Pith. sign in

REVIEW 2 major objections 5 minor 35 references

This paper claims that the poor performance of open-vocabulary segmentation on remote sensing imagery is largely caused by the text query—the class name is a weak address into the vision-language embedding space—and that learning a pseudo-w

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

2026-08-01 02:02 UTC pith:JSS4EVRR

load-bearing objection A plausible, clearly-written demonstration that textual inversion can fix weak text queries for frozen open-vocabulary segmenters on overhead imagery; the central claim is probably right but two controls are missing. the 2 major comments →

arxiv 2607.25563 v1 pith:JSS4EVRR submitted 2026-07-28 cs.CV

Few-Shot Open-Vocabulary Remote Sensing Segmentation via Textual Inversion

Junhyuk Heo , Junghwan Park This is my paper
classification cs.CV
keywords open-vocabulary segmentationremote sensingtextual inversionfew-shot adaptationtext querypseudo-word embeddingfrozen modelsemantic 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.

The paper argues that the gap in open-vocabulary segmentation on overhead imagery is mostly a problem of the text query, not the segmentation model. A class name acts as an address into a shared vision-language space, and for many aerial categories that address is weak, so the model ranks the region without committing to it. The authors show that a better name repairs some failures, but others require an address that no natural-language rephrasing provides. They recover that address by textual inversion on a fully frozen model, learning pseudo-word embeddings from a few support masks and using those learned tokens as the text query at inference with no support present. On a representative benchmark this raises the mean IoU on affected categories from 3.9 to 39.4, and across eight datasets it beats few-shot methods that instead inject visual prompts.

Core claim

The central claim is that the text query, not the segmentation model, is the bottleneck for open-vocabulary segmentation on overhead imagery. A class name lands at a point in the vision-language embedding space, and when that landing point is weak the model produces a diffuse, non-committal response. The paper demonstrates that renaming recovers categories with nominal failures, while categories like ground track field resist all tested rephrasings and are reached only by a learned address. Using textual inversion on a frozen open-vocabulary segmenter specialized for overhead imagery, the authors fit a compact pseudo-word per category from a handful of masks, then discard the support and que

What carries the argument

The central object is the learned pseudo-word embedding: a small set of input token embeddings (two by default) that replaces the class name in the frozen text encoder's input sequence. Initialized from the class name's own embeddings, these tokens are optimized by textual inversion—fitting a concept from a few examples while the model stays frozen—against a segmentation objective built from the smooth maximum of the instance and semantic readouts gated by a presence score, plus a cross-category negative term. The learned tokens become the query representation that selects a category in the shared embedding space; because every stage of the model is frozen, the learned address acts only by r

Load-bearing premise

The method assumes that for a failing category, the frozen model's image features already rank the true region above its surroundings—so the class name has high AUROC even where IoU is near zero—meaning the only missing ingredient is a text address that concentrates the response; if the features themselves fail to separate the category, a learned token cannot recover it.

What would settle it

A concrete test: pick a category on a dataset outside the paper's main benchmark whose name query has near-chance AUROC (features do not rank the region), fit a learned token from several masks, and check whether IoU rises. The paper predicts it will not, because the failure would be visual rather than nominal. Conversely, if a category with high name-AUROC and low name-AUPRC fails to improve after learning a token, the method's effectiveness is not purely about the query.

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

If this is right

  • If the diagnosis is correct, a large part of the remote-sensing gap in open-vocabulary segmentation can be closed by repairing the text query rather than retraining the model or supplying visual prompts at inference.
  • Enrolling a category is a one-time cost; once the token is learned, inference runs at the speed and memory footprint of a zero-shot text query, with no support image, mask, or feature bank retained.
  • Because the learned text address and ordinary class names are scored by the same full-vocabulary assignment rule, enrolled categories compete directly with named categories without separate calibration.
  • The text-only inference design isolates the contribution of the address itself: any gain over the class name must come from the query, not from visual evidence added at test time.
  • The per-class diagnosis separating nominal from visual failures provides a practical guide for choosing between renaming, learned tokens, and visual grounding.

