Pith. sign in

REVIEW 2 major objections 5 minor 32 references

SAM3 pseudo-labels can train a compact UAV segmenter that keeps most of the foundation model's accuracy at a small fraction of its size and compute.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 00:10 UTC pith:UXU4TCA6

load-bearing objection Solid one-stage SAM3-to-lightweight transfer for UAV targets; the two-stage "fine-grained" claim is under-supported by the paper's own numbers. the 2 major comments →

arxiv 2607.03754 v1 pith:UXU4TCA6 submitted 2026-07-04 cs.CV

Exploring SAM Supervision for Fine-Grained UAV Target Segmentation under Data Scarcity

Le-Anh Tran This is my paper
classification cs.CV
keywords UAV target segmentationSAM3pseudo-label supervisionlightweight segmentationIdentityFormerdata scarcityfine-grained masksfoundation model annotation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Segmenting drones from aerial video is hard: the targets are tiny, their looks change with distance and angle, backgrounds are cluttered, and dense pixel labels are scarce. This paper shows that masks produced by the large Segment Anything Model 3 can replace those missing labels when training a small network called IPS-Seg. Two uses of SAM3 are compared: taking its full-image masks as-is, and a two-stage process that first localizes the object then re-segments a tight crop so boundaries become sharper and false positives can be dropped. IPS-Seg combines a lightweight IdentityFormer backbone, multi-scale context pooling, and PixelShuffle upsampling. On a public UAV dataset the student retains most of SAM3's zero-shot scores while using only a few million parameters, arguing that foundation models can serve as automatic annotators for deployable networks when hand labels are unavailable.

Core claim

A lightweight network (IPS-Seg) trained only on SAM3-generated pseudo-labels reaches IoU 0.7941 and Dice 0.8737, close to SAM3's own zero-shot scores of 0.8201 / 0.8991, while needing just 2.69 M parameters and 9.72 GFLOPs. A two-stage refinement that re-applies SAM3 to object-centric crops recovers fine structures such as propellers and landing gear; the paper attributes the accompanying drop in region-overlap scores to the coarseness of the dataset's ground-truth masks rather than to worse supervision.

What carries the argument

The two-stage SAM3-guided pseudo-label pipeline: full-image SAM3 yields coarse masks for localization; candidate patches are cropped and re-segmented by SAM3 (using its presence head to discard false positives), producing fine masks that, together with the coarse ones, supervise IPS-Seg.

Load-bearing premise

The claim that two-stage masks are better supervision rests on the premise that their lower IoU and Dice scores are caused by coarse ground-truth annotations, not by genuinely worse masks.

What would settle it

Re-annotate a held-out validation subset with fine propeller- and landing-gear-level boundaries, then recompute IoU and Dice; if the two-stage models still score lower than the one-stage models, the refinement-superiority claim fails.

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

If this is right

  • Compact UAV segmentation models can be trained without dense manual pixel labels.
  • Foundation-model pseudo-labels can transfer open-vocabulary localization into networks small enough for onboard deployment.
  • Patch-wise re-segmentation recovers geometric detail that full-image prompting misses.
  • Conventional region-overlap metrics can undervalue fine-boundary gains when ground-truth masks themselves are coarse.
  • The same teacher-student pattern can be reused for other low-label dense-prediction tasks.

Where Pith is reading between the lines

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

  • The presence-head verification step may matter more than boundary refinement for cutting false positives on small aerial objects.
  • Boundary-aware metrics would be required before the two-stage strategy can be ranked above one-stage on leaderboards.
  • The pipeline could transfer to other small-object aerial tasks (birds, ground vehicles) where foundation models already localize coarsely.
  • Filtering SAM3 masks by confidence before student training could close the remaining teacher-student accuracy gap.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes a SAM3-guided pseudo-label framework for training a lightweight UAV-target segmentation network (IPS-Seg) under annotation scarcity. Two paradigms are studied: direct one-stage use of SAM3 masks, and a two-stage pipeline that first localizes candidates then re-segments cropped patches for finer boundaries and false-positive rejection. IPS-Seg combines an IdentityFormer backbone, ASPP bottleneck and PixelShuffle decoder. Fully-supervised results on a subsampled UAV Semantic Segmentation dataset show competitive IoU/Dice at low parameter/FLOP cost; one-stage pseudo-supervision transfers most of SAM3’s zero-shot performance to the student; two-stage results are lower on region-overlap metrics but are argued (via qualitative examples) to recover finer structures such as propellers and landing gear.

