Pith. sign in

REVIEW 4 major objections 5 minor 4 cited by

F2Net: A Frequency-Fused Network for Ultra-High Resolution Remote Sensing Segmentation

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

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

Pith's one-line read F2Net claims state-of-the-art semantic segmentation on two ultra-high-resolution remote sensing benchmarks by decomposing images into high- and low-frequency streams, reporting 80.22 mIoU on DeepGlobe and 83.39 on Inria Aerial.

desk verdict Strong reported numbers and a genuinely new architecture, but the key balance loss is underspecified and the text overstates gains; worth refereeing, not yet citable. read the letter →

arxiv 2506.07847 v1 pith:QOWAXJSZ submitted 2025-06-09 cs.CV

Hengzhi Chen , Liqian Feng , Wenhua Wu , Xiaogang Zhu , Shawn Leo , Kun Hu This is my paper
classification cs.CV
keywords ultra-high-resolutionsegmentationremotesensingfrequencydecompositionmulti-branchnetworkcross-frequencylossgradientbalancingstate-spacemodelsemantic
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

F2Net sets out to show that the resolution-versus-context trade-off in ultra-high-resolution remote sensing segmentation can be resolved by separating each image into high- and low-frequency components and processing them in parallel: a full-resolution branch preserves fine boundaries while a downsampled branch captures semantics. To keep these branches from fighting during training, it introduces two auxiliary objectives: an alignment loss that forces semantic consistency across frequencies and a balance loss that equalizes gradient magnitudes across branch parameter sets. On the DeepGlobe and Inria Aerial benchmarks the paper reports state-of-the-art mIoU of 80.22 and 83.39, making F2Net the only method in its comparison to surpass 80% mIoU on DeepGlobe. If these numbers hold, the framework offers a practical accuracy-memory trade-off for images well beyond 4K resolution.

What carries the argument

The load-bearing mechanism is the Adaptive Frequency Decomposition: the input is convolved 1x1, split into channel groups, and for each group a softmax-normalized convolution produces a spatially-varying low-pass kernel; the high-pass kernel is the residual after subtracting it from an all-ones identity kernel, so the decomposition is learned and input-adaptive. Around it, three specialized encoders (a full-resolution state-space high-frequency branch, a CNN short-range branch, and a ViT long-range branch), a Hybrid-Frequency Fusion module using channel and cross-branch attention, and two cross-frequency losses (a symmetric KL alignment loss and a gradient-norm balance loss over branch parameter sets) carry the argument. The balance loss is the least specified link: it is written as $\sum_{\Theta}|G_\Theta - \bar G|$ with $G_\Theta = \|\nabla_\Theta L_{CE}\|_2$, which would require second-order derivatives to optimize as part of $L = \lambda_1 L_{CFAL} + \lambda_2 L_{CFBL} + \lambda_3 L_{CE}$.

What would settle it

Retrain F2Net on DeepGlobe with the paper's stated splits and backbones, with and without CFBL, and check whether (a) the 1.45-point gain (77.35 to 78.80) reproduces, and (b) training with the literal second-order CFBL converges at any choice of $\lambda_2$. If CFBL either diverges or gives no gain, the central claim that gradient balancing drives the improvement is falsified; a simpler falsifier is to run the released code on the standard DeepGlobe validation split and compare mIoU to 80.22.

Watch

Extended reading notes

Core claim

The central claim is that frequency decomposition, rather than spatial cropping or simple downsampling, is the right way to organize multi-branch segmentation of ultra-high-resolution imagery. F2Net learns spatially-adaptive low-pass filters, derives complementary high-pass filters by subtraction, and routes the two components through specialized branches—a state-space model at full resolution for structure, and CNN/ViT sub-branches on a quarter-resolution low-frequency input for local and global semantics. A hybrid fusion module combines the streams, and two cross-frequency losses align their semantics and balance their gradient magnitudes. The paper reports that the complete system reaches 80.22% mIoU on DeepGlobe and 83.39% on Inria Aerial, with ablations attributing 2.87 points of the DeepGlobe gain to the two losses and showing the full three-branch configuration surpasses any single branch by a wide margin.

