Pith. sign in

REVIEW 4 major objections 5 minor 57 references

A single click can be propagated to every same-class object in a remote-sensing scene, cutting clicks per image from about 39 to under 11 on dense building data.

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 ↗

ISRS-DETR combines a CrossCut segmenter with an RF-DETR detector and a dynamic top-K click selector to propagate a single user click to multiple same-class instances, reducing per-image click counts by 3–28 on three benchmarks.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A useful, well-scoped idea (propagate one click to same-class instances via a detector) with broad experiments, but the headline click reduction lives entirely in a new, underspecified metric and the paper has polish gaps. the 4 major comments →

arxiv 2608.02468 v1 pith:WUUFWLN4 submitted 2026-08-03 cs.CV

ISRS-DETR: Detection-Guided Click Propagation for Remote Sensing Interactive Segmentation

Thanh Duc Pham , Anh Nguyen , Duong Duc Hieu , Minh-Tan Pham This is my paper
classification cs.CV
keywords interactive segmentationremote sensingclick propagationdetection-guided segmentationDETRsimulated clicksNumber of Clicks per Imageclass-aware 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 reading

The paper tries to establish that interactive segmentation of remote-sensing imagery should exploit inter-object correlation instead of treating each object of a class as a separate target. Its method, ISRS-DETR, adds a DETR-based detection branch to an existing interactive segmentation backbone: after a user clicks one object, the branch localises other instances of the same class, filters unreliable boxes, and converts each reliable box into a simulated click that costs the user nothing. The authors show this propagation sharply reduces the Number of Clicks per Image: on the WHU-Building benchmark, 10.71 clicks reach 70% IoU versus 38.69 for the strongest baseline, and two real clicks already reach 91.0 mIoU. A sympathetic reader would care because it suggests scene-scale annotation of repetitive remote-sensing objects can be nearly linear in the number of classes rather than the number of instances.

Core claim

The central claim is that one positive user click can be turned into many free clicks by detector-guided propagation, and that this is what makes scene-level interactive segmentation affordable. The method projects the user click's relative position inside its own bounding box onto each detected same-class box, generating four orientation candidates per box, and then a Dynamic Top-K selection keeps only candidates whose fused-feature similarity to the user click forms a clear high-similarity group. These simulated clicks are fed to the segmentation branch as additional positive prompts and are not counted in the interaction budget. Experiments on three benchmarks report consistent NoC-I redu

What carries the argument

The load-bearing mechanism is the Dynamic Top-K Simulated Click Selection strategy. Candidate clicks are ranked by softmax-normalised inner-product similarity of their interpolated fused features to the clicked point; the sorted similarity curve is treated as a sigmoid, and the optimal number of retained simulated clicks k* is chosen as the minimum of its second-order discrete difference. This chooses, per image, how many detected same-class instances are reliable enough to receive a propagated click, replacing a fixed top-k.

Load-bearing premise

The whole click-saving effect rests on the detection branch proposing reliable same-class boxes; if the detector misses or mislabels a category, propagation cannot help and the method degrades toward the baseline (the paper's own oracle experiment shows a 4.99 vs 10.71 NoC-I gap).

What would settle it

Measure NoC-I on a category the detector localises poorly (e.g., a class with low detection recall) and compare with the same interaction loop using ground-truth boxes. If the real-detector curve approaches the baseline rather than the oracle, the central propagation claim fails for that category. Equivalently, if removing the Dynamic Top-K rule and using a fixed k=1 or k=5 yields the same or better NoC-I, the specific mechanism is not the source of the gain.

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

If this is right

  • Click annotation cost for repetitive classes can drop by roughly an order of magnitude, because each real click supervises all selected co-occurring instances.
  • The interaction budget stops scaling with instance count and instead scales with class count and scene diversity.
  • Per-image NoC-I becomes a practical evaluation protocol for scene-scale interactive segmentation, complementing per-object NoC.
  • Detector quality becomes the limiting factor: improving same-class proposal recall directly translates into further click savings.
  • At high IoU targets the advantage narrows, indicating the mechanism accelerates instance coverage more than boundary refinement.

