Pith. sign in

REVIEW 4 major objections 4 minor 38 references

This paper argues that few-shot segmentation of remote sensing images can be moved from pixel-level prediction to entity-level reasoning, and that this shift, combined with a physics-inspired advection refinement, lets a frozen foundation m

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 · deepseek-v4-flash

2026-08-03 10:08 UTC pith:JGIHCLDR

load-bearing objection A promising training-free few-shot segmenter for remote sensing with large reported gains, but evaluation gaps—no code, no error bars, and a zero-shot baseline comparison—mean the headline should be read with caution. the 4 major comments →

arxiv 2607.29278 v1 pith:JGIHCLDR submitted 2026-07-31 cs.CV

Training-Free Entity-Level Few-Shot Segmentation of Remote Sensing Images with Advection Refinement

Xueting Bai , Huan Ni This is my paper
classification cs.CV
keywords few-shot segmentationremote sensingtraining-freeentity-level inferenceSAM3advection refinementsemantic potential fieldcross-domain segmentation
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.

This paper argues that few-shot segmentation of remote sensing images is better done at the level of whole regions rather than individual pixels, and that this shift plus a physics-inspired smoothing step is enough to make a frozen foundation model competitive with trained methods. The framework first uses SAM3's automatic masks to partition the query image into category-agnostic regions, then scores each region by combining support-image prototypes, category text embeddings, and dense text responses into a per-category semantic potential field. A two-dimensional advection equation then propagates high-confidence semantic information through both the feature space and the response maps, followed by entity-level classification. Under the 5-shot setting, the authors report a mean intersection-over-union (mIoU) of 36.37% averaged over five remote sensing benchmarks, 10.13 percentage points above the best compared method, with no training or parameter updates.

Core claim

The central claim is that a training-free pipeline can adapt SAM3 to the remote sensing domain by changing the unit of inference. Dividing the query image into spatially disjoint entity primitives, computing per-category foreground and background prototypes from a few target-domain support masks, and fusing those with text-driven responses yields a multimodal semantic potential field. Solving an advection equation on that field moves information from confident regions to uncertain ones, and the authors prove that the gated update never decreases the semantic potential. In the paper's reported numbers, this lifts SAM3 from 8.33% mIoU on the GID-15 dataset to 24.40% in the 5-shot setting, and

What carries the argument

The central object is the category-agnostic entity primitive: SAM3's automatic mask generator produces candidate masks, which are filtered by area and greedily assigned in quality-score order so that every pixel belongs to exactly one primitive, with leftover pixels forming a residual region. The semantic potential field is a weighted combination of foreground similarity, background similarity, and text similarity, normalized to the unit interval. On this field, a two-dimensional advection equation with finite-difference gradients defines a velocity field pointing toward higher semantic confidence; an explicit Euler update with a directional gate transports features and then response maps fr

Load-bearing premise

The framework's gains depend on SAM3's automatic masks being geometrically faithful to real objects; if a mask splits one semantic object or merges two, every pixel in that primitive receives the same label and advection only reweights responses inside the fixed primitive, so the error is locked in.

What would settle it

Compare ELFSS-AR against its own pixel-level baseline on query images where SAM3's masks are known to be poor, such as adjacent farmland parcels or roads with lane markings. If the mIoU gap does not track mask quality, or if deliberately repairing the masks does not improve results, then the entity-primitive assumption is not doing the work claimed.

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

If this is right

  • No source-domain training, target fine-tuning, or prompt engineering is needed; official category names and one to five support masks suffice, so deploying the method on a new remote sensing domain is inexpensive.
  • Region-level aggregation suppresses fragmented false predictions and improves boundary consistency, which is reflected in higher Boundary IoU alongside mIoU.
  • The advection refinement provably does not decrease category confidence at any step, giving a monotone semantic correction mechanism.
  • The method handles heterogeneous sensor and spectral configurations, treating aerial and satellite imagery at different resolutions as unseen target domains.
  • Because all parameters are frozen, the approach can be updated by swapping the underlying foundation model without retraining.