Load-bearing premise

The load-bearing premise is that the Cross-Frequency Balance Loss can actually be optimized as written—that the sum of absolute deviations of branch gradient norms can be differentiated and trained stably—and that its weight and the other loss weights are set to values that reproduce the reported 80.22/83.39 scores; the paper neither states the computation nor reports the weights.

Editorial extensions

If this is right

  • F2Net's reported results make it the strongest published baseline on DeepGlobe and Inria Aerial, so subsequent UHR segmentation work must match or beat 80.22/83.39 to claim progress.
  • Because the low-frequency branches operate at quarter resolution, the accuracy lost by dropping them to 1/8 is only 1.7 mIoU with 23% less memory, suggesting lightweight deployment variants are available.
  • The two cross-frequency losses, if implementable, transfer to any multi-branch architecture and could stabilize training beyond frequency-based designs.
  • The dynamic frequency decomposition is end-to-end learnable and input-adaptive, so the same module can be dropped into other dense prediction tasks.

Reading between the lines

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

  • A reader should not take the CFBL ablation at face value until the paper specifies how gradients of gradient norms are computed; as written, Eq. 18 inside Eq. 19 requires second-order derivatives, and the weights $\lambda_1,\lambda_2$ are absent.
  • The same frequency-split recipe could be extended to a multi-band or wavelet decomposition, which the authors themselves note as future work; such an extension would test whether the benefit comes from the high/low split or from having any multi-scale specialization.
  • The reported 'only method to surpass 80%' status is protocol-dependent: it holds against the baselines and splits listed in Tables 1 and 2, so reproducing with different training schedules or data partitions could shift the ordering.
  • The gradient-balance loss is a generic mechanism: one could attach it to any multi-branch segmentation network and measure whether it reduces training variance, without any frequency decomposition.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. F2Net is a multi-branch network for ultra-high-resolution (UHR) remote sensing segmentation. The input image is decomposed by a learned adaptive frequency decomposition into high- and low-frequency components. A high-frequency branch (based on VMamba) processes full-resolution features, while a low-frequency branch contains a short-range CNN sub-branch and a long-range Transformer sub-branch operating on downsampled inputs. A Hybrid-Frequency Fusion (HFF) module combines the branch outputs. Two auxiliary losses are proposed: Cross-Frequency Alignment Loss (CFAL), a symmetric KL divergence between branch features, and Cross-Frequency Balance Loss (CFBL), which regularizes the gradient norms of each branch. Experiments on DeepGlobe and Inria Aerial report state-of-the-art mIoU of 80.22 and 83.39, with ablations on branch configurations, input resolutions, loss terms, fusion strategies, and frequency decomposition methods. The paper is written in a workshop style and includes a limitation section.

Significance. If the reported results are reproducible, F2Net would outperform existing UHR segmentation methods on two widely used benchmarks by a substantial margin, and the frequency-decomposition design is a plausible way to balance detail preservation and computational cost. The work also provides a useful set of ablations that isolate the contributions of each branch and each loss. However, the significance is currently tempered by missing implementation details for the proposed CFBL loss (which requires second-order gradients), unreported hyperparameters and variance, and numerically inaccurate claims about the size of the gains. The central idea is interesting, but the evidence supporting the novel loss components is incomplete.

