Pith. sign in

REVIEW 3 major objections 5 minor 61 references

ZoRI: Towards Discriminative Zero-Shot Remote Sensing Instance Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read A new framework, ZoRI, claims state-of-the-art zero-shot instance segmentation for remote sensing by refining CLIP text channels, adapting only part of the image encoder, and injecting aerial visual prototypes.

desk verdict New task and benchmarks, reasonable DEC/KMA ideas, but the headline SOTA claim is compromised by a transductive PIP loop that uses the model's own test-set predictions to re-score the same test images. read the letter →

arxiv 2412.12798 v1 pith:DMOAOZHM submitted 2024-12-17 cs.CV

classification cs.CV
keywords zero-shotinstancesegmentationremotesensingvision-languagemodelsCLIPadaptationcachebankchannelselectioniSAIDNWPU-VHR-10
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

The paper introduces zero-shot remote sensing instance segmentation as a new task and proposes ZoRI to solve it. It claims that ZoRI segments aerial object classes never seen during training, reaching an unseen-class mAP of 9.30 on iSAID under the generalized zero-shot setting, compared with 4.91 for the FC-CLIP baseline. The method combines three mechanisms: channel selection that sharpens the text classifier, partial adaptation of the vision encoder that preserves pretrained vision-language alignment, and a cache bank of aerial visual prototypes that reweights the final predictions. If the claim holds, it would let vision-language models recognize novel aerial categories without any mask annotations for those categories.

What carries the argument

The load-bearing mechanism is a three-part pipeline built on FC-CLIP. DEC (Discrimination-Enhanced Classifier) selects the top-300 text-embedding channels that minimize inter-class similarity and maximize cross-class variance, sharpening the zero-shot classifier. KMA (Knowledge-Maintained Adaptation) freezes the 160 backbone channels most aligned with semantic content and trains only the remaining 32 channels, so the pretrained vision-language alignment survives while features adapt to bird's-eye imagery. PIP (Prior-Injected Prediction) builds a cache bank of CLIP visual prototypes from training instances and top-1 test predictions, then adds a cosine-similarity-based logit to the text classifier output through a weighted sum.

What would settle it

Re-run the GZSRI evaluation on iSAID with the PIP cache bank built only from seen-class training prototypes and no test-image pseudo-labels for unseen classes; if the unseen-class mAP falls back toward the 4.91 FC-CLIP baseline, the reported 9.30 result would be attributable to self-labeling on the test set rather than to the domain adaptation methods.

Watch

Extended reading notes

Core claim

The central claim is that zero-shot remote sensing instance segmentation can be made effective by adapting a frozen CLIP model to the aerial domain instead of using it directly. On iSAID and NWPU-VHR-10, under both the zero-shot and generalized zero-shot protocols, ZoRI reports higher unseen-class mAP and harmonic-mean mAP than existing zero-shot segmentation methods including FC-CLIP, ZSI, and D2Zero. The authors attribute the gain to three components: a discrimination-enhanced classifier that keeps only the most class-discriminative channels of the text embeddings, a knowledge-maintained adaptation that freezes semantic-related vision channels while training the rest to capture aerial cues, and a prior-injected prediction that merges text-based scores with a cache bank of visual prototypes from seen and pseudo-labeled unseen instances.

Load-bearing premise

The performance gain rests on using the model's own top-1 predictions on the test images to build the cache bank for unseen classes, which assumes this counts as legitimate zero-shot evaluation rather than test-set access.

Editorial extensions

If this is right

  • Unseen aerial classes such as swimming pools, helicopters, and tennis courts can be segmented without any mask annotations for those classes, using only text prompts and visual prototypes.
  • The channel-selection criterion used by DEC can be applied to any CLIP-based zero-shot classifier, not just instance segmentation, whenever classes are confusable.
  • Partially freezing the vision encoder is claimed to outperform both full fine-tuning and visual prompt tuning for adapting CLIP to a specialized domain while retaining zero-shot ability.
  • The cache bank injection shows that a small number of visual prototypes, here four per class, is enough to improve classification when text embeddings are too generic.
  • The reported state-of-the-art results establish the first benchmarks for zero-shot remote sensing instance segmentation, so future methods can be compared on the same splits and metrics.