Where Pith is reading between the lines

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

  • The 'address' framing suggests a general recipe for any domain shift where a frozen vision-language model underperforms on named categories: if the name ranks the region but lacks precision, textual inversion on a few examples could recover it without changing the inference interface.
  • A testable extension: measure the name query's AUROC and AUPRC per category before enrollment; categories with high AUROC but low AUPRC should be exactly the ones the learned token rescues, giving a cheap predictor of benefit.
  • The paper's own limitation (baseball diamond and soccer field) points to an ensemble strategy: route a category to a learned token when its failure is nominal and to visual prompting when it is visual, decided by the same ranking-versus-precision diagnostic.
  • Since the learned token is fixed per category, conditioning it on the input image (while still entering as text) might handle appearance shifts that a single static address cannot follow.

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. This paper studies open-vocabulary semantic segmentation on remote sensing imagery. It argues that a substantial part of the gap to everyday-scene performance is caused by the text query rather than by the segmentation model: for many overhead categories the class name is a weak address in the shared vision-language embedding space, producing diffuse responses. The authors support this with a diagnosis on iSAID, then propose to recover the address by textual inversion on a frozen model: a small set of pseudo-word embeddings per category is optimized on a few support masks, and at inference the category is queried with the learned text token alone. The evaluation covers eight remote sensing datasets and compares against zero-shot queries, a name-ensemble renaming baseline, and two few-shot visual-prompting baselines. On five iSAID categories with near-zero zero-shot IoU, mean IoU rises from 3.9 to 39.4, and the average over eight datasets improves at 5 and 10 shots.

Significance. If the central claim holds, the paper makes a useful contribution: it shows that a frozen open-vocabulary segmenter can be adapted to remote sensing categories through the text interface alone, without carrying visual prompts into inference. The experimental protocol is careful—support masks are drawn from a disjoint pool, the official evaluation split is not used for selection or fitting, and all methods share the same frozen backbone and full-vocabulary argmax. Releasing code is a plus. The per-class separation between categories that a better name repairs (storage tank) and categories that require a learned address (ground track field) is a clean and actionable finding, and the paper is honest about categories where visual prompting remains superior.

major comments (2)
  1. [§3.3, Eq. (1)/(5), §4.3] The central claim needs a calibration check. Eq. (1) feeds p·max(I,S) into a full-vocabulary argmax with a fixed background threshold. Eq. (5) includes a presence term that pushes p upward and segmentation losses that push the soft score on support pixels to 1, with no constraint that the learned token's score distribution match class-name queries. A louder learned query can then claim pixels from quieter name-based categories and background, inflating mIoU independent of any address improvement. The §4.3 AUPRC analysis does not rule this out because AUPRC is rank-based and invariant to rescaling, and the assertion that a per-category name threshold cannot reproduce the gain is not tested. Please report score/presence distributions for learned vs name queries, and run a version with learned-token scores calibrated to the name-query scale before argmax, or a per-category threshold on the
  2. [§4.3, Table 1] The domain-level diagnosis rests on a narrow base. The AUROC/AUPRC analysis covers nine iSAID categories at five shots, and the headline '3.9→39.4' is on categories chosen for zero-shot IoU<10. The method improves the eight-dataset average, but on LoveDA and UDD5 it does not beat zero-shot at five shots, and on baseball diamond and soccer field the visual baselines are better (Table 2). This is consistent with the paper's own limitations, but it means 'much of the gap traces to the text query' is demonstrated only for a subset of iSAID. Please apply the same diagnostic on at least one other dataset (e.g., VDD or OpenEarthMap) or temper the abstract to present the query-side diagnosis as an iSAID result and the eight-dataset comparison as a method evaluation.
minor comments (5)
  1. [Fig. 3b] The PCA projection of token embeddings would be more informative with the percentage of variance explained by the two principal components.
  2. [Table 1] The fully supervised reference is listed alongside few-shot methods without a footnote; consider adding a note that it is not part of the few-shot comparison and is taken from SegEarth-OV3.
  3. [Eq. (4)] The text says the smooth maximum 'converges to the hard maximum as τ approaches zero'; please specify τ→0+ and define the domain of τ.
  4. [§4.1] The language model used to generate candidate names for Name Ensemble is not specified; stating it would improve reproducibility of the renaming baseline.
  5. [Fig. 5 caption] 'crop land' should be 'cropland' for consistency with OpenEarthMap labels.