major comments (4)
  1. [§3.7, Eq. (18)–(19)] The Cross-Frequency Balance Loss L_CFBL = Σ_Θ |G_Θ − Ḡ| with G_Θ = ||∇_Θ L_CE||_2 requires differentiating through the gradient norm of the segmentation loss, i.e., second-order derivatives (Hessian–vector products or an equivalent double-backward). The paper does not state how this is implemented, whether it is stabilized (e.g., gradient clipping, detaching, or using a surrogate), or what values of λ1, λ2, λ3 in Eq. (19) are used. Table 5 attributes a 1.45 mIoU gain to CFBL; without these details, that ablation result cannot be reproduced or verified. Please provide the optimization implementation, the chosen hyperparameters, and ideally a reference implementation or pseudocode.
  2. [§4.3, Tables 1–2] The text claims "nearly a 7% improvement compared to X" on DeepGlobe and "a 7% gain in mIoU" on Inria Aerial. Comparing against the previous best UHR methods (SGNet, 75.44 on DeepGlobe and 81.21 on Inria), the absolute mIoU gains are 4.78 and 2.18, respectively; relative gains are about 6.3% and 2.7%. The stated 7% figure is therefore inaccurate and should be corrected. Additionally, the symbol "X" in the DeepGlobe sentence is undefined.
  3. [§3.7, Eq. (17)] CFAL uses the symmetric KL divergence D_KL(F_sl ∥ F_m) + D_KL(F_m ∥ F_sl). The KL divergence is defined for probability distributions, but F_sl and F_m are described as feature maps with no normalization (e.g., softmax along the channel or spatial dimension). As written, the loss is not well-defined. Please specify the normalization step or replace the divergence with a distribution-free similarity measure (e.g., L2 or cosine distance) and justify the choice.
  4. [§4.4, Tables 3–6] All ablation results are reported as single numbers without error bars, multiple seeds, or statistical significance tests. Several key comparisons are close: CFBL+CE (78.80) vs. CFAL+CE (79.20) differ by 0.40 mIoU, and the full model (80.22) is 1.02 mIoU above CFAL+CE. Given the central SOTA claim rests on these differences, the absence of variance reporting makes it unclear whether the observed gains are robust. Please report the mean and standard deviation across at least three runs, or otherwise justify why these differences are meaningful.
minor comments (5)
  1. [Abstract and Introduction] There are several typographical errors: "perservation" in the introduction, "leraning" in the contribution list, and "disting" in the qualitative analysis. These should be corrected.
  2. [§3.6, Eq. (14)–(15)] The cross-branch attention matrix M = σ(A_s A_l^T) has shape C_s × C_l, but in Eq. (15) MLP(M) is added to A_s (shape C_s) and A_l (shape C_l). Please clarify how the dimensions are matched; if an MLP is applied to each row or column, that should be stated explicitly.
  3. [§3.7, Eq. (18)] The notation for the gradient norm is inconsistent: the text writes G_i(t) = ||∇_Θ L_CE||_2, but the summation is over Θ. Please use a consistent notation, e.g., G_Θ = ||∇_Θ L_CE||_2, and specify exactly which parameter sets Θ (high-frequency branch, short-range sub-branch, long-range sub-branch) are included in the sum.
  4. [§4.2] The implementation details state that the long-range sub-branch uses a 6-layer ViT-tiny architecture, but Section 3.5 says it is based on TinyViT. Please reconcile these descriptions and cite the appropriate reference.
  5. [Tables 1–2] The tables have inconsistent formatting: some entries use "—" while others are blank, and SGNet's memory usage is missing on Inria. Please fill in all entries or explain why a value is not available.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: F2Net's SOTA claim is an empirical benchmark comparison; no equation reduces to fitted parameters or self-citation chains.

full rationale

F2Net's contributions are an architecture (frequency decomposition, HFF) and two auxiliary losses. The reported mIoU values are obtained by training on DeepGlobe/Inria Aerial and evaluating on held-out splits, with ablations in Tables 3-7. There is no derivation in which a quantity is defined in terms of the target result, and no fitted parameter is renamed as a prediction. CFBL (Eq. 18) is a regularizer whose optimization may need second-order derivatives, and the weights λ are not reported; this is an implementation/reproducibility gap, not a circularity. The paper contains no load-bearing self-citations; references are to external prior methods. Therefore no circular step can be exhibited.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities. Its central claim relies on untested architectural assumptions and several unreported hyperparameters, plus an implicit comparability of benchmark numbers.