Reading between the lines

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

  • Beyond the paper, the top-1 pseudo-labels used to build the unseen-class cache bank likely make the method transductive: the cache is created from the very test images being scored, so the reported gains over inductive baselines may shrink if the cache is built only from seen-class training prototypes.
  • A testable extension would be to replace the pseudo-labeled unseen prototypes with class-description prototypes generated from large language models, keeping the cache training-free while removing any reliance on test predictions.
  • The same channel-decoupling idea could be transferred to other vision-language backbones that are not Convolutional CLIP, such as transformer-based encoders, to check whether the preservation of semantic channels generalizes beyond the tested architecture.
  • The benchmark splits themselves are a contribution that invites scrutiny: the paper derives the iSAID split from a detection paper and clusters semantic embeddings for NWPU-VHR-10, so results across future methods may be sensitive to how that clustering was done.
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

3 major / 5 minor

Summary. The paper proposes ZoRI, a framework for zero-shot remote sensing instance segmentation. It formulates a new task and establishes benchmarks on iSAID, NWPU-VHR-10, and SIOR. The method combines three components: a Discrimination-Enhanced Classifier (DEC) that selects discriminative text-embedding channels, a Knowledge-Maintained Adaptation (KMA) that partially fine-tunes the CLIP vision backbone while freezing semantic-related channels, and a Prior-Injected Prediction (PIP) that uses a cache bank of visual prototypes to re-weight the CLIP zero-shot classifier. The authors report state-of-the-art results over FC-CLIP and older zero-shot instance segmentation methods under both ZSRI and GZSRI settings, with code and benchmarks released.

Significance. If the evaluations were conducted under a standard inductive zero-shot protocol, the paper would be a meaningful contribution: it is the first to formulate zero-shot instance segmentation for remote sensing, it releases code and benchmarks, and the DEC and KMA components alone yield substantial gains over the FC-CLIP baseline (unseen mAP 8.64 vs 4.91 on iSAID in the GZSRI setting). However, the headline claim is compromised because the PIP module constructs its unseen-class cache bank from the model's own top-1 predictions on the test set and then uses that bank to re-score the same test segments. This is a transductive self-training loop, and the paper does not disclose it or compare with transductive baselines. The feature-space mismatch between the cache keys and the query adds further uncertainty about the source of the reported gains. As evaluated, the contribution does not support the zero-shot claim made in the abstract and title.