Circularity Check

0 steps flagged

No circularity: the few-shot token is fit on disjoint support and evaluated on held-out splits; the only self-citation is non-load-bearing.

full rationale

The paper's central claim is that a pseudo-word learned by textual inversion on a frozen segmenter recovers categories whose class-name queries fail. This is not circular: the token v is optimized in Eq. (5) against support masks via Lseg, Lpre, and Lneg, then evaluated with the frozen inference rule of Eq. (1) on an official split disjoint from the support pool. No quantity that defines the method is also used as its prediction. On iSAID, the hard-category set is selected on a held-out selection split before support masks are drawn, so the 3.9-to-39.4 result is a forward evaluation on the official split, not a refit of the reported number. The Section 4.3 diagnostic uses the frozen model's own responses: the name's high AUROC with low AUPRC and the learned token's higher AUPRC are measured on the model's outputs, and although the AUPRC gain is plausibly correlated with the training objective, it is not treated as an input to the fit. The claim that a per-category threshold on the name cannot reproduce the gain may be under-supported, but that is a correctness/validity concern, not a by-construction equivalence. The only self-citation is reference [6], a related-work sentence about diffusion-augmented support; it is not load-bearing. The skeptic's score-scale-shift worry is a legitimate calibration threat to interpretation, but it does not make Eqs. (1)-(5) reduce to their own inputs. Therefore no circular step is established.

Axiom & Free-Parameter Ledger

9 free parameters · 5 axioms · 0 invented entities

The central claim rests on fixed hyperparameters (M=2, lr=1e-2, 300 steps, tau=0.1, lambda_p=0.5, lambda_n=0.5), held constant across datasets, and on the post-hoc selection of iSAID categories with zero-shot IoU<10, which defines the headline gain. No invented physical or model entities beyond the learned token parameter.

free parameters (9)
  • M (number of pseudo-word tokens) = 2
    Chosen by hand as a small budget; ablation shows 4 tokens gives the highest mean (42.2) but they adopt 2 to keep the term compact.
  • learning rate = 1e-2
    Adam, fixed across all datasets.
  • optimization steps = 300
    Fixed across all datasets.
  • EMA decay = 0.99
    Exponential moving average of token over optimization steps.
  • smooth-max temperature tau = 0.1
    Replaces hard max in Eq. (4) during fitting only.
  • lambda_p (presence weight) = 0.5
    Weight on the presence-score supervision term.
  • lambda_n (negative weight) = 0.5
    Weight on the cross-category suppression term.
  • category selection threshold = zero-shot IoU < 10
    The five iSAID 'affected categories' are selected post hoc on a held-out split by this criterion; the headline 3.9→39.4 is measured on this selected set.
  • minimum support mask area = 50 pixels and 1% of patch
    Removes degenerate patches from the fit.
axioms (5)
  • standard math Smooth maximum m_tau converges to hard maximum as tau approaches 0 and stays on the same normalized score scale.
    Standard log-sum-exp property used in Eq. (4).
  • domain assumption The frozen SegEarth-OV3 model's text encoder can interpret pseudo-word embeddings placed in the embedding table as ordinary words.
    Central premise of textual inversion in Section 3.2; if the encoder rejects out-of-vocabulary tokens, the method fails.
  • domain assumption The image encoder's features already separate the target regions (high AUROC), so the failure is one of query precision.
    Section 4.3 shows this on nine iSAID categories at five shots; the cross-dataset claim assumes it holds elsewhere.
  • domain assumption Support masks are representative of the test distribution.
    Few-shot assumption; support drawn from a disjoint pool/split and removed from evaluation.
  • domain assumption Initialization from the class-name embedding is a useful prior.
    Random initialization collapses on baseball diamond and soccer field in Table 3.