Significance. If the claims hold, the work supplies a practical recipe for distilling open-vocabulary foundation-model masks into compact, deployable UAV segmentors, addressing a real annotation bottleneck. Strengths include clean architectural ablations (Tables 3–4) that isolate backbone/bottleneck/decoder contributions, an explicit efficiency–accuracy Pareto plot (Fig. 1), public code, and a transparent teacher–student transfer experiment that nearly matches SAM3 zero-shot IoU (0.7941 vs 0.8201) with only 2.69 M parameters. These elements make the one-stage result immediately usable for resource-constrained platforms.

major comments (2)
  1. Section 4.2.2 and Table 2: the two-stage strategy is presented as the key contribution for “fine-grained” supervision, yet SAM3-2S and IPS-Seg-2S drop sharply in IoU/Dice (0.6656/0.7945 and 0.6002/0.7291) relative to the one-stage numbers. The manuscript attributes the drop solely to coarse ground-truth annotations that penalize recovered propellers/landing gear. No quantitative evidence is supplied that the GT masks are systematically coarser, that the added structures are correct rather than hallucinated, or that boundary-aware metrics (e.g., boundary F-score, Hausdorff distance) improve. Without such evidence the claim that two-stage refinement yields superior supervision remains unanchored and load-bearing for the paper’s narrative.
  2. Section 3.2.2 and Fig. 4: during two-stage training the student is supervised by both full-image coarse masks and patch-wise fine masks, while inference follows the same hierarchical crop-and-refine procedure. The paper never reports an ablation that isolates the contribution of the fine-mask branch (or of the presence-head filtering) to final accuracy; consequently it is unclear whether the extra stage is necessary or merely adds complexity that later hurts region-overlap scores.
minor comments (5)
  1. Tables 3–4 repeatedly misspell “IdentityFormer” as “IndentityFormer”; correct throughout.
  2. Figure numbering and captions are inconsistent: two distinct figures are both labeled “Figure 1”, and the caption of the second “Figure 1” refers to pseudo-label masks while the image shows network outputs.
  3. Throughout the text “UAV” appears as “UA V” with an extraneous space; standardize.
  4. Section 4.1: the 8 k-image subsample from the 300 k-image corpus is described only as “randomly sampled”; a short statement on stratification by altitude/scale/background would strengthen reproducibility.
  5. Eq. (12) uses plain BCE; a brief note on whether class imbalance (tiny UAV area) was handled (e.g., weighted BCE or Dice loss) would be helpful.

Circularity Check

0 steps flagged

No circular derivation chain; purely empirical teacher-student transfer from an external foundation model evaluated on held-out public data.

full rationale

The paper presents an empirical engineering framework (SAM3 as black-box pseudo-label oracle + lightweight IPS-Seg student) rather than any first-principles derivation or uniqueness claim. Equations (1)–(12) simply define the data set, the one-/two-stage SAM3 masking operators, the IdentityFormer/ASPP/PixelShuffle architecture, and the BCE training loss; none of them equates a fitted quantity to a claimed prediction by construction. SAM3 is treated as an external, frozen foundation model whose zero-shot masks are used as supervision; IPS-Seg parameters are optimized against those masks and then measured against independent ground-truth annotations on a held-out split of a public dataset. Self-citations (e.g., the author’s prior ITE-U-Net, DualKAN, dehazing networks) appear only as baselines or optional ablation modules and are not load-bearing for the central performance claims. No uniqueness theorem, ansatz smuggled via self-citation, or renaming of a known empirical pattern is invoked. The quantitative–qualitative discrepancy discussion in §4.2.2 is an interpretive hypothesis about metric limitations, not a circular reduction. Consequently the derivation chain is self-contained against external benchmarks and exhibits zero circularity of the kinds enumerated.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central transfer claim rests on a handful of standard ML training choices, the unproved but widely accepted zero-shot competence of SAM3 on aerial small objects, and the architectural hypothesis that an identity-mixer backbone plus ASPP plus PixelShuffle is sufficient for the task. No free physical constants; free parameters are ordinary hyper-parameters. Invented entities are the named network and the two-stage procedure themselves, validated only inside this paper.

