Pith. sign in

REVIEW 1 major objections 5 minor 52 references

GeoSelect shows that executing a typed spatial program over candidate boxes more than doubles prior training-free accuracy on aerial referring segmentation.

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 ↗

A training-free pipeline synthesises referring expressions into a typed geometric DSL, executes them over scored candidate boxes, and reaches 58.86 mIoU on RRSIS-D—over twice the previous training-free best.

T0 review reviewed 2026-07-11 challenge →

load-bearing objection Clean training-free RRSIS result: typed spatial programs over a scored candidate set roughly double the prior ceiling and the gain is isolated to the executor, not the backbone. the 1 major comments →

arxiv 2607.03869 v1 pith:6QRDYSQK submitted 2026-07-04 cs.CV cs.AI

GeoSelect: Spatial-Program Execution for Training-Free Referring Remote Sensing Image Segmentation

Yuhang Jiang , Guohui Deng , Miaozhong Xu , Chao Ruan , Jinling Zhao , Linsheng Huang This is my paper
classification cs.CV cs.AI
keywords referring image segmentationremote sensingtraining-freeneuro-symbolic reasoningprogram synthesisvision-language modelsspatial reasoning
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 reading

Aerial referring expressions often name one object among many near-identical ones by spatial, comparative, or ordinal cues such as the largest ship or the second court from the left. Existing training-free systems resolve these phrases only through implicit vision-language similarity, which cannot execute those relations. GeoSelect instead has a frozen text-only language model compile the expression into a small typed program, checks that program for well-formedness, and runs a deterministic executor over detector proposals. Continuous geometric fields handle position and proximity; discrete set and order operators add extrema, ranks, counted unions, and binary relations that fields alone cannot express. A reliability ladder falls back to a field-only special case whenever synthesis or execution fails, so every expression still returns an answer. With no referring supervision and on a single GPU, the method reaches 58.86 mIoU on RRSIS-D and 55.27 mIoU on RISBench—more than twice and about 1.7 times the best prior training-free results—and a controlled comparison with candidates and segmenter fixed attributes the gain to explicit execution itself.

Core claim

Training-free referring remote sensing image segmentation is solved by reframing the task as explicit execution of a typed spatial program: a frozen text-only language model synthesises the expression into a small domain-specific language, a well-formedness checker accepts it, and a deterministic executor runs continuous geometric fields and discrete set/order operators over a single scored-candidate-set type. This yields 58.86 mIoU on RRSIS-D test and 55.27 mIoU on RISBench under one frozen configuration, more than twice and about 1.7 times the best prior training-free methods, with the improvement shown to come from explicit geometric execution rather than the backbone.

What carries the argument

The scored candidate set: one closed type of boxes each carrying a non-negative score, under which every operator composes. Continuous geometric fields re-weight scores by position and proximity; discrete operators (extremum, ordinal, counted-union, binary relation) add constructions fields alone cannot express. A reliability ladder degrades any failing program to a field-only special case so an answer is always returned.

Load-bearing premise

The method assumes an aerial-pretrained open-vocabulary detector already proposes a box that overlaps the true object for most expressions; if that recall collapses, selection has nothing correct to choose.

What would settle it

Hold the same candidate boxes and segmenter fixed, replace the typed executor with the best implicit region-text or confidence selector, and check whether the large mIoU gain on spatial and compositional expressions disappears; if explicit execution no longer wins under that control, the central attribution fails.

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

If this is right

  • Training-free aerial referring can reach roughly 93 percent of a strong supervised specialist without referring masks or per-benchmark training.
  • Explicit spatial-program execution, not region-text similarity, is the main lever for resolving spatial and ordinal expressions under same-class clutter.
  • Residual error localises to detector recall on RRSIS-D and to program selection on RISBench, so further gains target those stages separately.
  • A single frozen configuration transfers across benchmarks with no test-set recalibration.
  • Every intermediate program, field, and ranking is inspectable, enabling stage-level failure analysis.