pith-pipeline@v1.3.0-alltime-deepseek · 12623 in / 12005 out tokens · 106447 ms · 2026-08-01T02:02:57.937494+00:00 · methodology

0 comments
read the original abstract

Open-vocabulary segmentation labels arbitrary categories from a text query without per-class training, yet on remote sensing imagery it underperforms on categories it handles reliably elsewhere. We find that much of this gap traces to the text query rather than to the segmentation model. Because these models are not specialized for overhead imagery, the class name that serves as the query is often a weak address into the vision-language embedding space. We show that a better name repairs part of the gap, while the remaining failures call for an address that the tested natural-language rephrasings do not provide. We recover that address from a few examples through textual inversion on a frozen model, keeping inference text only. On a representative benchmark this raises the mean intersection over union on the affected categories from 3.9 to 39.4, and across eight remote sensing datasets it improves over few-shot methods that instead inject visual prompts at inference.

Figures

Figures reproduced from arXiv: 2607.25563 by Junghwan Park, Junhyuk Heo.

Figure 1
Figure 1. Figure 1: Motivation. Three categories segmented by the same frozen SegEarth-OV3 model under different text queries. Each row is a category and the columns show the query image, the prediction from the class name, the prediction from a better name, the prediction from a token learned on a few support masks, and the ground truth. White marks the predicted region in the prediction columns and the labeled region in the… view at source ↗
Figure 2
Figure 2. Figure 2: Compared approaches. Five ways to specify a category to the same frozen SegEarth-OV3 model. (a) Zero-shot queries with the class name. (b) Name Ensemble queries with several candidate names, namely synonyms and descriptive phrases, and takes the pixelwise maximum over their masks. (c) SegRAG turns the support masks into point prompts through a DINOv3 feature bank. (d) FSS-SAM3 arranges the sup￾port and the… view at source ↗
Figure 3
Figure 3. Figure 3: Framework and analysis. The pseudo-word token is optimized while all other modules remain frozen, so gradients flow only to the learnable token through the red dashed path, and the learned token is used as the query at inference without the support mask. We further analyze the learned tokens on iSAID under the five-shot setting by projecting the trajectories of the learnable input token embeddings into two… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison. Qualitative comparison on iSAID under full￾vocabulary argmax. The columns show the input, the zero-shot name query, the Name Ensemble query, a visual few-shot baseline, the learned token, and the ground truth. The rows show the categories whose zero-shot IoU is below ten. Support masks are drawn from the training split on the seven additional datasets and from the disjoint support p… view at source ↗
Figure 5
Figure 5. Figure 5: Effect of support size. Pre￾dictions of the learned token as the number of support masks grows, shown in cyan against the ground truth in ma￾genta. On iSAID, where open-vocabulary segmentation fails on categories it han￾dles well in everyday scenes, enrolling the five categories whose zero-shot IoU is be￾low ten recovers them while leaving the rest in place. The full sixteen-way mIoU rises from 27.4 to 42.… view at source ↗
Figure 6
Figure 6. Figure 6: Ranking and precision. Analysis on iSAID at five shots. For nine categories, the AUROC of the name query and the AUPRC of the name query and of the learned token, against the ground truth. For this analysis we enroll a token for nine iSAID categories rather than the five used for the main result, so that the rank corre￾lation rests on enough points. We read the model’s own matching response against the gro… 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

35 extracted references · 10 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2605.17630 (2026)

    Boudiaf, A., Hussain, I., Javed, S.: Segrag: Training-free retrieval-augmented se- mantic segmentation. arXiv preprint arXiv:2605.17630 (2026)

    Pith/arXiv arXiv 2026

  2. [2]

    Journal of Visual Communication and Image Repre- sentation109, 104429 (2025)