Where Pith is reading between the lines

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

  • The same propagation idea could transfer to other repetitive-object domains, such as cell colonies or crop rows, where inter-object correlation is also high; the paper does not test this.
  • Because simulated clicks inherit the detector's class predictions, false-positive boxes could actively mislead the segmentation branch; a conservative confidence threshold or a human verification round might be needed in production annotation.
  • The Dynamic Top-K rule assumes a single elbow in the similarity curve; scenes with several distinct sub-groups of the same class (different orientations or lighting) could break that assumption and produce either too many or too few simulated clicks.
  • The reported two-click mIoU advantage suggests an interactive 'click once per class' workflow, which could be tested by measuring annotation time rather than click count.
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

4 major / 5 minor

Summary. The paper proposes ISRS-DETR, an interactive segmentation framework for remote sensing that augments the CrossCut segmentation backbone with an RF-DETR-based detection guide branch. The detector localizes same-class object instances from a single user click; a Dynamic Top-K Click Selection strategy (Eqs. 2–4) converts reliable detections into simulated clicks that do not count toward the interaction budget. Training uses pseudo ground-truth boxes derived from connected components of the segmentation masks, and the detection and segmentation branches are jointly optimized. Experiments on iSAID, WHU-Building, and NWPU VHR-10 report a new per-image metric (NoC-I), mIoU-vs-clicks curves, per-object NoC, and ablations including an oracle proposal-quality study (Table 3). The central claim is that ISRS-DETR achieves state-of-the-art accuracy while substantially reducing user clicks, e.g., 10.71 vs 38.69 NoC-I@70 on WHU-Building.

Significance. If the central claim is valid, the paper addresses a real practical bottleneck in remote sensing interactive segmentation: per-object interaction does not scale to scenes with many instances of the same class. The idea of propagating a single user click through a detector and a feature-similarity selection mechanism is clear and well motivated, and the oracle experiment in Table 3 is a useful upper-bound analysis. The paper also reports extensive comparisons on three benchmarks and states that code and data splits will be released, which is important for reproducibility. However, the primary evaluation metric NoC-I is introduced in this paper and is not clearly defined, and the reported numbers are internally inconsistent with the stated 20-click budget in Table A2. Because the headline click-reduction claim rests entirely on this unvalidated metric, the current evidence does not yet support the paper's strongest conclusion.