Where Pith is reading between the lines

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

  • The same typed executor can be swapped onto other open-vocabulary detectors or promptable segmenters without re-training, making aerial referring modular.
  • If synthesis quality is the RISBench bottleneck, lightweight semantic checks of the program against the candidate set may close more of the oracle gap than simply scaling the parser.
  • Oriented-box geometry for elongated objects such as ships and bridges is a direct next step the paper flags; it would likely lift superlative accuracy where axis-aligned boxes fail.
  • The reliability-ladder pattern—expressive program with a guaranteed field-only floor—may generalise to other training-free neuro-symbolic vision tasks that currently fail silently on malformed plans.
Share X Bluesky LinkedIn Reddit HN

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

1 major / 5 minor

Summary. GeoSelect reframes training-free referring remote sensing image segmentation as the execution of a typed spatial program. A frozen text-only LLM synthesises a natural-language expression into a small DSL; a well-formedness checker and reliability ladder accept or degrade the program; and a deterministic executor evaluates continuous geometric fields and discrete set/order operators over a single scored-candidate-set type. With no referring supervision, the method reports 58.86 mIoU on RRSIS-D test and 55.27 mIoU on RISBench under one frozen configuration, more than twice the best prior training-free result on RRSIS-D. A controlled selector ablation with candidates and segmenter fixed attributes the gain to explicit execution; an oracle decomposition localises residual error to detection recall on RRSIS-D and selection on RISBench; and an image-id exposure audit addresses detector pretraining overlap.

Significance. If the results hold, the paper substantially raises the training-free ceiling on aerial referring segmentation and supplies a clean experimental template for attributing gains to the selection mechanism rather than the backbone. Strengths include the same-candidate selector ablation (Table III, +12.06 mIoU over the best implicit selector), per-stratum operator gains with bootstrap CIs (Table IV), the oracle decomposition that separates detection from selection (Table V), the reliability ladder that keeps every expression answerable, and the image-level exposure audit (Table VIII). The method is inspectable at every intermediate stage, runs on a single GPU, and transfers across two public benchmarks without test-set recalibration. These are concrete, falsifiable contributions to neuro-symbolic remote-sensing vision.

major comments (1)
  1. The central attribution claim is well supported by the controlled selector ablation (Table III) and does not require further major revision. The load-bearing precondition is detector recall: the oracle analysis (Table V) shows a high-IoU box is already present for 86.6% of RRSIS-D and 81.8% of RISBench expressions. The paper states this clearly and measures it; once granted, the +12.06 mIoU gain over implicit selectors is cleanly isolated. No additional load-bearing inconsistency appears in the derivation or the experimental design.
minor comments (5)
  1. §V-C / Table IV: the ordinal stratum has only n=52 on RISBench and the CI is wide ([-17.3, +8.7]). The text already cautions against over-interpretation; a single sentence in the abstract or conclusion that the ordinal claim is coverage rather than measured accuracy would prevent misreading.
  2. §V-I: the exposure audit is performed only on RRSIS-D. The manuscript correctly declines a leakage-robust claim for RISBench because of re-encoded crops without source IDs; stating this limitation once more in the conclusion would make the scope fully explicit.
  3. §VI: the orientation sensitivity of axis-aligned boxes for elongated objects (ships, bridges) is quantified (IoU 57.9 vs 23.9). Elevating this to a short dedicated paragraph or a small table would help readers who work with oriented aerial objects.
  4. Fig. 3 grammar and Algorithm 1: a one-line note that SELECT is the only Phrase o C leaf and that every other operator is C o C would make the type closure immediately visible without reading the surrounding prose.
  5. Typos / consistency: abstract and body both use “more than twice” for RRSIS-D; keep the exact factor (2.08×) or the rounded phrase consistently. A few arXiv venue labels (e.g., “arXiv’26”) will need updating once the cited works appear in print.

Circularity Check

0 steps flagged

No significant circularity: headline gains are measured on external public benchmarks under a frozen configuration, with the controlled selector ablation isolating execution from the shared backbone.

full rationale

GeoSelect's central claim is an empirical attribution: with candidates and segmenter held fixed, explicit typed spatial-program execution outperforms implicit selectors by +12.06 mIoU on RRSIS-D val and yields 58.86 / 55.27 mIoU on the two public test sets. Design choices are locked on the validation split before any test evaluation; field-shape parameters are analytically inert under centre scoring; the few free parameters (detection threshold, SAHI tile) sit on flat plateaus and are not re-selected on test. The few-shot synthesis schema is programmatically verified disjoint from val/test expressions. The reliability ladder and field-only special case are deliberate fallbacks, not fitted predictions. The only mild self-referential element is the authors' own field-only baseline, which is an ablation of the same system rather than a load-bearing external uniqueness claim. No equation reduces reported mIoU to a fitted constant by construction, and no uniqueness theorem is imported from prior author work. Score 1 reflects ordinary self-citation of the method's own special case, not circular derivation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 3 invented entities