free parameters (5)
  • Adam initial learning rate = 1e-4
    Set to 1e-4 with cosine decay; directly controls optimization trajectory of IPS-Seg under both supervision regimes.
  • training epochs = 100
    Fixed at 100; determines how thoroughly the student fits the pseudo-labels.
  • input resolution = 256x256
    All images resized to 256x256; affects both SAM3 prompt quality and student FLOPs.
  • dataset subsample size and split = ~8k / 90-10
    ~8 k images randomly drawn from >300 k, 90/10 train/val; the particular draw is never re-seeded or released.
  • crop padding and text prompt set = small padding; three text prompts
    ‘Small spatial padding’ plus prompts {drone, UAV, quadcopter}; chosen by hand to keep context and trigger SAM3.
axioms (4)
  • domain assumption SAM3 zero-shot masks (with the listed text prompts) are sufficiently accurate teachers for small UAV targets despite the paper’s own admission that SAMs struggle with aerial small objects.
    Invoked throughout §3.2 and Table 2; without it the pseudo-label pipeline collapses.
  • domain assumption Binary cross-entropy against hard pseudo-masks is an adequate training objective for transferring fine boundary detail.
    Eq. (12) and §3.4; no soft-label or boundary-aware loss is explored.
  • ad hoc to paper An identity token-mixer (IdentityFormer) preserves enough spatial information for sub-10-pixel UAV targets.
    §3.3.1 and Table 4; chosen after comparing residual/PoolFormer/ConvFormer variants.
  • ad hoc to paper Lower IoU/Dice of two-stage masks is an artifact of coarse human ground truth rather than true quality loss.
    Explicit explanatory paragraph in §4.2.2; load-bearing for the claim that refinement is beneficial.
invented entities (2)
  • IPS-Seg architecture no independent evidence
    purpose: Lightweight encoder-bottleneck-decoder that can be trained from SAM3 pseudo-labels and still run under tight compute budgets.
    Defined only in this paper; no external validation or independent re-implementation yet.
  • two-stage SAM3-guided pseudo-label pipeline no independent evidence
    purpose: Coarse localization followed by patch-wise re-segmentation to recover fine structure and discard false positives via SAM3’s presence head.
    Core methodological contribution; evidence is internal qualitative figures and the disputed metric explanation.

pith-pipeline@v1.1.0-grok45 · 16555 in / 3039 out tokens · 36816 ms · 2026-07-12T00:10:42.661993+00:00 · methodology

0 comments
read the original abstract

Unmanned aerial vehicle (UAV) target segmentation remains challenging due to the small size of objects, appearance variations, cluttered backgrounds, and the scarcity of densely annotated data. These factors hinder the performance and practical deployment of lightweight segmentation models in real-world UAV applications. To address this problem, this paper investigates the use of SAM3 (Segment Anything Model 3) as a pseudo-label generator for training compact segmentation networks. Specifically, two supervision paradigms are explored: (i) direct pseudo-supervision using unaltered SAM3-generated masks, and (ii) a refinement strategy that re-applies SAM3 to localized image patches for improved mask quality. Based on these paradigms, a two-stage SAM3-guided pseudo-label generation framework is proposed. In the first stage, SAM3 generates coarse masks for initial object localization. The localized regions are subsequently cropped into patches and processed by SAM3 again to generate fine masks with accurate object boundaries and discard false positives. The resulting coarse and fine masks are then used as pseudo-labels to optimize a lightweight network, termed IPS-Seg, which consists of three components: an IdentityFormer backbone for feature extraction, an Atrous Spatial Pyramid Pooling module for multi-scale context aggregation, and a PixelShuffle-based decoder for spatial resolution recovery. Extensive experiments under multiple supervision settings demonstrate the effectiveness of the proposed framework. The results show that IPS-Seg achieves a favorable trade-off between segmentation accuracy and computational efficiency while benefiting consistently from the proposed pseudo-label generation strategy. These findings highlight the potential of large-scale foundation models as annotation sources for training compact task-specific segmentation networks in low-label vision domains.