major comments (3)
  1. [Prior-Injected Prediction (PIP), Supplementary B (Testing), Eqs. (5)-(7), Table 1] The unseen-class cache bank is built from the model's own top-1 predictions on the test set. Supplementary B states: 'For unseen classes, we use predictions with the highest probability after incorporating DEC and KMA to get visual samples' and 'we only use the top-1 probability prediction as the pseudo visual sample for unseen classes.' These pseudo-labeled features are concatenated into the cache bank F in Eq. (4), and Eqs. (5)-(7) then use the mask-pooled feature of the same test segment as the query f to compute logits_pip, which is added to the final score. The exact test images therefore contribute their own predicted labels to the classifier that scores them. This is a transductive self-training loop, not inductive zero-shot inference, and the paper neither discloses it as transductive nor compares with transductive baselines such as test-time adaptation or transductive ZSL methods. The component analysis in Table 1 shows that the PIP increment (iSAID unseen mAP 8.64 to 9.30; NWPU 9.32 to 12.26) is precisely the quantity built from test predictions, so the reported margin over inductive baselines such as FC-CLIP (Table 5) is not a fair zero-shot comparison. I recommend either removing the pseudo-labeled unseen cache bank or explicitly reframing the method as transductive zero-shot and evaluating against transductive baselines.
  2. [Cache Bank Construction, Eq. (4) vs Eq. (5), Supplementary B (Testing)] There is an apparent feature-space mismatch in the cache bank prediction. The cache bank keys F in Eq. (4) are 'obtained using the pretrained CLIP image encoder Evis', while the query f in Eq. (5) is the 'mask-pooled CLIP-extracted backbone feature' from the KMA-adapted backbone used at inference. If Evis is the original frozen CLIP image encoder and the query comes from the partially fine-tuned backbone, the cosine similarity in Eq. (5) compares features from two different encoders, with the KMA trainable/frozen channel split making the two feature spaces differ along at least the 32 trainable channels. The paper should clarify which encoder is used for the unseen-class pseudo samples and for the query, and should justify that the cross-encoder similarity is meaningful; otherwise the PIP gains in Tables 1 and 5 may be an artifact of the encoder mismatch.
  3. [Experimental Setup and Implementation Details, Tables 2 and 4] Hyperparameters for the final model, including the DEC channel count (300), cache size (4), and α (0.5), are selected using the same iSAID split on which the final results are reported (Tables 2 and 4; Implementation Details). No held-out validation split is described for this selection. Because the reported SOTA numbers are the result of test-set-based model selection, the comparison with baselines is not a clean evaluation of the method's inductive performance. The authors should either use a separate validation split for hyperparameter selection or clearly disclose that the test set was used for model selection.
minor comments (5)
  1. [Abstract and Introduction] 'state-of-art' should be 'state-of-the-art' in the abstract and in the contribution bullet.
  2. [Figure 2 caption and Method (KMA)] 'ground-level prospective' should be 'ground-level perspective' in both places this phrase appears.
  3. [Related Work] 'untilized' should be 'utilized' in the paragraph on zero-shot learning in remote sensing.
  4. [Method (DEC)] 'discirminative power' is a typo for 'discriminative power'.
  5. [Figures 6 and 7 captions] 'habor' should be 'harbor' in the captions of Figures 6 and 7.

Circularity Check

1 steps flagged · score 6.0 of 10

PIP's unseen-class cache bank is built from the model's own top-1 test-set predictions and then re-scores the same test segments, making the zero-shot SOTA comparison transductive rather than inductive.

  1. other [Method, Prior-Injected Prediction (PIP), Cache Bank Construction (Eqs. 4-7); Supplementary B, Testing]
    "For N u unseen classes whose ground truth annotations are assumed unavailable, we use P predictions with the highest probability as visual samples for unseen categories. Then, visual embeddings are also retrieved using the pretrained CLIP image encoder and concatenated to the cache bank, which is updated to obtain F ∈ R(KN s+P Nu)×D."

    The unseen-class entries of cache bank F are produced by the model's own test-set predictions (Supplementary B: 'we only use the top-1 probability prediction as the pseudo visual sample for unseen classes'). Equations (5)-(7) then take the mask-pooled feature f of a test segment as query, compute logits_cb = softmax(fF^T)L, and form logits_pip = fW^T + alpha*logits_cb. Thus the final score for a test instance includes a term built from features of test images labeled by the model itself. This is a transductive self-training loop, not an inductive zero-shot prediction; the PIP increment (iSAID unseen mAP 8.64->9.30; NWPU 9.32->12.26) is produced by test-set feedback and is not comparable to inductive baselines under the same protocol.

full rationale

ZoRI's DEC and KMA components are self-contained: DEC selects text channels via Eqs. (1)-(3) using only class-name embeddings, and KMA freezes semantic channels chosen from seen-class training features; neither uses test labels. The circularity is confined to PIP. The main text extends the cache bank with features of unseen-class instances selected by 'P predictions with the highest probability' on the test set, and Supplementary B confirms these are the model's own top-1 predictions. Equations (5)-(7) then use F as keys and the same test segment's feature as query, adding logits_cb to the CLIP score, so the final unseen-class score is partly computed from test-set features labeled by the model being evaluated. This is not a tautology (the loop can hurt recall, as the paper's own tables show), but it invalidates the claimed same-protocol SOTA comparison against inductive baselines such as FC-CLIP, ZSI, and D2Zero. Importantly, the component table shows DEC+KMA alone already improve iSAID unseen mAP from 4.91 to 8.64 and NWPU from 6.86 to 9.32, so the core contributions retain independent content; only the PIP increment is produced by the test-feedback loop. No load-bearing self-citation or imported uniqueness argument appears in the paper.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The free parameters are all hand-set or tuned via ablation, and the axioms indicate that the method rests on several heuristic assumptions about the usefulness of the channel-selection criterion, the compatibility of frozen and adapted feature spaces, and the accuracy of test-time pseudo-labels. There are no invented physical entities; the cache bank is a data structure, not a new entity.