The central empirical claim rests on a small set of frozen public models, a handful of hand-chosen but flat hyperparameters, and the domain assumption that an aerial-pretrained open-vocabulary detector supplies usable candidates. No new physical entities are postulated; the DSL and scored-candidate-set type are engineering abstractions whose behaviour is fully defined by the paper’s equations.

free parameters (4)
  • detection threshold τ_det = 0.15
    Set to 0.15 after a val sweep that shows a flat plateau (0.10–0.25); frozen before test evaluation.
  • SAHI tile size = 768
    Chosen as 768 px on val; 512 collapses recall while 1024 is equivalent.
  • directional exponent γ / centre width σ_c / proximity length L_A = γ=1, σ_c=0.25 min(H,W), L_A=1.5·mean side
    Analytically shown to be irrelevant under centre scoring (selection invariant to strictly increasing field reshaping); reported values are conventional defaults.
  • extremal softmax temperature β = 1/σ_u
    Set to inverse axis std-dev of live candidates; inert on RRSIS-D (almost no superlatives) and left un-tuned.
axioms (4)
  • domain assumption An aerial-pretrained open-vocabulary detector (LAE-DINO on LAE-1M) yields candidate boxes whose best-IoU already exceeds 0.5 for >80 % of expressions.
    Stated as the premise of the oracle analysis (§V-D) and confirmed by the exposure audit (§V-I); without it the selection stage has nothing to rank.
  • domain assumption A frozen text-only LLM can emit a well-formed program in the given DSL from language alone with high enough rate that the reliability ladder recovers the remainder.
    Measured program rates 83–91 % (§V-E); the ladder guarantees an answer even when the assumption fails.
  • standard math Centre sampling of geometric fields is a valid scoring rule for ranking boxes (selection invariant to monotone field reshaping).
    Derived in §IV-C and verified by exact invariance sweeps (§V-H).
  • ad hoc to paper The expression taxonomy derived solely from text is method-independent and partitions every system identically.
    Defined in §III-E; used only for analysis, not for training or selection.
invented entities (3)
  • scored candidate set type C = List[(box, score)] no independent evidence
    purpose: Single closed type under which continuous fields and discrete set/order operators compose freely.
    Defined by Eq. (2); every operator has signature C → C. Purely an engineering abstraction with no external ontological claim.
  • spatial-program DSL (SELECT / FILTER / ARGMAX / NTH / RELATE / …) no independent evidence
    purpose: Makes superlative, ordinal and compositional constructions executable rather than approximated by similarity.
    Grammar in Fig. 3; semantics given by the executor equations. Behaviour fully specified inside the paper.
  • reliability ladder (program o field-only fallback) no independent evidence
    purpose: Guarantees every expression returns an answer even when synthesis or execution fails.
    Algorithm 1 and §IV-H; measured program rates reported separately from accuracy.

reviewed 2026-07-11 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GeoSelect: Spatial-Program Execution for Training-Free Referring Remote Sensing Image Segmentation." pith.science (2026). https://pith.science/paper/6QRDYSQK

@misc{pith2026260703869,
  author       = {Pith},
  title        = {Pith review of: GeoSelect: Spatial-Program Execution for Training-Free Referring Remote Sensing Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6QRDYSQK}},
  note         = {Machine review of arXiv:2607.03869}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Referring remote sensing image segmentation isolates the object named by a natural-language expression in an aerial image. Existing training-free methods resolve the expression through implicit vision-language activations or region-text similarity, which gives weak control over the spatial, comparative, and ordinal relations that dominate aerial referring: they cannot represent constructions such as the largest ship or the second court from the left. We propose GeoSelect, a training-free pipeline that reframes referring as the execution of a typed spatial program. A frozen, text-only language model synthesises the expression into a small domain-specific language, a well-formedness checker accepts the program, and a deterministic executor runs it. The central abstraction is a single scored candidate set type under which every operator composes: continuous geometric fields realise position and proximity as dense pixel-level maps, while discrete set and order operators add the extremum, ordinal, counted-union, and relational constructions that fields alone cannot express. Because execution is explicit, every intermediate program, field, and ranking is inspectable, and a reliability ladder degrades any failing program to a field-only special case, so every expression returns an answer. GeoSelect attains 58.86 mIoU on RRSIS-D test and 55.27 mIoU on RISBench test, more than twice the best prior training-free method on RRSIS-D, with no referring supervision and on a single GPU. A controlled comparison with candidates and segmenter fixed attributes the gain to explicit execution, not the backbone; an oracle decomposition localises the residual gap to detection recall on RRSIS-D and selection on RISBench, and an exposure audit confirms robustness to pretraining leakage. Code will be released upon acceptance at the project page https://avalon-s.github.io/GeoSelect/.