free parameters (5)
  • lambda_1 (CFAL weight) = not reported
    Weight for Cross-Frequency Alignment Loss in Eq. 19; chosen by hand but value absent.
  • lambda_2 (CFBL weight) = not reported
    Weight for Cross-Frequency Balance Loss in Eq. 19; chosen by hand but value absent.
  • lambda_3 (CE weight) = not reported
    Weight for segmentation cross-entropy loss in Eq. 19; chosen by hand but value absent.
  • filter kernel size k = not reported
    Kernel size in Eqs. 4-5 for adaptive frequency decomposition; affects filter behavior, value not reported.
  • number of channel groups N = not reported
    Number of groups in Eq. 3 for channel-wise decomposition; affects decomposition granularity, value not reported.
assumptions (3)
  • domain assumption Softmax-based dynamic kernels produce a valid low-pass filter and identity-minus-low-pass produces a valid high-pass decomposition.
    Sec 3.3 assumes this separation preserves useful segmentation information; no formal analysis provided.
  • ad hoc to paper CFBL is differentiable and optimizable through second-order gradients.
    Eqs. 18-19 define a loss on gradient norms; the required second-order optimization is not described or justified.
  • domain assumption Evaluation protocols and train/validation/test splits match those in [9] and [12] for all baselines.
    Section 4.1 states the splits but does not show that baseline numbers were produced under identical preprocessing and inference settings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of F2Net: A Frequency-Fused Network for Ultra-High Resolution Remote Sensing Segmentation." pith.science (2026). https://pith.science/paper/QOWAXJSZ

@misc{pith2026250607847,
  author       = {Pith},
  title        = {Pith review of: F2Net: A Frequency-Fused Network for Ultra-High Resolution Remote Sensing Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QOWAXJSZ}},
  note         = {Machine review of arXiv:2506.07847}
}
read the original abstract

Semantic segmentation of ultra-high-resolution (UHR) remote sensing imagery is critical for applications like environmental monitoring and urban planning but faces computational and optimization challenges. Conventional methods either lose fine details through downsampling or fragment global context via patch processing. While multi-branch networks address this trade-off, they suffer from computational inefficiency and conflicting gradient dynamics during training. We propose F2Net, a frequency-aware framework that decomposes UHR images into high- and low-frequency components for specialized processing. The high-frequency branch preserves full-resolution structural details, while the low-frequency branch processes downsampled inputs through dual sub-branches capturing short- and long-range dependencies. A Hybrid-Frequency Fusion module integrates these observations, guided by two novel objectives: Cross-Frequency Alignment Loss ensures semantic consistency between frequency components, and Cross-Frequency Balance Loss regulates gradient magnitudes across branches to stabilize training. Evaluated on DeepGlobe and Inria Aerial benchmarks, F2Net achieves state-of-the-art performance with mIoU of 80.22 and 83.39, respectively. Our code will be publicly available.

Figures

Figures reproduced from arXiv: 2506.07847 by the authors.

Figure 1
Figure 1. Comparison of multi-branch architectures for UHR [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed F2Net multi-branch architecture. It begins with a dynamic frequency decomposition [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of branch-relevant heatmaps. (a) High [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Illustrations of qualitative examples from the DeepGlobe dataset. Masks with different colors represent distinct [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Illustrations of qualitative examples from the Inria Aerial dataset. White and black pixels represent building and [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Training convergence with different loss combina [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. WeaveEarth: Structured Evidence Construction and Reasoning for Training-Free UHR Remote Sensing Understanding

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A training-free evidence-selection and topology-preserving packaging pipeline improves frozen VLMs on ultra-high-resolution remote sensing VQA without multi-round search.

  2. GDGS: 3D Gaussian Splatting Via Geometry-Guided Initialization And Dynamic Density Control

    cs.CV 2025-07 conditional novelty 4.0 of 10

    A 3DGS variant that adds MLP initialization, normal alignment, and region-aware density control reports consistent but modest quality gains over vanilla 3DGS on three standard benchmarks.

  3. Building Lightweight Semantic Segmentation Models for Aerial Images Using Dual Relation Distillation

    eess.IV 2025-06 conditional novelty 4.0 of 10

    A student segmentation network trained with spatial and channel relation distillation from a PSPNet ResNet101 teacher gains about 3 to 5 mIoU points on Vaihingen, Potsdam, and Cityscapes.