Where Pith is reading between the lines

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

  • The paper does not examine what happens when SAM3's masks fail; a testable follow-up is to degrade mask quality deliberately and show that the mIoU gain over pixel-level inference shrinks accordingly.
  • The advection refinement is a generic spatial regularization operator; the same gated Euler scheme could be transferred to other dense prediction problems where a semantic potential field can be defined, such as change detection or instance-level refinement.
  • The method's reliance on official category names as prompts may underestimate its headroom; prompt variants or per-category paraphrases could widen the gap over trained baselines, though the paper intentionally avoids prompt engineering.
  • Since the framework is training-free, it can serve as a strong baseline or initialization for hybrid approaches that later fine-tune on target data; the reported gains may not persist after training.

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

4 major / 4 minor

Summary. ELFSS-AR proposes a training-free, entity-level few-shot segmentation framework for remote sensing images. SAM3's automatic mask generation partitions the query image into category-agnostic entity primitives; support prototypes and category text embeddings are combined into a multi-modal semantic potential field; an advection-equation refinement propagates information in feature and response spaces; and each entity is assigned one label. Experiments on GID-15, FBP, Potsdam, Vaihingen, and iSAID under 1-shot and 5-shot settings report average mIoU of 31.70% and 36.37%, respectively, outperforming trained FSS/CDFSS methods and training-free OVSS baselines, with a 10.13 pp margin over IFA at 5-shot. Ablations on three datasets support the contribution of entity-level inference and of feature/response advection. An appendix formalizes entity generation and gives a theoretical analysis of feature advection.

Significance. If the reported results are reproducible, the paper makes a useful empirical contribution: it demonstrates that a foundation-model-based, training-free pipeline can achieve strong cross-domain few-shot segmentation without source-domain training, and the ablation study helps isolate the proposed components. The use of five public remote-sensing benchmarks and the inclusion of a direct SAM3 zero-shot baseline are strengths. However, the central quantitative claims are currently not fully supported by the manuscript as submitted: no code or technical supplement is available, no error bars or significance tests are reported, all sixteen scalar hyperparameters are set empirically without a validation protocol or sensitivity analysis, the entity-level ablation omits two of the five datasets, and the theoretical guarantee in Appendix C is not aligned with the actual update equations. These issues are addressable and should be fixed before the paper can be accepted.