Figures

Figures reproduced from arXiv: 2607.03869 by Chao Ruan, Guohui Deng, Jinling Zhao, Linsheng Huang, Miaozhong Xu, Yuhang Jiang.

Figure 1
Figure 1. Figure 1: Why explicit execution. For a spatial expression over several same-class objects, implicit region–text matching (left, GeoRSCLIP) selects the wrong instance, whereas GeoSelect synthesises and executes a typed spatial program that scores the candidate set (middle, brighter = higher score; the chosen box in green) and returns the instance matching the ground truth (right). Same candidate boxes and segmenter;… view at source ↗
Figure 2
Figure 2. Figure 2: GeoSelect executes a typed spatial program over a single scored candidate set, with a real per-operator trace above each stage (RISBench, “the westernmost storage tank in the line of tanks”). A frozen text-only LLM synthesises the program π; the executor evaluates it bottom-up: SELECT returns the detected candidates, the continuous FILTER (teal) re-weights them by a geometric field, and the discrete ARGMIN… view at source ↗
Figure 3
Figure 3. Figure 3: The spatial-program DSL. Every expr has type C→ C (scored candi￾date set to scored candidate set) except Select, which produces the initial set from the detector. Leaves are Select; Filter and Relate apply contin￾uous fields and binary relations; Argmax/Argmin/Nth/RestrictCount are the discrete set and order operators; And/Or/Not compose by field algebra. remain as design choices. For an image-relative pre… view at source ↗
Figure 4
Figure 4. Figure 4: The continuous geometric field G(x, y) of the FILTER operator. For image-relative predicates the field is computed in closed form against the image frame (Eqs. 4–5); a candidate is scored at its centre (Eq. 6). Input (left), field as a heat map (middle, warm = high), and the resulting prediction (right; cyan = mask, yellow = ground truth) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Pipeline traces on RRSIS-D test, one case per row across the columns input, candidate boxes, scored candidate set (brighter = higher score, chosen box in green), GeoSelect mask (cyan), and ground truth (yellow). Rows span the strata: (a) image-relative, (b, d) object-anchored/compositional, (c) attribute. Small targets are circled in red for visibility. The referring expression of each row is printed below… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison on RRSIS-D test. Columns: SAM3 [18] (full-expression prompt), RSVG-ZeroOV [16], RSRefSeg2 [14] (supervised), GeoSelect (ours), ground truth. Predictions in cyan, GT in yellow. The training-free baselines SAM3 and RSVG-ZeroOV over-segment multiple same-class regions or pick the wrong instance, whereas GeoSelect, also training-free, resolves the queried instance comparably to the super… view at source ↗
Figure 7
Figure 7. Figure 7: Pipeline traces on RISBench test (frozen RRSIS-D configuration), one case per row across input, candidate boxes, scored candidate set (brighter = higher score, chosen box in green), GeoSelect mask (cyan), and ground truth (yellow). Rows span (a) superlative, (b) compositional, (c) ordinal (“the second vehicle”), and (d) object-anchored proximity (a vehicle near a building). Small targets are circled in red… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on RISBench test. Columns: SAM3 [18] (full-expression prompt), RSVG-ZeroOV [16], RSRefSeg2 [14] (supervised), GeoSelect (ours), ground truth. Predictions in cyan, GT in yellow. The longer, more constrained RISBench expressions widen the gap to the implicit baselines. Red ellipses mark a small or method-distinguishing target region. The referring expression of each row is printed belo… view at source ↗
Figure 9
Figure 9. Figure 9: Failure modes by pipeline stage (cyan = prediction, yellow = ground truth; small targets circled in red). (a) Detection: the referent is too small to be proposed. (b) Selection: a near-perfect candidate exists but the monotone ramp picks the wrong instance. (c, d) Synthesis: the correct box is present but the synthesised program is wrong. Per-case expressions and IoUs are in Section V-K [PITH_FULL_IMAGE:f… view at source ↗
Figure 10
Figure 10. Figure 10: Failure causes by pipeline stage on the two test sets. Each expression is Success (IoU ≥ 0.5) or attributed to the earliest failing stage: Detection (no candidate reaches IoU 0.5), Selection (the field-only fallback picks the wrong existing box), or Synthesis (an executed program picks the wrong existing box). Per-stage shares and their cross-dataset shift are read in Section VI. of that, the discrete ope… 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

