in vhr optical remote sensing imagery,

R. Dong, D. Xu, J. Zhao, L. Jiao, and J. An, “Sig-nms-based faster r-cnn combining transfer learning for small target detection in vhr optical remote sensing imagery,” IEEE Trans. Geosci. Remote Sens. , vol. 57, no. 11, pp. 8534–8545, 2019

2019
[38]

Small object intelligent detection method based on adaptive recursive feature pyramid,

J. Zhang, Y. Liu, J. Zhang, X. Lin, and Y. Fan, “Small object intelligent detection method based on adaptive recursive feature pyramid,” Sensors, vol. 21, no. 18, p. 6139, 2021

2021
[39]

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,” arXiv preprint arXiv:2207.02696, 2022

2022 arXiv
[40]

Focal loss for dense object detection,

T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Doll´ ar, “Focal loss for dense object detection,” IEEE transactions on pattern analysis and machine intelligence , vol. 42, no. 2, pp. 318–327, 2020

2020
[41]

Objects as points,

X. Zhou, D. Wang, and P. Kr¨ ahenb¨ uhl, “Objects as points,” arXiv preprint arXiv:1904.07850, 2019

1904 arXiv
[42]

A convolution with transformer attention module integrating local and global features for remote sensing object detection,

L. Xiao, Q. Zhao, and H. Liu, “A convolution with transformer attention module integrating local and global features for remote sensing object detection,” Remote Sensing, vol. 16, no. 5, p. 906, 2024

2024
[43]

Rs-featfusenet: An integrated remote sensing object detection network based on enhanced spatial hierarchical attention,

Y. Wang, J. Liu, and P. Zhang, “Rs-featfusenet: An integrated remote sensing object detection network based on enhanced spatial hierarchical attention,”Remote Sensing, vol. 17, no. 1, p. 61, 2024. 21

2024

Pith tools

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

[1] [1]

Military target detection method based on efficientdet and transfer learning,

L. Li et al. , “Military target detection method based on efficientdet and transfer learning,” Engineering Applications of Artificial Intelligence , vol. 126, p. 107413, 2024

work page 2024

[2] [2]

An online continual object detector on vhr remote sensing images via replay-based approach,

X. Zhang et al. , “An online continual object detector on vhr remote sensing images via replay-based approach,” Engineering Applications of Artificial Intelligence , vol. 114, p. 105091, 2022

work page 2022

[3] [3]

Enhancing aerial object detection with se- lective frequency interaction network,

W. Weng, M. Wei, J. Ren, and F. Shen, “Enhancing aerial object detection with se- lective frequency interaction network,” IEEE Transactions on Artificial Intelligence , vol. 1, no. 01, pp. 1–12, 2024

work page 2024

[4] [4]

A novel multi- frequency coordinated module for sar ship detection,

C. Qiao, F. Shen, X. Wang, R. Wang, F. Cao, S. Zhao, and C. Li, “A novel multi- frequency coordinated module for sar ship detection,” in 2022 IEEE 34th Interna- tional Conference on Tools with Artificial Intelligence (ICTAI) . IEEE, 2022, pp. 804–811

work page 2022

[5] [5]

Enhanced few-shot object detection for remote sensing images via pro- gressive integration learning,

P. Li et al. , “Enhanced few-shot object detection for remote sensing images via pro- gressive integration learning,” Engineering Applications of Artificial Intelligence , vol. 126, p. 107435, 2024

work page 2024

[6] [6]

High-resolution network for static infrared weak and small targets detection,

H. Yue et al. , “High-resolution network for static infrared weak and small targets detection,” Engineering Applications of Artificial Intelligence , vol. 113, p. 105040, 2024

work page 2024

[7] [7]

Unmanned aerial vehicles advances in object detection and tracking for environmental monitoring applications,

M. A. Al-Garadi et al. , “Unmanned aerial vehicles advances in object detection and tracking for environmental monitoring applications,” Engineering Applications of Ar- tificial Intelligence , vol. 126, p. 107385, 2024

work page 2024

[8] [8]

Scfnet: Semantic correction and focus network for remote sensing object detection,

J. Li, W. Wang, T. Zhang, and S. Yang, “Scfnet: Semantic correction and focus network for remote sensing object detection,” Engineering Applications of Artificial Intelligence, vol. 120, p. 105970, 2023

work page 2023