major comments (4)
  1. [Methods, Eqs. (6), (13)–(17), (21); Appendix A] Sixteen scalar hyperparameters (λF_fg=1.20, λF_bg=0.75, λF_txt=0.05, δF=0.06, γF=0.18, NF=2, γR=0.18, ρ=0.50, w_fg=1.25, w_bg=1.35, w_txt=0.03, w_sem=0.25, w_sam=0.01, w_ctx=0.01, θ_area=10, N_max=500) are all described as ‘empirically set’ with no search procedure, no validation split, and no sensitivity analysis. Since the headline claim is a 10.13 pp margin over IFA (Table 2), test-set tuning cannot be excluded. Please report how these values were chosen, a sensitivity study for the most influential parameters, or evidence that performance is stable under reasonable perturbations.
  2. [Tables 2–4] All mIoU/BIoU numbers are point estimates without standard deviations, number of support/query splits, or significance tests. The paper states a 5-shot average of 36.37% versus 26.24% for IFA, but single-run results do not establish that component improvements in Table 4 (e.g., ELFSS vs ELFSS+FA: 59.44→60.21 on Vaihingen, 19.52→20.66 on iSAID) are statistically reliable. Please report mean±std over at least three random splits (support selection and query set) and, if feasible, a paired significance test.
  3. [Appendix A, Table 4, Conclusion] The entity-level mechanism rests on SAM3 masks being reliable category-agnostic partitions. Appendix A constructs a hard partition (Eqs. 31–36) and Eq. (22) forces all pixels in an entity to share one label; over- or under-segmentation is therefore locked in. The Conclusion acknowledges that ‘the reliability of entity primitive construction could be further enhanced,’ but Table 4 ablates entity-level FSS only on Vaihingen, iSAID, and GID-15, omitting Potsdam and FBP where the full method is evaluated. No quantitative mask-quality measure is reported. Please add mask-quality diagnostics (e.g., achievable segmentation accuracy, boundary precision/recall) and extend the entity-level ablation to all five datasets, or qualify the attribution of the gains on Potsdam/FBP.
  4. [Appendix C, Proposition 2, Eq. (14)] Proposition 2 claims a non-decreasing semantic potential for the ‘discretized advection update,’ but the proof analyzes only the intermediate update in Eq. (40). The implemented feature update is Eq. (14), which includes the residual interpolation with γF=0.18 and channel-wise L2 normalization; the response update Eq. (16) includes clipping and the final fusion Eq. (17) with ρ=0.50. The proof also treats the potential U_F_c as a local function of the feature while Eq. (5) applies spatial min-max normalization. The stated guarantee is therefore not established for the actual algorithm. Please state precisely which invariant is proved, align the proof with the deployed update, or downgrade the claim to a heuristic motivation supported by the ablations.
minor comments (4)
  1. [Reproducibility] The manuscript repeatedly refers to a ‘Technical Supplement’ that is not included with the arXiv submission, and the code link states the code ‘will be publicly available’ rather than providing it. Please include the supplement and an anonymized code release for review.
  2. [Table 3 / Section 'Results'] The comparison with training-free OVSS methods is not zero-shot for ELFSS-AR, since it additionally receives one or five support images/masks. The caption notes this, but the main text should state this task difference more prominently to avoid overclaiming a zero-shot comparison.
  3. [Appendix C, Proposition 2 heading] Typo: ‘Poof’ should be ‘Proof’.
  4. [Eq. (16)] The response advection update clamps to [−1, 1] after each iteration, but there is no discussion of numerical stability or a CFL-type condition for the chosen δF and γR. A brief stability note would help readers apply the method to other resolutions.

Circularity Check

1 steps flagged

Advection-refinement 'guarantee' is a tautology of the gate definition; the headline mIoU gains rest on external benchmarks, so circularity is partial, not structural.

specific steps
  1. self definitional [Appendix C, Proposition 2 (Eqs. 46–53); main paper Eqs. (8)–(12)]
    "Consequently, only candidate advection updates that increase the semantic potential are accepted, whereas updates that would decrease the semantic confidence are discarded. Therefore, throughout the iterative discretization of the advection equation, the semantic confidence encoded by the semantic potential field is guaranteed to be non-decreasing."

    The 'semantic confidence' is defined to be the semantic potential field U (Eqs. 5–6), and the velocity field is defined as v = -∇U (Eqs. 8–9). Therefore the directional derivative d = v·∇U is non-positive by construction, so the gate 1[d<0] in Eq. (12) is always active. Proposition 2's non-decrease conclusion is thus a restatement of the gate's own definition, not an independent derivation. It also does not establish that increasing U improves segmentation accuracy; that claim is supported only by the external ablation study.

full rationale

The central experimental claim — 36.37% average mIoU under 5-shot, +10.13 points over IFA — is validated against external trained FSS/CDFSS baselines and training-free OVSS baselines on five datasets, so it is not a fitted parameter renamed as a prediction. The entity-level reformulation is a design choice built on SAM3's masks, and the paper explicitly acknowledges that 'the reliability of entity primitive construction could be further enhanced,' which is an attribution/robustness risk rather than a circular step. The only reduction-by-construction we identified is the theoretical 'guarantee' in Appendix C: semantic confidence is equated with the semantic potential field, the velocity is its negative gradient, and the gate accepts exactly the updates that increase that potential, making the non-decrease claim true by definition. This tautology does not drive the headline numbers, which come from external benchmarks and ablations. No load-bearing self-citation was found; the only overlapping-author citation (SegEarth-OV, Bai X.) appears in a related-work list and is not used to justify the framework. Overall, the paper's empirical contribution is self-contained, but one supporting theoretical argument is circular by construction, giving a moderate score.