52 extracted references · 10 linked inside Pith

  1. [1]

    Rotated multi-scale interaction network for referring remote sensing image seg- mentation,

    S. Liu, Y . Ma, X. Zhang, H. Wang, J. Ji, X. Sun, and R. Ji, “Rotated multi-scale interaction network for referring remote sensing image seg- mentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 26 658–26 668

    2024

  2. [2]

    Cross-modal bidirectional interaction model for referring remote sensing image segmentation,

    Z. Dong, Y . Sun, T. Liu, W. Zuo, and Y . Gu, “Cross-modal bidirectional interaction model for referring remote sensing image segmentation,” arXiv preprint arXiv:2410.08613, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Exploring fine-grained image-text alignment for referring remote sensing image segmentation,

    S. Lei, X. Xiao, T. Zhang, H.-C. Li, Z. Shi, and Q. Zhu, “Exploring fine-grained image-text alignment for referring remote sensing image segmentation,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  4. [4]

    Visual programming: Compositional visual reasoning without training,

    T. Gupta and A. Kembhavi, “Visual programming: Compositional visual reasoning without training,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 14 953– 14 962

    2023

  5. [5]

    ViperGPT: Visual inference via python execution for reasoning,

    D. Sur ´ıs, S. Menon, and C. V ondrick, “ViperGPT: Visual inference via python execution for reasoning,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 11 888– 11 898

    2023

  6. [6]

    Locate anything on earth: Advancing open-vocabulary object detection for remote sensing community,

    J. Pan, Y . Liu, Y . Fu, M. Ma, J. Li, D. P. Paudel, L. Van Gool, and X. Huang, “Locate anything on earth: Advancing open-vocabulary object detection for remote sensing community,” inProceedings of the AAAI Conference on Artificial Intelligence, 2025

    2025

  7. [7]

    CRIS: CLIP-driven referring image segmentation,

    Z. Wang, Y . Lu, Q. Li, X. Tao, Y . Guo, M. Gong, and T. Liu, “CRIS: CLIP-driven referring image segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 11 686–11 695

    2022

  8. [8]

    LA VT: Language-aware vision transformer for referring image segmentation,

    Z. Yang, J. Wang, Y . Tang, K. Chen, H. Zhao, and P. H. S. Torr, “LA VT: Language-aware vision transformer for referring image segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 18 155–18 165. PREPRINT 19

    2022

  9. [9]

    Referring image segmentation via recurrent refinement networks,

    R. Li, K. Li, Y .-C. Kuo, M. Shu, X. Qi, X. Shen, and J. Jia, “Referring image segmentation via recurrent refinement networks,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 5745–5753

    2018

  10. [10]

    Bi-directional relationship inferring network for referring image segmentation,

    Z. Hu, G. Feng, J. Sun, L. Zhang, and H. Lu, “Bi-directional relationship inferring network for referring image segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 4424–4433

    2020

  11. [11]

    Cross-modal pro- gressive comprehension for referring segmentation,

    S. Liu, T. Hui, S. Huang, Y . Wei, B. Li, and G. Li, “Cross-modal pro- gressive comprehension for referring segmentation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 9, pp. 4761– 4775, 2021

    2021

  12. [12]

    RRSIS: Referring remote sensing image segmentation,

    Z. Yuan, L. Mou, Y . Hua, and X. X. Zhu, “RRSIS: Referring remote sensing image segmentation,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  13. [13]

    Referring remote sensing image segmentation via bidirectional alignment guided joint prediction,