free parameters (6)
  • lambda = 0.7
    Balance factor in Eq. (3) between similarity and variance for channel selection; set empirically in Implementation Details, no sensitivity analysis shown.
  • Number of DEC channels k = 300
    Number of top channels selected from 768-dim text embeddings; tuned via ablation on iSAID (Table 2), then applied to all datasets.
  • alpha = 0.5
    Balance factor in Eq. (7) for combining CLIP zero-shot score with cache bank logits; set empirically in Implementation Details.
  • KMA trainable channels = 32
    Number of backbone channels left trainable out of 192; set empirically in Implementation Details, no ablation reported.
  • Cache size per seen class = 4
    Number of visual prototypes per seen class in the cache bank; tuned via ablation (Table 4) on iSAID.
  • Pseudo samples per unseen class = 1 (top-1)
    Number of test predictions used as pseudo visual samples for each unseen class; set in Supplementary B, Testing.
assumptions (4)
  • ad hoc to paper The channel-selection criterion in Eq. (3) identifies the most discriminative channels: low inter-class similarity and high variance are assumed to be good proxies for discriminative power.
    No derivation is provided that these two statistics select the best channels; the criterion is a heuristic validated only by ablation on one dataset.
  • ad hoc to paper The same criterion used for text embeddings in DEC also applies to visual backbone features in KMA, so channels with low similarity and high variance across classes are semantic-related and should be frozen.
    The paper assumes transferability of the heuristic from text space to visual feature space without independent evidence; the justification is post hoc from overall results.
  • domain assumption Cosine similarity between query features from the KMA-adapted backbone and cache-bank features from the frozen CLIP encoder is a meaningful measure of semantic similarity.
    The paper does not address the feature-space mismatch; it simply uses L2-normalized cosine similarity in Eq. (5). If the two encoders produce incompatible spaces, the cache bank logits become unreliable.
  • domain assumption The model's own top-1 predictions on test images for unseen classes are accurate enough to serve as pseudo-labels for cache-bank construction.
    This is the transductive self-training assumption; no validation of pseudo-label accuracy is provided, and the paper itself notes 'numerous inaccurate predictions' before resorting to top-1 only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ZoRI: Towards Discriminative Zero-Shot Remote Sensing Instance Segmentation." pith.science (2026). https://pith.science/paper/DMOAOZHM

@misc{pith2026241212798,
  author       = {Pith},
  title        = {Pith review of: ZoRI: Towards Discriminative Zero-Shot Remote Sensing Instance Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DMOAOZHM}},
  note         = {Machine review of arXiv:2412.12798}
}
abstract

Instance segmentation algorithms in remote sensing are typically based on conventional methods, limiting their application to seen scenarios and closed-set predictions. In this work, we propose a novel task called zero-shot remote sensing instance segmentation, aimed at identifying aerial objects that are absent from training data. Challenges arise when classifying aerial categories with high inter-class similarity and intra-class variance. Besides, the domain gap between vision-language models' pretraining datasets and remote sensing datasets hinders the zero-shot capabilities of the pretrained model when it is directly applied to remote sensing images. To address these challenges, we propose a $\textbf{Z}$ero-Sh$\textbf{o}$t $\textbf{R}$emote Sensing $\textbf{I}$nstance Segmentation framework, dubbed $\textbf{ZoRI}$. Our approach features a discrimination-enhanced classifier that uses refined textual embeddings to increase the awareness of class disparities. Instead of direct fine-tuning, we propose a knowledge-maintained adaptation strategy that decouples semantic-related information to preserve the pretrained vision-language alignment while adjusting features to capture remote sensing domain-specific visual cues. Additionally, we introduce a prior-injected prediction with cache bank of aerial visual prototypes to supplement the semantic richness of text embeddings and seamlessly integrate aerial representations, adapting to the remote sensing domain. We establish new experimental protocols and benchmarks, and extensive experiments convincingly demonstrate that ZoRI achieves the state-of-art performance on the zero-shot remote sensing instance segmentation task. Our code is available at https://github.com/HuangShiqi128/ZoRI.