major comments (4)
  1. [Evaluation Metrics / Table 1 / Table A2] NoC-I is the paper's primary metric, but its definition is ambiguous and its reported values are inconsistent with the stated budget. The text says NoC-I measures 'the number of user clicks required to reach a target IoU threshold of every target instances of that image,' which could mean (a) every instance individually must reach the threshold, or (b) the image-level mIoU must reach the threshold; these give very different click counts in multi-instance scenes. More concretely, Table A2 states 'Max clicks per sample 20,' yet Table 1 reports NoC-I@80 values as high as 95.68 (SimpleClick on WHU-Building) and 54.24 (iSAID). This means the reported values cannot be direct click counts under the stated protocol; they must be extrapolated, interpolated, or the budget is not enforced at inference. None of this is described. Since the 28-click reduction over CrossCut on WHU-Building is the cent
  2. [Eq. (4) and Section 'Dynamic Top-K Simulated Click Selection'] The Dynamic Top-K rule is a load-bearing component: it determines the number of simulated clicks, and hence the strength of the click-propagation effect. The paper justifies k* as the minimum second-order discrete derivative by asserting that 'the score distribution typically exhibits a sharp transition' and that the sorted similarity curve 'appears alike a sigmoid function.' No quantitative evidence is provided for this claim — no example curve, no distribution statistics, and no ablation comparing Eq. (4) against fixed-k alternatives or a confidence-threshold-based rule. Table 2 shows that the average number of simulated clicks varies substantially across datasets (7.33 on iSAID, 2.01 on WHU-Building, 0.59 on NWPU VHR-10), but it does not show whether Eq. (4) selects the correct number or whether errors in k* hurt segmentation. The heuristic should be validated with an ablation and, at
  3. [Experimental Protocol / Table 1] All main results are single runs with no reported variance. State-of-the-art claims on three datasets and nine protocol columns are based on one training run per method. Given 55 epochs, multi-dataset mixture training, and a new selection heuristic, it is essential to report results over multiple seeds (at least 3) with mean and standard deviation, or a statistical significance test. Without this, it is not possible to know whether the improvements, some of which are small (e.g., NWPU VHR-10 NoC-I@70: 6.94 vs 8.52), are reliable.
  4. [Table 3 and Conclusion] The oracle experiment in Table 3 shows that replacing the detector with ground-truth proposals reduces WHU-Building NoC-I@70 from 10.71 to 4.99, which is a larger gap than the difference between the proposed method and the CrossCut baseline. The conclusion acknowledges that detector quality bounds the benefit, but the paper does not report per-category or per-dataset detection recall / precision, nor does it characterize failure cases. Since the central click-reduction claim depends on the detector reliably finding same-class instances, the manuscript should provide detection-quality numbers (e.g., recall of the filtered proposals on each benchmark) and ideally NoC-I broken down by object category. Without this, the claim that the method 'substantially reduces' clicks is too broad for categories where the detector localizes poorly.
minor comments (5)
  1. [Abstract / Related Work] The phrase 'MS-COCO (?)' in the supplementary material has a missing citation placeholder. Also 'Following(Robinson et al. 2026)' has a missing space.
  2. [Section 'Training Procedures'] Algorithm 1 line description says 'Simulated Click Generation (Lines10–1)' which is clearly a typo for 'Lines 10–16' or similar. Also 'exsisting' should be 'existing'.
  3. [Table 2] The caption refers to 'per class episode,' but the term 'episode' is not defined earlier. Clarify whether it means per interaction round or per sequence of interactions in an image.
  4. [Figure 4] The qualitative figure is informative, but the caption mentions 'same automatic clicking strategy' without specifying the click-simulation rule. Please state the rule (e.g., largest error region) in the caption or refer to the algorithm.
  5. [Supplementary / Macroscopic Correlation] The MC metric is properly defined, but the comparison with MS-COCO uses an uncited reference and the sample size (4.4K) is not justified. Also, the claim that remote sensing images 'concentrate at markedly higher MC values' is only supported by a density plot, not by a statistical test; consider reporting means and standard deviations.

Circularity Check

1 steps flagged

Headline NoC-I reduction is partly defined into existence: simulated clicks are excluded from the click budget by construction.

specific steps
  1. self definitional [Evaluation Metrics (Experiments) and Simulated Click Generation (Method)]
    "we introduce NoC-I (Number of Clicks per Image). Similar to the traditional NoC metric, NoC-I measures the number of user clicks required to reach a target IoU threshold of every target instances of that image. ... Since they are generated automatically from a single positive user click, they are not counted toward the interaction budget."

    NoC-I is the paper's primary efficiency metric and the source of the headline margins (e.g., WHU-Building @70: 38.69 for CrossCut vs 10.71 for ISRS-DETR). Yet NoC-I counts 'user clicks', and the method's core mechanism is declared to be non-user: every detector-generated simulated click is 'not counted toward the interaction budget' by definition. The reduction is therefore partly arithmetic—the method's added click-equivalents are simply placed outside the numerator. The conventional per-object NoC in Table A1 shows a much smaller advantage on the same comparison (4.76 vs 1.32), confirming that the dramatic 28-click gap is substantially produced by the counting rule rather than by segmentation accuracy.

full rationale

The only circular element I find is in the headline efficiency claim. NoC-I is introduced by this paper, and simulated clicks are excluded from the click count by fiat, which makes the large NoC-I reduction partly an artifact of metric definition. This is a real but partial circularity: the mIoU/AuC curves and per-object NoC are independent, externally comparable measurements, and they still show genuine gains. I do not see additional circularity: the RF-DETR decoder and Dynamic Top-K are standard supervised/selection components; the pseudo-GT boxes are extracted from the same masks used for segmentation, but that is co-adaptation rather than a fitted output renamed as a prediction, and Table 3 honestly exposes the detector bottleneck by reporting a GT-proposal oracle. The paper itself acknowledges detector dependence as a limitation. The NoC-I definition's ambiguity and the inconsistency between 'Max clicks per sample 20' in Table A2 and NoC-I values above 20 in Table 1 are correctness and verification risks, not circular derivation steps.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No new physical or conceptual entities are introduced. The core free parameters are hand-set thresholds and the heuristic k* rule; the main ad hoc assumptions are the pseudo-box training signal and the sigmoid sharp-transition claim underlying Eq. 4.