    T. Zhang, Z. Wen, B. Kong, K. Liu, Y . Zhang, P. Zhuang, and J. Li, “Referring remote sensing image segmentation via bidirectional alignment guided joint prediction,”arXiv preprint arXiv:2502.08486, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    RSRefSeg 2: Decoupling referring remote sensing image segmentation with founda- tion models,

    K. Chen, C. Liu, B. Chen, J. Zhang, Z. Zou, and Z. Shi, “RSRefSeg 2: Decoupling referring remote sensing image segmentation with founda- tion models,”IEEE Transactions on Geoscience and Remote Sensing, vol. 64, pp. 1–20, 2026

    2026

  15. [15]

    Zero-shot referring remote sensing image segmentation with expert knowledge priors and hierarchical relational matching,

    C. Sun, H. Han, J. Li, Y . Song, and B. Zhang, “Zero-shot referring remote sensing image segmentation with expert knowledge priors and hierarchical relational matching,”IEEE Geoscience and Remote Sensing Letters, vol. 23, pp. 1–5, 2026

    2026

  16. [16]

    RSVG-ZeroOV: Exploring a training-free framework for zero-shot open-vocabulary visual grounding in remote sensing images,

    K. Li, D. Wang, T. Wang, F. Dong, Y . Zhang, L. Zhang, X. Wang, S. Li, and Q. Wang, “RSVG-ZeroOV: Exploring a training-free framework for zero-shot open-vocabulary visual grounding in remote sensing images,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 8, 2026, pp. 6288–6296

    2026

  17. [17]

    DGL-RSIS: Decou- pling global spatial context and local class semantics for training-free remote sensing image segmentation,

    B. Li, C. Zhang, R. M. Timmerman, and W. Bao, “DGL-RSIS: Decou- pling global spatial context and local class semantics for training-free remote sensing image segmentation,”International Journal of Applied Earth Observation and Geoinformation, vol. 146, p. 105113, 2026

    2026

  18. [18]

    SAM 3: Segment anything with concepts,

    N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V . Alwala, H. Khedr, A. Huang, J. Lei, T. Ma, B. Guo, A. Kalla, M. Marks, J. Greer, M. Wang, P. Sun, R. R¨adle, T. Afouras, E. Mavroudi, K. Xu, T.-H. Wu, Y . Zhou, L. Momeni, R. Hazra, S. Ding, S. Vaze, F. Porcher, F. Li, S. Li, A. Kamath, H. K. Cheng, P. Doll ´ar, N. Ravi, K. ...

    Pith/arXiv arXiv 2026

  19. [19]

    SegEarth-OV: Towards training-free open-vocabulary segmentation for remote sensing images,

    K. Li, R. Liu, X. Cao, X. Bai, F. Zhou, D. Meng, and Z. Wang, “SegEarth-OV: Towards training-free open-vocabulary segmentation for remote sensing images,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 10 545– 10 556

    2025

  20. [20]

    SegEarth-OV3: Exploring SAM 3 for open-vocabulary semantic seg- mentation in remote sensing images,

    K. Li, S. Zhang, Y . Wang, Y . Deng, Z. Wang, D. Meng, and X. Cao, “SegEarth-OV3: Exploring SAM 3 for open-vocabulary semantic seg- mentation in remote sensing images,” 2026

    2026

  21. [21]

    InstructSAM: A training-free framework for instruction-oriented remote sensing object recognition,

    Y . Zheng, W. Wu, Q. Li, X. Wang, X. Zhou, A. Ren, J. Shen, L. Zhao, G. Li, and X. Yang, “InstructSAM: A training-free framework for instruction-oriented remote sensing object recognition,”arXiv preprint arXiv:2505.15818, 2025

    arXiv 2025

  22. [22]

    SegEarth-R1: Geospatial pixel reasoning via large language model,

    K. Li, Z. Xin, L. Pang, C. Pang, Y . Deng, J. Yao, G. Xia, D. Meng, Z. Wang, and X. Cao, “SegEarth-R1: Geospatial pixel reasoning via large language model,”arXiv preprint arXiv:2504.09644, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    GeoGround: A unified large vision-language model for remote sensing visual grounding,