Figures

Figures reproduced from arXiv: 2412.12798 by the authors.

Figure 1
Figure 1. Illustration of zero-shot remote sensing instance [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) After refinement, the head of plane is highlighted and the activation map strictly follows its shape; For tennis court, the activation map is more focused and the missed one in the middle using original channels is also emphasized. (b) Classes such as basketball court and tennis court share similar color and shape, whereas instances from ship can have various appearances. (c) Remote sensing images are in bird’s … view at source ↗
Figure 3
Figure 3. We adopt the paradigm of FC-CLIP (Yu et al. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Overview of ZoRI framework. The CLIP image encoder is partially trained with knowledge-maintained adaptation (KMA) to extract backbone features, which are then fed into a mask generator to produce mask predictions and class embeddings. Discriminative-enhanced classifie…
Figure 4
Figure 4. Figure 4: Comparison of GZSRI results: (top row) ground truth, (middle row) FC-CLIP (Yu et al. 2023) and (bottom row) our [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: (Best viewed in color) t-SNE visualization of [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Comparison of GZSRI results on NWPU-VHR-10 dataset. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Comparison of GZSRI results on iSAID dataset. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 49 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Carvalho, O. L. F. d.; de Carvalho Junior, O. A.; Albuquerque, A. O. d.; Bem, P. P. d.; Silva, C. R.; Ferreira, P. H. G.; Moura, R. d. S. d.; Gomes, R. A. T.; Guimaraes, R. F.; and Borges, D. L. 2020. Instance segmentation for large, multi-channel remote sensing imagery using mask-RCNN and a mosaicking approach. Remote Sensing

  4. [4]

    Chen, K.; Liu, C.; Chen, H.; Zhang, H.; Li, W.; Zou, Z.; and Shi, Z. 2024. RSPrompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model. TGRS

  5. [5]

    Chen, Y.; Li, Y.; Mao, H.; Liu, G.; Chai, X.; and Jiao, L. 2023. A Novel Discriminative Enhancement Method for Few-Shot Remote Sensing Image Scene Classification. Remote Sensing

  6. [6]

    G.; Kirillov, A.; and Girdhar, R

    Cheng, B.; Misra, I.; Schwing, A. G.; Kirillov, A.; and Girdhar, R. 2022. Masked-attention Mask Transformer for Universal Image Segmentation. In CVPR

  7. [7]

    Cheng, G.; Han, J.; and Lu, X. 2017. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE

  8. [8]

    Cheng, G.; Han, J.; Zhou, P.; and Guo, L. 2014. Multi-class geospatial object detection and geographic image classification based on collection of part detectors. ISPRS Journal of Photogrammetry and Remote Sensing

Show all 61 references
  1. [9]

    Cheng, G.; Si, Y.; Hong, H.; Yao, X.; and Guo, L. 2021. Cross-Scale Feature Fusion for Object Detection in Optical Remote Sensing Images. GRSL

  2. [10]

    Cheng, G.; Xie, X.; Han, J.; Li, K.; and Xia, G.-S. 2020. Remote Sensing Image Scene Classification Meets Deep Learning: Challenges, Methods, Benchmarks, and Opportunities

  3. [11]

    O.; de Carvalho, O

    de Albuquerque, A. O.; de Carvalho, O. L. F.; e Silva, C. R.; de Bem, P. P.; Gomes, R. A. T.; Borges, D. L.; Guimar \ a es, R. F.; Pimentel, C. M. M.; and de Carvalho J \'u nior, O. A. 2021. Instance segmentation of center pivot irrigation systems using multi-temporal SENTINEL...