free parameters (6)
  • Confidence threshold θ = 0.30
    Boxes with confidence below 0.30 are discarded; hand-set, affects proposal quality and the simulated click set.
  • Minimum side length s_min = 0.004 (normalized)
    Discards tiny boxes; hand-set.
  • NMS IoU threshold = 0.6
    Used to remove duplicate detections; hand-set.
  • Max detections per episode = 64
    Limits the number of simulated clicks per episode; hand-set.
  • k* selection rule (Eq. 4) = min second-order discrete derivative
    The number of simulated clicks is chosen by the sharpest elbow of the sorted softmax similarity curve; the rule is asserted by observation and not otherwise validated.
  • Data mixture ratio = 0.40 : 0.35 : 0.25 (iSAID : WHU : NWPU)
    Training ratio for the three datasets; hand-chosen and affects the joint model.
axioms (4)
  • domain assumption Pearson correlation between bounding box descriptors captures meaningful inter-object correlation (Eq. A2)
    Used to motivate the approach in Fig. 1; assumes high MC implies informativeness of one object for another.
  • ad hoc to paper Pseudo ground-truth boxes obtained from connected components of the segmentation mask provide a valid training signal for the detector branch
    Section 'Training Procedures' and Algorithm A1: boxes from connected components can be irregular for touching or overlapping objects and may not match instance boundaries.
  • ad hoc to paper The sorted softmax similarity curve is sigmoid-like with a sharp transition (used in Eq. 4)
    The text asserts 'we observe that the score distribution typically exhibits a sharp transition'; no evidence is provided, and the optimality of the min second-derivative is assumed.
  • domain assumption Simulated clicks are 'free' and do not count toward the user interaction budget
    Evaluation protocol convention; if simulated clicks were counted, the NoC-I reductions would shrink.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ISRS-DETR: Detection-Guided Click Propagation for Remote Sensing Interactive Segmentation." pith.science (2026). https://pith.science/paper/WUUFWLN4

@misc{pith2026260802468,
  author       = {Pith},
  title        = {Pith review of: ISRS-DETR: Detection-Guided Click Propagation for Remote Sensing Interactive Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WUUFWLN4}},
  note         = {Machine review of arXiv:2608.02468}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Interactive segmentation reduces the prohibitive cost of pixel-level annotation by allowing users to delineate objects with a few clicks. However, applying this paradigm directly to remote sensing imagery is non-trivial: ultra-high resolutions, small object sizes, and sparse spatial distributions all degrade segmentation quality. Recent work has addressed the resolution barrier and achieved competitive results in interactive segmentation for remote sensing (ISRS). However, they treat all instances of a class within an image as a single objective target. Consequently, interactions spent on one object contribute nothing to its same-class neighbours, and satisfactory masks may demand up to 40 clicks per image, hindering the practicality of these frameworks. We observe that remote sensing scenes exhibit markedly strong inter-object correlation, meaning a single clicked object is highly informative about the rest of its category. Building on this, we propose ISRS-DETR, a detection-guided interactive segmentation framework that injects object-level evidence into both training and inference. Our ISRS-DETR employs an RF-DETR decoder with the interactive segmentation backbone to localise co-occurring same-class objects, and introduces a Dynamic Top-K Click Selection strategy that retains only reliable proposals and converts each into a simulated click, so one user interaction propagates across an entire class. Experiments on three standard remote sensing benchmarks show that ISRS-DETR achieves state-of-the-art accuracy while substantially reducing Number of Clicks per Image (NoC-I). All codes and data splits will be released for reproducibility upon acceptance.

Figures

Figures reproduced from arXiv: 2608.02468 by Anh Nguyen, Duong Duc Hieu, Minh-Tan Pham, Thanh Duc Pham.