    Y . Zhou, M. Lan, X. Li, L. Feng, Y . Ke, X. Jiang, Q. Li, X. Yang, and W. Zhang, “GeoGround: A unified large vision-language model for remote sensing visual grounding,”arXiv preprint arXiv:2411.11904, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Text4Seg++: Advancing image segmentation via generative lan- guage modeling,

    M. Lan, C. Chen, J. Xu, Z. Li, Y . Ke, X. Jiang, Y . Yu, Y . Zhao, and S. Bai, “Text4Seg++: Advancing image segmentation via generative lan- guage modeling,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  25. [25]

    Enabling training-free text-based remote sensing segmentation,

    J. Sosa, D. Rukhovich, A. Kacem, and D. Aouada, “Enabling training-free text-based remote sensing segmentation,”arXiv preprint arXiv:2602.17799, 2026

    arXiv 2026

  26. [26]

    Neural module networks,

    J. Andreas, M. Rohrbach, T. Darrell, and D. Klein, “Neural module networks,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 39–48

    2016

  27. [27]

    Neural-symbolic VQA: Disentangling reasoning from vision and lan- guage understanding,

    K. Yi, J. Wu, C. Gan, A. Torralba, P. Kohli, and J. B. Tenenbaum, “Neural-symbolic VQA: Disentangling reasoning from vision and lan- guage understanding,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 31, 2018

    2018

  28. [28]

    The neuro- symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision,

    J. Mao, C. Gan, P. Kohli, J. B. Tenenbaum, and J. Wu, “The neuro- symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision,” inInternational Conference on Learning Representations (ICLR), 2019

    2019

  29. [29]

    ReCLIP: A strong zero-shot baseline for referring ex- pression comprehension,

    S. Subramanian, W. Merrill, T. Darrell, M. Gardner, S. Singh, and A. Rohrbach, “ReCLIP: A strong zero-shot baseline for referring ex- pression comprehension,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL), 2022, pp. 5198– 5215

    2022

  30. [30]

    A new way to represent the relative position between areal objects,

    P. Matsakis and L. Wendling, “A new way to represent the relative position between areal objects,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 21, no. 7, pp. 634–643, 1999

    1999

  31. [31]

    Fuzzy spatial relationships for image processing and inter- pretation: a review,

    I. Bloch, “Fuzzy spatial relationships for image processing and inter- pretation: a review,”Image and Vision Computing, vol. 23, no. 2, pp. 89–110, 2005

    2005

  32. [32]

    Grounding spatial language in perception: An empirical and computational investigation,

    T. Regier and L. A. Carlson, “Grounding spatial language in perception: An empirical and computational investigation,”Journal of Experimental Psychology: General, vol. 130, no. 2, pp. 273–298, 2001

    2001

  33. [33]

    STAR: A first- ever dataset and a large-scale benchmark for scene graph generation in large-size satellite imagery,

    Y . Li, L. Wang, T. Wang, X. Yang, J. Luo, Q. Wang, Y . Deng, W. Wang, X. Sun, H. Li, B. Dang, Y . Zhang, Y . Yu, and J. Yan, “STAR: A first- ever dataset and a large-scale benchmark for scene graph generation in large-size satellite imagery,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 3, pp. 1832–1849, 2025

    2025

  34. [34]

    ReCon1M: A large-scale benchmark dataset for relation comprehension in remote sensing imagery,

    Q. Yan, C. Deng, C. Liu, Z. Hou, X. Liu, Y . Jiang, W. Lu, F. Yao, X. Liu, L. Hao, H. Yu, and X. Sun, “ReCon1M: A large-scale benchmark dataset for relation comprehension in remote sensing imagery,”IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1–22, 2025

    2025

  35. [35]

    AUG: A new dataset and an efficient model for aerial image urban scene graph generation,

    Y . Li, K. Li, Y . Zhang, L. Wang, and D. Zhang, “AUG: A new dataset and an efficient model for aerial image urban scene graph generation,” arXiv preprint arXiv:2404.07788, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    RSGround-R1: Rethinking remote sensing visual grounding through spatial reasoning,

    S. Huang, S. He, and B. Wen, “RSGround-R1: Rethinking remote sensing visual grounding through spatial reasoning,”arXiv preprint arXiv:2601.21634, 2026

    arXiv 2026

  37. [37]

    ProVG: Progressive visual grounding via language decoupling for remote sensing imagery,