  4. [12]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805

  5. [13]

    Ding, J.; Xue, N.; Xia, G.-S.; and Dai, D. 2022. Decoupling Zero-Shot Semantic Segmentation. In CVPR

  6. [14]

    He, S.; Ding, H.; and Jiang, W. 2023 a . Primitive Generation and Semantic-related Alignment for Universal Zero-Shot Segmentation. In CVPR

  7. [15]

    He, S.; Ding, H.; and Jiang, W. 2023 b . Semantic-Promoted Debiasing and Background Disambiguation for Zero-Shot Instance Segmentation. In CVPR

  8. [16]

    He, S.; Ding, H.; Jiang, X.; and Wen, B. 2024. SegPoint: Segment Any Point Cloud via Large Language Model. In ECCV

  9. [17]

    Huang, P.; Han, J.; Cheng, D.; and Zhang, D. 2022. Robust Region Feature Synthesizer for Zero-Shot Object Detection. arXiv preprint arXiv:2201.00103

  10. [18]

    Ilharco, G.; Wortsman, M.; Wightman, R.; Gordon, C.; Carlini, N.; Taori, R.; Dave, A.; Shankar, V.; Namkoong, H.; Miller, J.; Hajishirzi, H.; Farhadi, A.; and Schmidt, L. 2021. OpenCLIP

  11. [19]

    Jia, M.; Tang, L.; Chen, B.-C.; Cardie, C.; Belongie, S.; Hariharan, B.; and Lim, S.-N. 2022. Visual Prompt Tuning. In ECCV

  12. [20]

    C.; Lo, W.-Y.; Doll \'a r, P.; and Girshick, R

    Kirillov, A.; Mintun, E.; Ravi, N.; Mao, H.; Rolland, C.; Gustafson, L.; Xiao, T.; Whitehead, S.; Berg, A. C.; Lo, W.-Y.; Doll \'a r, P.; and Girshick, R. 2023. Segment Anything. arXiv:2304.02643

  13. [21]

    Li, A.; Lu, Z.; Wang, L.; Xiang, T.; and Wen, J.-R. 2017. Zero-shot scene classification for high spatial resolution remote sensing images. TGRS

  14. [22]

    Li, K.; Wan, G.; Cheng, G.; Meng, L.; and Han, J. 2020. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing

  15. [23]

    Li, X.; Wen, C.; Hu, Y.; and Zhou, N. 2023. RS-CLIP: Zero shot remote sensing scene classification via contrastive vision-language supervision. International Journal of Applied Earth Observation and Geoinformation

  16. [24]

    Li, Y.; Zhu, Z.; Yu, J.-G.; and Zhang, Y. 2021. Learning deep cross-modal embedding networks for zero-shot remote sensing image scene classification. TGRS

  17. [25]

    Liang, F.; Wu, B.; Dai, X.; Li, K.; Zhao, Y.; Zhang, H.; Zhang, P.; Vajda, P.; and Marculescu, D. 2023. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 7061--7070

  18. [26]

    Liu, F.; Chen, D.; Guan, Z.; Zhou, X.; Zhu, J.; Ye, Q.; Fu, L.; and Zhou, J. 2024 a . RemoteCLIP: A Vision Language Foundation Model for Remote Sensing. TGRS

  19. [27]

    Liu, X.; and Di, X. 2022. Global Context Parallel Attention for Anchor-Free Instance Segmentation in Remote Sensing Images. GRSL

  20. [28]

    Liu, Y.; Chen, D.; Ma, A.; Zhong, Y.; Fang, F.; and Xu, K. 2020. Multiscale U-Shaped CNN Building Instance Extraction Framework With Edge Constraint for High-Spatial-Resolution Remote Sensing Imagery. TGRS

  21. [29]