Figure 1
Figure 1. Figure 1: Statistical distribution of macroscopic correlation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An overview of our ISRS-DETR framework: Our framework consists of two branches: (i) The segmentation branch follows CrossCut to predict patch-level masks and reconstruct the final segmentation map. (ii) The detection guide branch leverages the RF-DETR decoder with dedicated modules to propagate a single user click to all remaining similar instances of the same category. rectly converting all candidate clic… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of the mean IoU scores according to the number of clicks on the iSAID (Waqas Zamir et al. 2019), [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on NWPU VHR-10 at the same interaction budget of [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 5 linked inside Pith

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  9. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  10. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  11. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  12. [12]

    Lin, Zheng and Zhou, Nan and Wang, Yuhan and Zhang, Bojian , booktitle =

  13. [13]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

    Attention Retention for Continual Learning with Vision Transformers , author =. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

  14. [14]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Rethinking Interactive Image Segmentation with Low Latency, High Quality, and Diverse Prompts , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  15. [15]

    Lee, Chaewon and Lee, Seon-Ho and Kim, Chang-Su , booktitle =

  16. [16]

    Relation

    Hou, Xiuquan and Liu, Meiqin and Zhang, Senlin and Wei, Ping and Chen, Badong and Lan, Xuguang , booktitle =. Relation. 2024 , doi =

    2024

  17. [17]

    Robinson, Isaac and Robicheaux, Peter and Popov, Matvei and Ramanan, Deva and Peri, Neehar , booktitle =

  18. [18]

    Graph cuts and efficient

    Boykov, Yuri and Funka-Lea, Gareth , journal =. Graph cuts and efficient

  19. [19]

    Interactive image segmentation using an adaptive

    Blake, Andrew and Rother, Carsten and Brown, Matthew and Perez, Patrick and Torr, Philip , booktitle =. Interactive image segmentation using an adaptive

  20. [20]

    Rother, Carsten and Kolmogorov, Vladimir and Blake, Andrew , journal =

  21. [21]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Graph cut based image segmentation with connectivity priors , author =. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  22. [22]

    European Conference on Computer Vision (ECCV) , year =

    Star shape prior for graph-cut image segmentation , author =. European Conference on Computer Vision (ECCV) , year =

  23. [23]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) , year =

    Random walks for image segmentation , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) , year =

  24. [24]

    European Conference on Computer Vision (ECCV) , year =

    Generative image segmentation using random walks with restart , author =. European Conference on Computer Vision (ECCV) , year =

  25. [25]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Geodesic star convexity for interactive image segmentation , author =. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  26. [26]

    International Journal of Computer Vision (IJCV) , year =

    Geodesic matting: A framework for fast interactive image and video segmentation and matting , author =. International Journal of Computer Vision (IJCV) , year =

  27. [27]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Geodesic graph cut for interactive image segmentation , author =. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  28. [28]

    IEEE International Conference on Image Processing (ICIP) , year =

    Reviving iterative training with mask guidance for interactive segmentation , author =. IEEE International Conference on Image Processing (ICIP) , year =

  29. [29]

    Chen, Xi and Zhao, Zhiyan and Zhang, Yilei and Duan, Manni and Qi, Donglian and Zhao, Hao , booktitle =

  30. [30]

    Liu, Qin and Xu, Zhenlin and Bertasius, Gedas and Niethammer, Marc , booktitle =

  31. [31]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Segment anything , author =. IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  32. [32]

    CVPR , pages=

    Deep interactive object selection , author=. CVPR , pages=

  33. [33]

    2025 , eprint=

    ROS-SAM: High-Quality Interactive Segmentation for Remote Sensing Moving Object , author=. 2025 , eprint=

    2025

  34. [34]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Conditional detr for fast training convergence , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  35. [35]

    arXiv preprint arXiv:2201.12329 , year=

    DAB-DETR: Dynamic anchor boxes are better queries for DETR , author=. arXiv preprint arXiv:2201.12329 , year=