Axiom & Free-Parameter Ledger

16 free parameters · 5 axioms · 0 invented entities

The framework introduces no new physical or ontological entities; 'entity primitives', 'semantic potential field', and 'semantic velocity field' are algorithmic constructions over SAM3 outputs and similarity maps. The central burden is the large number of hand-set fusion weights and advection parameters, plus the domain assumptions that SAM3 masks and cosine similarities are reliable on remote sensing data.

free parameters (16)
  • lambda_F_fg = 1.20
    Hand-set weight on foreground response in the feature-level potential field, Eq. (6); no sensitivity analysis.
  • lambda_F_bg = 0.75
    Hand-set weight on background suppression, Eq. (6).
  • lambda_F_txt = 0.05
    Hand-set weight on textual response, Eq. (6).
  • delta_F = 0.06
    Feature-advection time step, Eq. (13); empirically set, no stability analysis.
  • gamma_F = 0.18
    Feature-advection residual interpolation strength, Eq. (14).
  • N_F = 2
    Number of feature-advection iterations; empirically fixed.
  • gamma_R = 0.18
    Response-advection residual interpolation strength, Eq. (16).
  • rho = 0.50
    Retention weight fusing original and refined responses, Eq. (17).
  • w_fg = 1.25
    Fusion weight on aggregated foreground response, Eq. (21).
  • w_bg = 1.35
    Fusion weight on aggregated background response, Eq. (21).
  • w_txt = 0.03
    Fusion weight on aggregated textual response, Eq. (21).
  • w_sem = 0.25
    Fusion weight on response-level semantic potential, Eq. (21).
  • w_sam = 0.01
    Fusion weight on SAM3 quality score, Eq. (21).
  • w_ctx = 0.01
    Fusion weight on contextual support, Eq. (21).
  • theta_area = 10 pixels
    Minimum area threshold for entity primitives, Eq. (28).
  • N_max = 500
    Maximum number of entity primitives, Appendix A.
axioms (5)
  • domain assumption Remote sensing scenes are approximately piecewise-constant semantic functions over spatially continuous regions.
    Motivates entity-level aggregation; introduced in Appendix A, Eqs. (23)-(25). If semantic boundaries do not align with SAM3 entities, the variance-reduction argument does not translate into accuracy.
  • domain assumption SAM3's automatic mask generation produces reliable category-agnostic geometric priors for remote sensing objects.
    Entity primitives inherit SAM3's region boundaries, and no correction mechanism exists for over- or under-segmentation; the authors themselves flag this in the Conclusion.
  • domain assumption Cosine similarity between frozen SAM3 features and support prototypes is a meaningful semantic signal under domain shift.
    Eqs. (3)-(4) define the potential field; if frozen features are non-transferable to the target domain, advection refinement amplifies misleading similarities.
  • domain assumption Explicit Euler discretization of the advection equation with the chosen step sizes is numerically stable and preserves beneficial feature/response structure.
    Eqs. (13)-(16); no stability or boundedness analysis is supplied, and Proposition 2 drops O(delta_F^2) terms.
  • domain assumption Official remote sensing category names, passed as text prompts, are sufficiently interpretable by SAM3's text encoder to provide useful semantic responses.
    Textual response enters the potential field with weight lambda_txt=0.05, Eq. (6); the authors note official names are often highly specialized and hard for foundation models to interpret.

pith-pipeline@v1.3.0-daily-deepseek · 72 in / 13442 out tokens · 265024 ms · 2026-08-03T10:08:26.508688+00:00 · methodology

0 comments
read the original abstract

Existing cross-domain few-shot segmentation approaches suffer from high training costs due to source-domain episodic training and pixel-wise dense prediction, while often producing fragmented and noisy predictions. To overcome these issues, we propose a training-free entity-level few-shot segmentation framework for remote sensing images with advection refinement. Specifically, we first leverage SAM3's generic geometric priors to generate category-agnostic entity primitives. By reformulating few-shot inference from pixel-level prediction to entity-level reasoning, foreground and background prototypes are constructed and combined with dense textual semantic responses from SAM3 to build a multi-modal semantic potential field. Furthermore, an advection equation-based semantic refinement mechanism is introduced to propagate category-aware information across both feature and similarity spaces, enhancing semantic continuity and suppressing local texture noise. Extensive experiments on multiple remote sensing datasets demonstrate that the proposed framework effectively mitigates domain shift and local noise, substantially improving SAM3's adaptation capability for remote sensing few-shot segmentation without additional training. Our code will be publicly available at https://github.com/yu-ni1989/ELFSS-AR.