    Liu, Y.; Li, H.; Hu, C.; Luo, S.; Luo, Y.; and Chen, C. W. 2024 b . Learning to aggregate multi-scale context for instance segmentation in remote sensing images. IEEE Transactions on Neural Networks and Learning Systems

  22. [30]

    Liu, Z.; Mao, H.; Wu, C.-Y.; Feichtenhofer, C.; Darrell, T.; and Xie, S. 2022. A ConvNet for the 2020s. arXiv:2201.03545

  23. [31]

    P.; Liu, M

    Mall, U.; Phoo, C. P.; Liu, M. K.; Vondrick, C.; Hariharan, B.; and Bala, K. 2023. Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment

  24. [32]

    Mikolov, T.; Chen, K.; Corrado, G.; and Dean, J. 2013. Efficient Estimation of Word Representations in Vector Space. arXiv:1301.3781

  25. [33]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I. 2021. Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020

  26. [34]

    Rong, J.; Chen, H.; Chen, T.; Ou, L.; Yu, X.; and Liu, Y. 2023. Retrieval-enhanced visual prompt learning for few-shot classification. arXiv preprint arXiv:2306.02243

  27. [35]

    Su, H.; Huang, P.; Yin, J.; and Zhang, X. 2022. Faster and better instance segmentation for large scene remote sensing imagery. In IGARSS. IEEE

  28. [36]

    Su, H.; Wei, S.; Liu, S.; Liang, J.; Wang, C.; Shi, J.; and Zhang, X. 2020. HQ-ISNet: High-quality instance segmentation for remote sensing imagery. Remote Sensing

  29. [37]

    Su, H.; Wei, S.; Yan, M.; Wang, C.; Shi, J.; and Zhang, X. 2019. Object detection and instance segmentation in remote sensing imagery based on precise mask R-CNN. In IGARSS. IEEE

  30. [38]

    Van der Maaten, L.; and Hinton, G. 2008. Visualizing data using t-SNE. JMLR, 9(11)

  31. [39]

    Wang, C.; Peng, G.; and De Baets, B. 2021. A distance-constrained semantic autoencoder for zero-shot remote sensing scene classification. IEEE J-STARS

  32. [40]

    Wang, D.; Zhang, J.; Du, B.; Xu, M.; Liu, L.; Tao, D.; and Zhang, L. 2023. SAMRS: Scaling-up Remote Sensing Segmentation Dataset with Segment Anything Model. In NeurIPS

  33. [41]

    Wei, S.; Zeng, X.; Zhang, H.; Zhou, Z.; Shi, J.; and Zhang, X. 2022. LFG-Net: Low-level Feature Guided Network for Precise Ship Instance Segmentation in SAR Images. TGRS

  34. [42]

    Xia, G.-S.; Bai, X.; Ding, J.; Zhu, Z.; Belongie, S.; Luo, J.; Datcu, M.; Pelillo, M.; and Zhang, L. 2018. DOTA: A Large-Scale Dataset for Object Detection in Aerial Images. In CVPR

  35. [43]

    H.; Schiele, B.; and Akata, Z

    Xian, Y.; Lampert, C. H.; Schiele, B.; and Akata, Z. 2020. Zero-Shot Learning -- A Comprehensive Evaluation of the Good, the Bad and the Ugly. arXiv:1707.00600

  36. [44]

    Xu, M.; Zhang, Z.; Wei, F.; Hu, H.; and Bai, X. 2023. Side Adapter Network for Open-Vocabulary Semantic Segmentation. In CVPR

  37. [45]

    Xu, M.; Zhang, Z.; Wei, F.; Lin, Y.; Cao, Y.; Hu, H.; and Bai, X. 2022. A Simple Baseline for Open Vocabulary Semantic Segmentation with Pre-trained Vision-language Model. In ECCV

  38. [46]

    Xu, X.; Feng, Z.; Cao, C.; Li, M.; Wu, J.; Wu, Z.; Shang, Y.; and Ye, S. 2021. An improved swin transformer-based model for remote sensing object detection and instance segmentation. Remote Sensing

  39. [47]