  4. A Global-Local Cross-Attention Network for Ultra-high Resolution Remote Sensing Image Semantic Segmentation

    cs.CV 2025-06 reject novelty 3.0 of 10

    GLCANet is a dual-branch global-local attention network that reports top mIoU on DeepGlobe, Vaihingen, and Potsdam, but the method and experiments are internally inconsistent and lack code.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages · cited by 4 Pith papers

  1. [1]

    Ascher and E

    S. Ascher and E. Pincus. 1999.The Filmmaker’s Handbook: A Comprehensive Guide for the Digital Age. Plume

    work page 1999

  2. [2]

    Liang-Chieh Chen, Yi Yang, Jiang Wang, Wei Xu, and Alan L. Yuille. 2016. Atten- tion to Scale: Scale-Aware Semantic Image Segmentation. InCVPR

    work page 2016

  3. [3]

    Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. 2018. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. InECCV

    work page 2018

  4. [4]

    Wuyang Chen, Ziyu Jiang, Zhangyang Wang, Kexin Cui, and Xiaoning Qian

    work page

  5. [5]

    Ho Kei Cheng, Jihoon Chung, Yu-Wing Tai, and Chi-Keung Tang. 2020. Cas- cadePSP: Toward Class-Agnostic and Very High-Resolution Segmentation via Global and Local Refinement. InCVPR

    work page 2020

  6. [6]

    Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. 2018. DeepGlobe 2018: A Challenge to Parse the Earth through Satellite Images. InCVPRW

    work page 2018

  7. [7]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. InICLR

    2021

  8. [8]

    Mingyuan Fan, Shenqi Lai, Junshi Huang, Xiaoming Wei, Zhenhua Chai, Jun- feng Luo, and Xiaolin Wei. 2021. Rethinking BiSeNet for Real-Time Semantic Segmentation. InCVPR

    work page 2021

Show all 37 references
  1. [9]

    Shaohua Guo, Liang Liu, Zhenye Gan, Yabiao Wang, Wuhao Zhang, Chengjie Wang, Guannan Jiang, Wei Zhang, Ran Yi, Lizhuang Ma, and Ke Xu. 2022. ISDNet: Integrating Shallow and Deep Networks for Efficient Ultra-High Resolution Segmentation. InCVPR

    2022

  2. [10]

    Chuong Huynh, Anh Tuan Tran, Khoa Luu, and Minh Hoai. 2021. Progressive Semantic Segmentation. InCVPR

    2021

  3. [11]

    Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. 2020. PointRend: Image Segmentation As Rendering. InCVPR

    2020

  4. [12]

    Qi Li, Weixiang Yang, Wenxi Liu, Yuanlong Yu, and Shengfeng He. 2021. From Contexts to Locality: Ultra-High Resolution Image Segmentation via Locality- Aware Contextual Correlation. InICCV

    2021

  5. [13]

    Yifan Li, Ziqian Liu, Junli Yang, and Haopeng Zhang. 2023. Wavelet transform fea- ture enhancement for semantic segmentation of remote sensing images.Remote Sensing(2023)

    2023

  6. [14]

    Chen, and Bing Xu

    Ziming Li, Bin Chen, Shengbiao Wu, Mo Su, Jing M. Chen, and Bing Xu. 2024. Deep learning for urban land use category classification: A review and experi- mental assessment.Remote Sensing of Environment(2024)

    2024

  7. [15]

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. 2024. VMamba: Visual State Space Model. InNeurIPS

    2024

  8. [16]

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. 2015. Fully Convolutional Networks for Semantic Segmentation. InCVPR

    2015

  9. [17]

    Emmanuel Maggiori, Yuliya Tarabalka, Guillaume Charpiat, and Pierre Alliez

  10. [18]

    Salma Abdel Magid, Yulun Zhang, Donglai Wei, Won-Dong Jang, Zudi Lin, Yun Fu, and Hanspeter Pfister. 2021. Dynamic high-pass filtering and multi-spectral attention for image super-resolution. InICCV

    2021

  11. [19]