    Pith/arXiv arXiv

  36. [36]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Dn-detr: Accelerate detr training by introducing query denoising , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  37. [37]

    arXiv preprint arXiv:2203.03605 , year=

    Dino: Detr with improved denoising anchor boxes for end-to-end object detection , author=. arXiv preprint arXiv:2203.03605 , year=

    Pith/arXiv arXiv

  38. [38]

    arXiv preprint arXiv:2207.13080 , year=

    DETRs with Hybrid Matching , author=. arXiv preprint arXiv:2207.13080 , year=

    Pith/arXiv arXiv

  39. [39]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Anchor detr: Query design for transformer-based detector , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  40. [40]

    arXiv preprint arXiv:2203.16527 , year=

    Exploring plain vision transformer backbones for object detection , author=. arXiv preprint arXiv:2203.16527 , year=

    Pith/arXiv arXiv

  41. [41]

    Proceedings of the IEEE international conference on computer vision , pages=

    Fast r-cnn , author=. Proceedings of the IEEE international conference on computer vision , pages=

  42. [42]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Feature pyramid networks for object detection , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  43. [43]

    arXiv preprint arXiv:2010.04159 , year=

    Deformable detr: Deformable transformers for end-to-end object detection , author=. arXiv preprint arXiv:2010.04159 , year=

    Pith/arXiv arXiv 2010

  44. [44]

    2020 , eprint=

    End-to-End Object Detection with Transformers , author=. 2020 , eprint=

    2020

  45. [45]

    2024 , eprint=

    DETRs Beat YOLOs on Real-time Object Detection , author=. 2024 , eprint=

    2024

  46. [46]

    2024 , eprint=

    Relation DETR: Exploring Explicit Position Relation Prior for Object Detection , author=. 2024 , eprint=

    2024

  47. [47]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , month =

    Waqas Zamir, Syed and Arora, Aditya and Gupta, Akshita and Khan, Salman and Sun, Guolei and Shahbaz Khan, Fahad and Zhu, Fan and Shao, Ling and Xia, Gui-Song and Bai, Xiang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , month =. 2019 , pages =

    2019

  48. [48]

    The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Xia, Gui-Song and Bai, Xiang and Ding, Jian and Zhu, Zhen and Belongie, Serge and Luo, Jiebo and Datcu, Mihai and Pelillo, Marcello and Zhang, Liangpei , title =. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

  49. [49]

    2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS) , pages =

    Maggiori, Emmanuel and Tarabalka, Yuliya and Charpiat, Guillaume and Alliez, Pierre , title =. 2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS) , pages =. 2017 , doi =

    2017

  50. [50]

    ISPRS Journal of Photogrammetry and Remote Sensing , volume =

    Cheng, Gong and Han, Junwei and Zhou, Peicheng and Guo, Lei , title =. ISPRS Journal of Photogrammetry and Remote Sensing , volume =. 2014 , doi =

    2014

  51. [51]

    2017 , eprint=

    Adam: A Method for Stochastic Optimization , author=. 2017 , eprint=

    2017

  52. [52]

    2016 , eprint=

    The Cityscapes Dataset for Semantic Urban Scene Understanding , author=. 2016 , eprint=

    2016

  53. [53]

    Nature methods , volume=

    nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation , author=. Nature methods , volume=. 2021 , publisher=

    2021

  54. [54]

    Proceedings of the IEEE , volume=

    Remote sensing image scene classification: Benchmark and state of the art , author=. Proceedings of the IEEE , volume=. 2017 , publisher=

    2017

  55. [55]

    IEEE Geoscience and Remote Sensing Magazine , volume=

    Deep learning in remote sensing: A comprehensive review and list of resources , author=. IEEE Geoscience and Remote Sensing Magazine , volume=. 2017 , publisher=

    2017

  56. [56]

    ISPRS Journal of Photogrammetry and Remote Sensing , volume=

    Beyond RGB: Very high resolution urban remote sensing with multimodal deep networks , author=. ISPRS Journal of Photogrammetry and Remote Sensing , volume=. 2018 , publisher=

    2018

  57. [57]

    2025 , eprint=

    AerOSeg: Harnessing SAM for Open-Vocabulary Segmentation in Remote Sensing Images , author=. 2025 , eprint=

    2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.