    Yang, X.; Yan, J.; Liao, W.; Yang, X.; Tang, J.; and He, T. 2022. Scrdet++: Detecting small, cluttered and rotated objects via instance-level feature denoising and rotation loss smoothing. IEEE TPAMI

  40. [48]

    S.; Isiacik Colak, A

    Yasir, M.; Zhan, L.; Liu, S.; Wan, J.; Hossain, M. S.; Isiacik Colak, A. T.; Liu, M.; Islam, Q. U.; Raza Mehdi, S.; and Yang, Q. 2023. Instance segmentation ship detection based on improved Yolov7 using complex background SAR images. Frontiers in Marine Science

  41. [49]

    Ye, W.; Zhang, W.; Lei, W.; Zhang, W.; Chen, X.; and Wang, Y. 2023. Remote sensing image instance segmentation network with transformer and multi-scale feature representation. Expert Systems with applications

  42. [50]

    Yu, Q.; He, J.; Deng, X.; Shen, X.; and Chen, L.-C. 2023. Convolutions Die Hard: Open-Vocabulary Segmentation with Single Frozen Convolutional CLIP. In NeurIPS

  43. [51]

    W.; Arora, A.; Gupta, A.; Khan, S.; Sun, G.; Khan, F

    Zamir, S. W.; Arora, A.; Gupta, A.; Khan, S.; Sun, G.; Khan, F. S.; Zhu, F.; Shao, L.; Xia, G.-S.; and Bai, X. 2019. iSAID: A Large-scale Dataset for Instance Segmentation in Aerial Images. arXiv:1905.12886

  44. [52]

    Zang, Z.; Lin, C.; Tang, C.; Wang, T.; and Lv, J. 2024. Zero-Shot Aerial Object Detection with Visual Description Regularization. arXiv:2402.18233

  45. [53]

    Zavras, A.; Michail, D.; Demir, B.; and Papoutsis, I. 2024. Mind the Modality Gap: Towards a Remote Sensing Vision-Language Model via Cross-modal Alignment

  46. [54]

    Zhang, R.; Fang, R.; Gao, P.; Zhang, W.; Li, K.; Dai, J.; Qiao, Y.; and Li, H. 2021 a . Tip-Adapter: Training-free CLIP-Adapter for Better Vision-Language Modeling. arXiv preprint arXiv:2111.03930

  47. [55]

    Zhang, R.; Hu, X.; Li, B.; Huang, S.; Deng, H.; Li, H.; Qiao, Y.; and Gao, P. 2023. Prompt, Generate, then Cache: Cascade of Foundation Models makes Strong Few-shot Learners. arXiv preprint arXiv:2303.02151

  48. [56]

    Zhang, T.; Zhang, X.; Zhu, P.; Tang, X.; Li, C.; Jiao, L.; and Zhou, H. 2021 b . Semantic attention and scale complementary network for instance segmentation in remote sensing images. IEEE Transactions on Cybernetics

  49. [57]

    Zhang, X.; Han, L.; Han, L.; and Zhu, L. 2020. How well do deep learning-based methods for land cover classification and object detection perform on high resolution remote sensing imagery? Remote Sensing

  50. [58]

    Zheng, Y.; Wu, J.; Qin, Y.; Zhang, F.; and Cui, L. 2021. Zero-Shot Instance Segmentation. arXiv:2104.06601

  51. [59]

    C.; and Dai, B

    Zhou, C.; Loy, C. C.; and Dai, B. 2022. Extract Free Dense Labels from CLIP. In ECCV

  52. [60]

    Zhou, Z.; Lei, Y.; Zhang, B.; Liu, L.; and Liu, Y. 2023. ZegCLIP: Towards adapting CLIP for zero-shot semantic segmentation. In CVPR

  53. [61]

    Zhu, X.; Zhang, R.; He, B.; Zhou, A.; Wang, D.; Zhao, B.; and Gao, P. 2023. Not All Features Matter: Enhancing Few-shot CLIP with Adaptive Prior Refinement. arXiv:2304.01195

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.