    Botong Ou, Gang Shao, Baijian Yang, and Songlin Fei. 2025. FocalSR: Revisit- ing image super-resolution transformers with fourier-transform cross attention layers for remote sensing image enhancement.Geomatica(2025)

    2025

  12. [20]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation. InMICCAI

    2015

  13. [21]

    Binbin Song, Songhan Min, Hui Yang, Yongchuang Wu, and Biao Wang. 2023. A Fourier frequency domain convolutional neural network for remote sensing crop classification considering global consistency and edge specificity.Remote Sensing(2023)

    2023

  14. [22]

    Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. 2021. Segmenter: Transformer for Semantic Segmentation. InICCV

    2021

  15. [23]

    Yuxi Sun, Shanshan Feng, Xutao Li, Yunming Ye, Jian Kang, and Xu Huang. 2022. Visual Grounding in Remote Sensing Images. InACM International Conference on Multimedia

    2022

  16. [24]

    Di Wang, Jing Zhang, Bo Du, Minqiang Xu, Lin Liu, Dacheng Tao, and Liangpei Zhang. 2023. SAMRS: Scaling-up Remote Sensing Segmentation Dataset with Segment Anything Model. InNeurIPS

    2023

  17. [25]

    Sai Wang, Yutian Lin, Yu Wu, and Bo Du. 2024. Toward Real Ultra Image Seg- mentation: Leveraging Surrounding Context to Cultivate General Segmentation Model. InNeurIPS

    2024

  18. [26]

    Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. 2022. TinyViT: Fast Pretraining Distillation for Small Vision Transformers. InECCV

    2022

  19. [27]

    Tong Wu, Zhenzhen Lei, Bingqian Lin, Cuihua Li, Yanyun Qu, and Yuan Xie. 2020. Patch Proposal Network for Fast Semantic Segmentation of High-Resolution Conference’17, July 2017, Washington, DC, USA Hengzhi Chen, Liqian Feng, Wenhua Wu, Xiaogang Zhu, Shawn Leo, and Kun Hu Image...

    2020

  20. [28]

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. 2021. SegFormer: Simple and efficient design for semantic segmentation with transformers.Advances in neural information processing systems34 (2021), 12077–12090

    2021

  21. [29]

    Hui Yang, Caili Zhou, Xiaoyu Xing, Yongchuang Wu, and Yanlan Wu. 2024. A High-Resolution Remote Sensing Road Extraction Method Based on the Coupling of Global Spatial Features and Fourier Domain Features.Remote Sensing(2024)

    2024

  22. [30]

    Yunsong Yang, Genji Yuan, and Jinjiang Li. 2024. Sffnet: A wavelet-based spatial and frequency domain fusion network for remote sensing segmentation.IEEE Transactions on Geoscience and Remote Sensing(2024)

    2024

  23. [31]

    Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Jiajun Shen, Jia Li, and Xiaojuan Qi. 2022. Towards efficient and scale-robust ultra-high-definition image demoiréing. In ECCV

    2022

  24. [32]

    Qinyan Zeng, Bin Hui, Zhaoji Liu, Zheng Xu, and Miao He. 2025. A Method Com- bining Discrete Cosine Transform with Attention for Multi-Temporal Remote Sensing Image Matching.Sensors(2025)

    2025

  25. [33]

    Cheng Zhang, Wanshou Jiang, Yuan Zhang, Wei Wang, Qing Zhao, and Chenjie Wang. 2022. Transformer and CNN hybrid deep neural network for semantic segmentation of very-high-resolution remote sensing imagery.IEEE Transactions on Geoscience and Remote Sensing60 (2022), 1–20

    2022

  26. [34]

    Hengshuang Zhao, Xiaojuan Qi, Xiaoyong Shen, Jianping Shi, and Jiaya Jia

  27. [2017]

    InIGARSS

    Can semantic labeling methods generalize to any city? the inria aerial image labeling benchmark. InIGARSS

  28. [2018]

    ICNet for Real-Time Semantic Segmentation on High-Resolution Images. InECCV

  29. [2019]

    Collaborative Global-Local Networks for Memory-Efficient Segmentation of Ultra-High Resolution Images. InCVPR

Pith tools

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