Figures

Figures reproduced from arXiv: 2607.03754 by Le-Anh Tran.

Figure 1
Figure 1. Figure 1: Performance-efficiency trade-off of various segmentation networks on the UAV Semantic [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The proposed framework. 2.4 Research Gap Although recent studies have partially reduced the dependence on manual annotations, two challenges remain. First, pseudo-labels generated by foundation models often exhibit inaccurate boundaries and fragmented masks when applied to small and distant objects like UAV targets. Second, many high-performance segmentation models remain computationally demanding for depl… view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparison of masks generated by SAM3 and the proposed two-stage strategy. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the proposed two-stage generation/prediction strategy. Candidate regions ex [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 1
Figure 1. Figure 1: Visual comparison of the pseudo-label masks generated by one-stage and two-stage Figure 5: Qualitative comparison of different segmentation approaches [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visual comparison of the pseudo-label masks generated by one-stage and two-stage Figure 6: Qualitative comparison between the one-stage and two-stage strategies [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 6 linked inside Pith

  1. [1]

    A semantic segmentation dataset and real-time localization model for anti-uav applications.Applied Sciences, 15(13):7183, 2025

    Sang-Chul Kim and Yeong Min Jang. A semantic segmentation dataset and real-time localization model for anti-uav applications.Applied Sciences, 15(13):7183, 2025

    2025

  2. [2]

    Improved u-net with identity transformer encoder for efficient uav semantic segmentation.IEEE Access, 13:208962–208972, 2025

    Truong-Dong Do, Le-Anh Tran, Junyong Lee, Sung Kyung Hong, et al. Improved u-net with identity transformer encoder for efficient uav semantic segmentation.IEEE Access, 13:208962–208972, 2025

    2025

  3. [3]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. InProceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

    2023

  4. [4]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. InInternational Conference on Learning Representations, volume 2025, pages 28085–28128, 2025

    2025

  5. [5]

    Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Metaformer baselines for vision.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(2):896–912, 2023

    Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer baselines for vision.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(2):896–912, 2023. 12

    2023

  7. [7]

    Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834–848, 2017

    2017

  8. [8]

    Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network

    Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1874–1883, 2016

    2016

  9. [9]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015

    2015

  10. [10]

    Robust u-net-based road lane markings detection for autonomous driving

    Le-Anh Tran and My-Ha Le. Robust u-net-based road lane markings detection for autonomous driving. In 2019 International Conference on System Science and Engineering (ICSSE), pages 62–66. IEEE, 2019

    2019

  11. [11]

    A novel encoder-decoder network with guided transmission map for single image dehazing.Procedia Computer Science, 204:682–689, 2022

    Le-Anh Tran, Seokyong Moon, and Dong-Chul Park. A novel encoder-decoder network with guided transmission map for single image dehazing.Procedia Computer Science, 204:682–689, 2022

    2022

  12. [12]

    Encoder-decoder networks with guided transmission map for effective image dehazing.The Visual Computer, 41:359–382, 2025

    Le-Anh Tran and Dong-Chul Park. Encoder-decoder networks with guided transmission map for effective image dehazing.The Visual Computer, 41:359–382, 2025

    2025

  13. [13]

    Spatial pyramid pooling in deep convolutional networks for visual recognition.IEEE transactions on pattern analysis and machine intelligence, 37(9): 1904–1916, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition.IEEE transactions on pattern analysis and machine intelligence, 37(9): 1904–1916, 2015

    1904

  14. [14]

    Transunet: Transformers make strong encoders for medical image segmentation.arXiv preprint arXiv:2102.04306, 2021

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation.arXiv preprint arXiv:2102.04306, 2021

    Pith/arXiv arXiv 2021

  15. [15]