    K. Li, T. Wang, D. Wang, Y . Zhu, Y . Zhang, T. Lei, and Q. Wang, “ProVG: Progressive visual grounding via language decoupling for remote sensing imagery,”arXiv preprint arXiv:2604.01893, 2026

    arXiv 2026

  38. [38]

    Grounded language-image pre- training,

    L. H. Li, P. Zhang, H. Zhang, J. Yang, C. Li, Y . Zhong, L. Wang, L. Yuan, L. Zhang, J.-N. Hwanget al., “Grounded language-image pre- training,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  39. [39]

    Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, C. Li, J. Yang, H. Su, J. Zhu, and L. Zhang, “Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection,” inProceedings of the European Conference on Computer Vision (ECCV), 2024

    2024

  40. [40]

    RS5M and GeoRSCLIP: A large-scale vision-language dataset and a large vision-language model for remote sensing,

    Z. Zhang, T. Zhao, Y . Guo, and J. Yin, “RS5M and GeoRSCLIP: A large-scale vision-language dataset and a large vision-language model for remote sensing,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  41. [41]

    RemoteCLIP: A vision language foundation model for remote sensing,

    F. Liu, D. Chen, Z. Guan, X. Zhou, J. Zhu, Q. Ye, L. Fu, and J. Zhou, “RemoteCLIP: A vision language foundation model for remote sensing,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1– 16, 2024

    2024

  42. [42]

    SkyScript: A large and semantically diverse vision-language dataset for remote sens- ing,

    Z. Wang, R. Prabha, T. Huang, J. Wu, and R. Rajagopal, “SkyScript: A large and semantically diverse vision-language dataset for remote sens- ing,” inProceedings of the AAAI Conference on Artificial Intelligence, 2024

    2024

  43. [43]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll´ar, and R. Girshick, “Segment anything,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  44. [44]

    SAM 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafsonet al., “SAM 2: Segment anything in images and videos,” inInternational Conference on Learning Repre- sentations (ICLR), vol. 2025, 2025, pp. 28 085–28 128

    2025

  45. [45]

    RemoteSAM: Towards segment anything for earth observation,

    L. Yao, F. Liu, D. Chen, C. Zhang, Y . Wang, Z. Chen, W. Xu, S. Di, and Y . Zheng, “RemoteSAM: Towards segment anything for earth observation,” inProceedings of the 33rd ACM International Conference on Multimedia (MM), 2025

    2025

  46. [46]

    Qwen3 technical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Slicing aided hyper inference and fine-tuning for small object detection,

    F. C. Akyon, S. O. Altinuc, and A. Temizel, “Slicing aided hyper inference and fine-tuning for small object detection,” inProceedings of the IEEE International Conference on Image Processing (ICIP), 2022, pp. 966–970. PREPRINT 20

    2022

  48. [48]

    Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture- of-LoRAs,

    A. Abouelenin, A. Ashfaq, A. Atkinson, H. Awadalla, N. Bach, J. Bao, A. Benhaim, M. Cai, V . Chaudhary, C. Chenet al., “Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture- of-LoRAs,”arXiv preprint arXiv:2503.01743, 2025

    Pith/arXiv arXiv 2025

  49. [49]

    Falcon mamba: The first competitive attention-free 7B language model,

    J. Zuo, M. Velikanov, D. E. Rhaiem, I. Chahed, Y . Belkada, G. Kunsch, and H. Hacid, “Falcon mamba: The first competitive attention-free 7B language model,”arXiv preprint arXiv:2410.05355, 2024

    Pith/arXiv arXiv 2024

  50. [50]

    The Llama 3 herd of models,

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughanet al., “The Llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  51. [51]

    CARIS: Context-aware referring image segmentation,

    S.-A. Liu, Y . Zhang, Z. Qiu, H. Xie, Y . Zhang, and T. Yao, “CARIS: Context-aware referring image segmentation,” inProceedings of the 31st ACM International Conference on Multimedia (MM), 2023, pp. 779– 788

    2023

  52. [52]

    Generalized re- ferring expression segmentation on aerial photos,

    L. Marnoto, A. Bernardino, and B. Martins, “Generalized re- ferring expression segmentation on aerial photos,”arXiv preprint arXiv:2512.07338, 2025

    arXiv 2025

This paper was first reviewed by grok-4.5 on July 11, 2026.