Figures

Figures reproduced from arXiv: 2607.29278 by Huan Ni, Xueting Bai.

Figure 1
Figure 1. Figure 1: Overview of the proposed Training-Free Entity-Level Few-Shot Segmentation framework with Advection Refinement [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Segmentation results under the 5-shot setting. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy curves of the ablation study under the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of the ablation study under the 5-shot setting. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Category-wise feature maps before and after ad [PITH_FULL_IMAGE:figures/full_fig_p007_5.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

38 extracted references · 1 canonical work pages

  1. [1]

    Everingham, Mark and Eslami, S. M. Ali and Van Gool, Luc and Williams, Christopher K. I. and Winn, John and Zisserman, Andrew , date =. The Pascal Visual Object Classes Challenge: A Retrospective , url =. International Journal of Computer Vision , number =. 2015 , bdsk-url-1 =. doi:10.1007/s11263-014-0733-5 , id =

    doi:10.1007/s11263-014-0733-5 2015

  2. [2]

    Results of the ISPRS benchmark on urban object detection and 3D building reconstruction , url =

    Franz Rottensteiner and Gunho Sohn and Markus Gerke and Jan Dirk Wegner and Uwe Breitkopf and Jaewook Jung , date-added =. Results of the ISPRS benchmark on urban object detection and 3D building reconstruction , url =. ISPRS Journal of Photogrammetry and Remote Sensing , keywords =. 2014 , bdsk-url-1 =. doi:https://doi.org/10.1016/j.isprsjprs.2013.10.004...

    doi:10.1016/j.isprsjprs.2013.10.004 2014

  3. [3]

    Enabling country-scale land cover mapping with meter-resolution satellite imagery , url =

    Xin-Yi Tong and Gui-Song Xia and Xiao Xiang Zhu , date-added =. Enabling country-scale land cover mapping with meter-resolution satellite imagery , url =. ISPRS Journal of Photogrammetry and Remote Sensing , keywords =. 2023 , bdsk-url-1 =. doi:https://doi.org/10.1016/j.isprsjprs.2022.12.011 , issn =

    work page doi:10.1016/j.isprsjprs.2022.12.011 2023

  4. [4]

    Land-cover classification with high-resolution remote sensing images using transferable deep models , url =

    Xin-Yi Tong and Gui-Song Xia and Qikai Lu and Huanfeng Shen and Shengyang Li and Shucheng You and Liangpei Zhang , date-added =. Land-cover classification with high-resolution remote sensing images using transferable deep models , url =. Remote Sensing of Environment , keywords =. 2020 , bdsk-url-1 =. doi:https://doi.org/10.1016/j.rse.2019.111322 , issn =

    arXiv 2020

  5. [5]

    ProxyCLIP: Proxy Attention Improves CLIP for Open-Vocabulary Segmentation , year =

    Lan, Mengcheng and Chen, Chaofeng and Ke, Yiping and Wang, Xinjiang and Feng, Litong and Zhang, Wayne , booktitle =. ProxyCLIP: Proxy Attention Improves CLIP for Open-Vocabulary Segmentation , year =

  6. [6]

    Explore the Potential of CLIP for Training-Free Open Vocabulary Semantic Segmentation , year =

    Shao, Tong and Tian, Zhuotao and Zhao, Hang and Su, Jingyong , booktitle =. Explore the Potential of CLIP for Training-Free Open Vocabulary Semantic Segmentation , year =

  7. [7]

    ClearCLIP: Decomposing CLIP Representations for Dense Vision-Language Inference , year =