    Cai, W., Jin, K., Hou, J., Guo, C., Wu, L., Yang, W.: Vdd: Varied drone dataset for semantic segmentation. Journal of Visual Communication and Image Repre- sentation109, 104429 (2025)

    2025

  3. [3]

    In: Chinese Conference on Pattern Recognition and Computer Vision (PRCV)

    Chen, Y., Wang, Y., Lu, P., Chen, Y., Wang, G.: Large-scale structure from motion with semantic constraints of aerial images. In: Chinese Conference on Pattern Recognition and Computer Vision (PRCV). pp. 347–359. Springer (2018)

    2018

  4. [4]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Cho,S.,Shin,H.,Hong,S.,Arnab,A.,Seo,P.H.,Kim,S.:Cat-seg:Costaggregation for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4113–4123 (2024)

    2024

  5. [5]

    Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image genera- tionusingtextualinversion.In:TheEleventhInternationalConferenceonLearning Representations

  6. [6]

    arXiv preprint arXiv:2503.03785 (2025)

    Immanuel, S.A., Cho, W., Heo, J., Kwon, D.: Tackling few-shot segmentation in remote sensing via inpainting diffusion model. arXiv preprint arXiv:2503.03785 (2025)

    Pith/arXiv arXiv 2025

  7. [7]

    In: International conference on machine learning

    Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904–4916. PMLR (2021)

    2021

  8. [8]

    arXiv preprint arXiv:2201.03546 (2022)

    Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546 (2022)

    Pith/arXiv arXiv 2022

  9. [9]

    arXiv preprint arXiv:2512.08730 (2025)

    Li, K., Zhang, S., Wang, Y., Deng, Y., Wang, Z., Meng, D., Cao, X.: Segearth- ov3: Exploring sam 3 for open-vocabulary semantic segmentation in remote sensing images. arXiv preprint arXiv:2512.08730 (2025)

    Pith/arXiv arXiv 2025

  10. [10]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Mar- culescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 7061–7070 (2023)

    2023

  11. [11]

    Ad- vances in Neural Information Processing Systems35, 17612–17625 (2022)

    Liang, V.W., Zhang, Y., Kwon, Y., Yeung, S., Zou, J.Y.: Mind the gap: Under- standing the modality gap in multi-modal contrastive representation learning. Ad- vances in Neural Information Processing Systems35, 17612–17625 (2022)

    2022

  12. [12]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–16 (2024)

    Liu, F., Chen, D., Guan, Z., Zhou, X., Zhu, J., Ye, Q., Fu, L., Zhou, J.: Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing62, 1–16 (2024)

    2024

  13. [13]

    arXiv preprint arXiv:2305.13310 (2023)

    Liu, Y., Zhu, M., Li, H., Chen, H., Wang, X., Shen, C.: Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310 (2023)

    Pith/arXiv arXiv 2023

  14. [14]

    ISPRS journal of photogrammetry and remote sensing165, 108–119 (2020)

    Lyu, Y., Vosselman, G., Xia, G.S., Yilmaz, A., Yang, M.Y.: Uavid: A semantic seg- mentation dataset for uav imagery. ISPRS journal of photogrammetry and remote sensing165, 108–119 (2020)

    2020

  15. [15]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Min, J., Kang, D., Cho, M.: Hypercorrelation squeeze for few-shot segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6941–6952 (2021)

    2021

  16. [16]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from Few-Shot Open-Vocabulary RS Segmentation via Textual Inversion 17 natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  17. [17]

    In: Artificial Intelligence for Security and Defence Applications III

    Ruis, F., Burghouts, G., Kuijf, H.: Textual inversion for efficient adaptation of open-vocabulary object detectors without forgetting. In: Artificial Intelligence for Security and Defence Applications III. vol. 13679, pp. 481–494. SPIE (2025)

    2025

  18. [18]

    arXiv preprint arXiv:2508.10104 (2025)

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khali- dov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., et al.: Dinov3. arXiv preprint arXiv:2508.10104 (2025)