    Semi-supervised semantic segmenta- tion needs strong, varied perturbations

    Geoffrey French, Samuli Laine, Timo Aila, and Michal Mackiewicz. Semi-supervised semantic segmenta- tion needs strong, varied perturbations. InBMVC, 2019

    2019

  16. [16]

    Semi-supervised semantic segmentation with cross- consistency training

    Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross- consistency training. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12674–12684, 2020

    2020

  17. [17]

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks

    Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. InICML Workshop on Challenges in Representation Learning, 2013

    2013

  18. [18]

    Self-supervised augmentation consistency for adapting semantic segmentation

    Nikita Araslanov and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15384–15394, 2021

    2021

  19. [19]

    Self-supervised learning of object parts for semantic segmentation

    Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14502–14511, 2022

    2022

  20. [20]

    Sammed: A medical image annotation framework based on large vision model.arXiv preprint arXiv:2307.05617, 3, 2023

    Chenglong Wang, Dexuan Li, Sucheng Wang, Chengxiu Zhang, Yida Wang, Yun Liu, and Guang Yang. Sammed: A medical image annotation framework based on large vision model.arXiv preprint arXiv:2307.05617, 3, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Segment anything is a good pseudo-label generator for weakly supervised semantic segmentation.arXiv preprint arXiv:2305.01275, 2023

    Peng-Tao Jiang and Yuqi Yang. Segment anything is a good pseudo-label generator for weakly supervised semantic segmentation.arXiv preprint arXiv:2305.01275, 2023

    Pith/arXiv arXiv 2023

  22. [22]

    Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation.arXiv preprint arXiv:2305.05803, 2023

    Tianle Chen, Zheda Mai, Ruiwen Li, and Wei-lun Chao. Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation.arXiv preprint arXiv:2305.05803, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  24. [24]

    Pyramid scene parsing network

    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2881–2890, 2017

    2017

  25. [25]

    Unet++: A nested u-net architecture for medical image segmentation

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. InInternational workshop on deep learning in medical image analysis, pages 3–11. Springer, 2018. 13

    2018

  26. [26]

    Road extraction by deep residual u-net.IEEE Geoscience and Remote Sensing Letters, 15(5):749–753, 2018

    Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual u-net.IEEE Geoscience and Remote Sensing Letters, 15(5):749–753, 2018

    2018

  27. [27]

    Encoder-decoder with atrous separable convolution for semantic image segmentation

    Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. InProceedings of the European conference on computer vision (ECCV), pages 801–818, 2018

    2018

  28. [28]

    Resunet++: An advanced architecture for medical image segmentation

    Debesh Jha, Pia H Smedsrud, Michael A Riegler, Dag Johansen, Thomas De Lange, Pål Halvorsen, and Håvard D Johansen. Resunet++: An advanced architecture for medical image segmentation. In2019 IEEE international symposium on multimedia (ISM), pages 225–2255. IEEE, 2019

    2019

  29. [29]

    Distilled pooling transformer encoder for efficient realistic image dehazing.Neural Computing and Applications, 37(6):5203–5221, 2025

    Le-Anh Tran and Dong-Chul Park. Distilled pooling transformer encoder for efficient realistic image dehazing.Neural Computing and Applications, 37(6):5203–5221, 2025

    2025

  30. [30]

    Fill-unet: extended composite semantic segmentation.Applied Soft Computing, 172:112891, 2025

    Qunpo Liu, Yi Zhao, Weiping Ding, Xuhui Bu, and Naohiko Hanajima. Fill-unet: extended composite semantic segmentation.Applied Soft Computing, 172:112891, 2025

    2025

  31. [31]

    Unpaired image dehazing via kolmogorov-arnold transformation of latent features.Pattern Recognition, page 113304, 2026

    Le-Anh Tran. Unpaired image dehazing via kolmogorov-arnold transformation of latent features.Pattern Recognition, page 113304, 2026

    2026

  32. [32]

    Improving convolutional networks with self-calibrated convolutions

    Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Changhu Wang, and Jiashi Feng. Improving convolutional networks with self-calibrated convolutions. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10096–10105, 2020. 14

    2020