    Lan, Mengcheng and Chen, Chaofeng and Ke, Yiping and Wang, Xinjiang and Feng, Litong and Zhang, Wayne , booktitle =. ClearCLIP: Decomposing CLIP Representations for Dense Vision-Language Inference , year =

  8. [8]

    SCLIP: Rethinking Self-Attention for Dense Vision-Language Inference , year =

    Wang, Feng and Mei, Jieru and Yuille, Alan , booktitle =. SCLIP: Rethinking Self-Attention for Dense Vision-Language Inference , year =

  9. [9]

    Extract Free Dense Labels from CLIP , year =

    Zhou, Chong and Loy, Chen Change and Dai, Bo , booktitle =. Extract Free Dense Labels from CLIP , year =

  10. [10]

    Segment and Recognize Anything at Any Granularity , year =

    Li, Feng and Zhang, Hao and Sun, Peize and Zou, Xueyan and Liu, Shilong and Li, Chunyuan and Yang, Jianwei and Zhang, Lei and Gao, Jianfeng , booktitle =. Segment and Recognize Anything at Any Granularity , year =

  11. [11]

    One-Shot Learning for Semantic Segmentation , url =

    Shaban, Amirreza and Bansal, Shray and Liu, Zhen and Essa, Irfan and Boots, Byron , booktitle =. One-Shot Learning for Semantic Segmentation , url =. 2017 , bdsk-url-1 =. doi:10.5244/C.31.167 , editor =

    doi:10.5244/c.31.167 2017

  12. [12]

    PANet: Few-Shot Image Semantic Segmentation With Prototype Alignment , year =

    Wang, Kaixin and Liew, Jun Hao and Zou, Yingtian and Zhou, Daquan and Feng, Jiashi , booktitle =. PANet: Few-Shot Image Semantic Segmentation With Prototype Alignment , year =. doi:10.1109/ICCV.2019.00929 , keywords =

    arXiv 2019

  13. [13]

    Prior Guided Feature Enrichment Network for Few-Shot Segmentation , volume =

    Tian, Zhuotao and Zhao, Hengshuang and Shu, Michelle and Yang, Zhicheng and Li, Ruiyu and Jia, Jiaya , date-modified =. Prior Guided Feature Enrichment Network for Few-Shot Segmentation , volume =. 2022 , bdsk-url-1 =. doi:10.1109/TPAMI.2020.3013717 , journal =

    arXiv 2022

  14. [14]

    PFENet++: Boosting Few-Shot Semantic Segmentation With the Noise-Filtered Context-Aware Prior Mask , volume =

    Luo, Xiaoliu and Tian, Zhuotao and Zhang, Taiping and Yu, Bei and Tang, Yuan Yan and Jia, Jiaya , date-modified =. PFENet++: Boosting Few-Shot Semantic Segmentation With the Noise-Filtered Context-Aware Prior Mask , volume =. 2024 , bdsk-url-1 =. doi:10.1109/TPAMI.2023.3329725 , journal =

    arXiv 2024

  15. [15]

    Self-support

    Fan, Qi and Pei, Wenjie and Tai, Yu-Wing and Tang, Chi-Keung , booktitle =. Self-support

  16. [19]

    Learning What Not to Segment: A New Perspective on Few-Shot Segmentation , year =

    Lang, Chunbo and Cheng, Gong and Tu, Binfei and Han, Junwei , booktitle =. Learning What Not to Segment: A New Perspective on Few-Shot Segmentation , year =. doi:10.1109/CVPR52688.2022.00789 , keywords =

    arXiv 2022

  17. [20]

    Prediction Calibration for Generalized Few-Shot Semantic Segmentation , volume =

    Lu, Zhihe and He, Sen and Li, Da and Song, Yi-Zhe and Xiang, Tao , date-modified =. Prediction Calibration for Generalized Few-Shot Semantic Segmentation , volume =. 2023 , bdsk-url-1 =. doi:10.1109/TIP.2023.3282070 , journal =

    arXiv 2023

  18. [22]