    Pith/arXiv arXiv 2025

  19. [19]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Singha, M., Pal, H., Jha, A., Banerjee, B.: Ad-clip: Adapting domains in prompt space using clip. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4355–4364 (2023)

    2023

  20. [20]

    ISPRS Ann

    Sohn, G., Jung, J., Gerke, M., Baillard, C., Benitez, S., Breitkopf, U.: The isprs benchmark on urban object classification and 3d building reconstruction. ISPRS Ann. Photogramm. Remote Sens (2012)

    2012

  21. [21]

    arXiv preprint arXiv:2509.26036 (2025)

    Timmermann, C., Lee, H., Lee, W.: Semobridge: Semantic modality bridge for efficient few-shot adaptation of clip. arXiv preprint arXiv:2509.26036 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    arXiv preprint arXiv:2604.05433 (2026)

    Tsai, Y.J., Lin, Y.Y., Wang, C.Y.: Few-shot semantic segmentation meets sam3. arXiv preprint arXiv:2604.05433 (2026)

    Pith/arXiv arXiv 2026

  23. [23]

    IEEE Transactions on Geoscience and Remote Sensing60, 1–18 (2021)

    Wang, B., Wang, Z., Sun, X., Wang, H., Fu, K.: Dmml-net: Deep metametric learning for few-shot geographic object segmentation in remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing60, 1–18 (2021)

    2021

  24. [24]

    arXiv preprint arXiv:2110.08733 (2021)

    Wang, J., Zheng, Z., Ma, A., Lu, X., Zhong, Y.: Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733 (2021)

    Pith/arXiv arXiv 2021

  25. [25]

    In: proceedings of the IEEE/CVF inter- national conference on computer vision

    Wang, K., Liew, J.H., Zou, Y., Zhou, D., Feng, J.: Panet: Few-shot image semantic segmentation with prototype alignment. In: proceedings of the IEEE/CVF inter- national conference on computer vision. pp. 9197–9206 (2019)

    2019

  26. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6830–6839 (2023)

    2023

  27. [27]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Towards segmenting everything in context. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1130–1140 (2023)

    2023

  28. [28]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops

    Waqas Zamir, S., Arora, A., Gupta, A., Khan, S., Sun, G., Shahbaz Khan, F., Zhu, F., Shao, L., Xia, G.S., Bai, X.: isaid: A large-scale dataset for instance segmentation in aerial images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 28–37 (2019)

    2019

  29. [29]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Xia, J., Yokoya, N., Adriano, B., Broni-Bediako, C.: Openearthmap: A bench- mark dataset for global high-resolution land cover mapping. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6254–6264 (2023)

    2023

  30. [30]

    Advances in neural information processing systems34, 12077–12090 (2021)

    Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems34, 12077–12090 (2021)

    2021

  31. [31]

    IEEE Transactions on Geoscience and Remote Sensing60, 1–11 (2021)

    Yao, X., Cao, Q., Feng, X., Cheng, G., Han, J.: Scale-aware detailed matching for few-shot aerial image semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing60, 1–11 (2021)

    2021

  32. [32]

    arXiv preprint arXiv:2405.20141 (2024) 18 J

    Yilmaz, G., Peng, S., Pollefeys, M., Engelmann, F., Blum, H.: Opendas: Open- vocabulary domain adaptation for 2d and 3d segmentation. arXiv preprint arXiv:2405.20141 (2024) 18 J. Heo and J. Park

    Pith/arXiv arXiv 2024

  33. [33]

    In: International Conference on Learning Representations

    Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Qiao, Y., Peng, G., Li, H.: Personalize segment anything model with one shot. In: International Conference on Learning Representations. vol. 2024, pp. 18250–18279 (2024)

    2024

  34. [34]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision- language models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16816–16825 (2022)

    2022

  35. [35]

    International journal of computer vision130(9), 2337–2348 (2022)

    Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International journal of computer vision130(9), 2337–2348 (2022)

    2022