    Unlocking the Power of

    Qianxiong Xu and Lanyun Zhu and Xuanyi Liu and Guosheng Lin and Cheng Long and Ziyue Li and Rui Zhao , booktitle =. Unlocking the Power of. 2025 , bdsk-url-1 =

    2025

  19. [23]

    Cross-Domain Few-Shot Semantic Segmentation , year =

    Lei, Shuo and Zhang, Xuchao and He, Jianfeng and Chen, Fanglan and Du, Bowen and Lu, Chang-Tien , booktitle =. Cross-Domain Few-Shot Semantic Segmentation , year =

  20. [28]

    Bayesian Decomposition and Semantic Completion for Few-shot Semantic Segmentation , url =

    Shi, Guangchen and Wu, Yirui and Zhu, Wei and Wang, Tao and Zhang, Hao and Li, Bo and Lu, Tong , booktitle =. Bayesian Decomposition and Semantic Completion for Few-shot Semantic Segmentation , url =. 2026 , bdsk-url-1 =

    2026

  21. [31]

    Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation , url =

    Jintao Tong and Ran Ma and Yixiong Zou and Guangyao Chen and Yuhua Li and Ruixuan Li , booktitle =. Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation , url =. 2025 , bdsk-url-1 =

    2025

  22. [32]

    Deep Residual Learning for Image Recognition , year =

    He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian , booktitle =. Deep Residual Learning for Image Recognition , year =. doi:10.1109/CVPR.2016.90 , keywords =

    doi:10.1109/cvpr.2016.90 2016

  23. [33]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , url =

    Alexey Dosovitskiy and Lucas Beyer and Alexander Kolesnikov and Dirk Weissenborn and Xiaohua Zhai and Thomas Unterthiner and Mostafa Dehghani and Matthias Minderer and Georg Heigold and Sylvain Gelly and Jakob Uszkoreit and Neil Houlsby , booktitle =. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , url =. 2021 , bdsk-url-1 =

    2021

  24. [34]

    The Thirteenth International Conference on Learning Representations , date-modified =

    Nikhila Ravi and Valentin Gabeur and Yuan-Ting Hu and Ronghang Hu and Chaitanya Ryali and Tengyu Ma and Haitham Khedr and Roman R. The Thirteenth International Conference on Learning Representations , date-modified =. 2025 , bdsk-url-1 =

    2025

  25. [35]

    Generalized Decoding for Pixel, Image, and Language , year =

    Zou, Xueyan and Dou, Zi-Yi and Yang, Jianwei and Gan, Zhe and Li, Linjie and Li, Chunyuan and Dai, Xiyang and Behl, Harkirat and Wang, Jianfeng and Yuan, Lu and Peng, Nanyun and Wang, Lijuan and Lee, Yong Jae and Gao, Jianfeng , booktitle =. Generalized Decoding for Pixel, Image, and Language , year =. doi:10.1109/CVPR52729.2023.01451 , keywords =

    arXiv 2023

  26. [36]

    Segment Everything Everywhere All at Once , url =

    Xueyan Zou and Jianwei Yang and Hao Zhang and Feng Li and Linjie Li and Jianfeng Wang and Lijuan Wang and Jianfeng Gao and Yong Jae Lee , booktitle =. Segment Everything Everywhere All at Once , url =. 2023 , bdsk-url-1 =

    2023

  27. [37]

    A Simple Framework for Open-Vocabulary Segmentation and Detection , year =

    Zhang, Hao and Li, Feng and Zou, Xueyan and Liu, Shilong and Li, Chunyuan and Yang, Jianwei and Zhang, Lei , booktitle =. A Simple Framework for Open-Vocabulary Segmentation and Detection , year =. doi:10.1109/ICCV51070.2023.00100 , keywords =

    arXiv 2023

  28. [41]

    PixelLM: Pixel Reasoning with Large Multimodal Model , year =

    Ren, Zhongwei and Huang, Zhicheng and Wei, Yunchao and Zhao, Yao and Fu, Dongmei and Feng, Jiashi and Jin, Xiaojie , booktitle =. PixelLM: Pixel Reasoning with Large Multimodal Model , year =. doi:10.1109/CVPR52733.2024.02491 , keywords =

    arXiv 2024

  29. [42]

    Pay Attention to Your Neighbours: Training-Free Open-Vocabulary Semantic Segmentation , year =

    Hajimiri, Sina and Ayed, Ismail Ben and Dolz, Jose , booktitle =. Pay Attention to Your Neighbours: Training-Free Open-Vocabulary Semantic Segmentation , year =. doi:10.1109/WACV61041.2025.00495 , keywords =

    arXiv 2025

  30. [43]

    PEARL: Geometry Aligns Semantics for Training-Free Open-Vocabulary Semantic Segmentation , year =

    Pei, Gensheng and Jiang, Xiruo and Cai, Xinhao and Chen, Tao and Yao, Yazhou and Jeon, Byeungwoo , booktitle =. PEARL: Geometry Aligns Semantics for Training-Free Open-Vocabulary Semantic Segmentation , year =

  31. [45]

    Harnessing Vision Foundation Models for High-Performance, Training-Free Open Vocabulary Segmentation , year =

    Shi, Yuheng and Dong, Minjing and Xu, Chang , booktitle =. Harnessing Vision Foundation Models for High-Performance, Training-Free Open Vocabulary Segmentation , year =. doi:10.1109/ICCV51701.2025.02180 , keywords =

    arXiv 2025

  32. [46]

    SegEarth-OV: Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images , year =

    Li, Kaiyu and Liu, Ruixun and Cao, Xiangyong and Bai, Xueru and Zhou, Feng and Meng, Deyu and Wang, Zhi , booktitle =. SegEarth-OV: Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images , year =. doi:10.1109/CVPR52734.2025.00986 , keywords =

    arXiv 2025

  33. [47]

    Advances in neural information processing systems , title =

    Snell, Jake and Swersky, Kevin and Zemel, Richard , date-modified =. Advances in neural information processing systems , title =

  34. [48]

    ReAttnCLIP: Training-Free Open-Vocabulary Remote Sensing Image Segmentation via Re-defined Attention in CLIP , year =

    Niu, Xin and Zhao, Manqi and Jiang, Dongsheng and Wu, Yingying and Su, Bing , booktitle =. ReAttnCLIP: Training-Free Open-Vocabulary Remote Sensing Image Segmentation via Re-defined Attention in CLIP , year =

  35. [49]

    Junjue Wang and Zhuo Zheng and Ailong Ma and Xiaoyan Lu and Yanfei Zhong , date-modified =. Love. 2021 , bdsk-url-1 =

    2021

  36. [50]

    isaid: A large-scale dataset for instance segmentation in aerial images , year =

    Waqas Zamir, Syed and Arora, Aditya and Gupta, Akshita and Khan, Salman and Sun, Guolei and Shahbaz Khan, Fahad and Zhu, Fan and Shao, Ling and Xia, Gui-Song and Bai, Xiang , booktitle =. isaid: A large-scale dataset for instance segmentation in aerial images , year =

  37. [51]

    The Fourteenth International Conference on Learning Representations , date-modified =

    Nicolas Carion and Laura Gustafson and Yuan-Ting Hu and Shoubhik Debnath and Ronghang Hu and Didac Suris Coll-Vinent and Chaitanya Ryali and Kalyan Vasudev Alwala and Haitham Khedr and Andrew Huang and Jie Lei and Tengyu Ma and Baishan Guo and Arpit Kalla and Markus Marks and Joseph Greer and Meng Wang and Peize Sun and Roman R. The Fourteenth Internation...

    2026

  38. [52]

    Implicit CLIP Prior Decoupling for Few-Shot Remote Sensing Image Segmentation , volume =

    Jiang, Zhiyu and Yuan, Ye and Ma, Dandan and Wang, Qi and Yuan, Yuan , date-modified =. Implicit CLIP Prior Decoupling for Few-Shot Remote Sensing Image Segmentation , volume =. 2025 , bdsk-url-1 =. doi:10.1109/TGRS.2025.3617662 , journal =

    arXiv 2025