Pith. sign in

REVIEW 3 major objections 5 minor 54 references

SULAND v2: A Refined RGB Dataset and Deep Learning Object Detection Benchmark for UAV/UGV-Based SUrface LANDmine Detection Under Domain Shift

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Label errors in the SULAND landmine benchmark, not just model choice, drive what detector comparisons appear to show: re-annotation shifts mAP@50 by up to 44 points and flips which families look OOD-robust.

desk verdict A genuinely useful benchmark repair: the class-ID fix and cross-version experiments are convincing, but the new ground truth is a manual re-annotation without independent validation, so treat the headline mAP gains as indicative rather than exact. read the letter →

arxiv 2607.28996 v1 pith:RWCA4EAK submitted 2026-07-31 cs.CV

Sagar Lekhak , Prasanna Reddy Pulakurthi , Lalit Joshi , Ramesh Bhatta , Emmett J. Ientilucci This is my paper
classification cs.CV
keywords SULANDv2surfacelandminedetectionRGBobjectdatasetre-annotationbenchmarkreliabilityout-of-distributionevaluationdomainshiftYOLOandDETRdetectors
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

This paper claims that the widely used SULAND surface-landmine detection benchmark, as originally released, was too error-ridden to support its published conclusions about out-of-distribution (OOD) robustness. The authors manually re-annotated all 33,771 images under a single visibility rule, fixed an inverted class convention in the OOD split, and show that the relabeling alone moves YOLOv8 in-distribution scores by 14.6–19.6 mAP@50 points and OOD scores by roughly 25 points from the class-ID fix alone. On the corrected data they benchmark 35 detector configurations, finding YOLOv12-Small best in-distribution (0.908 mAP@50) but RF-DETR-Large and Faster R-CNN far ahead under domain shift (0.799 OOD mAP@50). The reason a sympathetic reader should care: in a safety-critical domain with few public datasets, benchmark label errors can silently misdirect which detectors get deployed.

What carries the argument

The load-bearing construction is SULAND v2 itself — a fully re-annotated label set — plus a cross-version evaluation protocol that separates label effects from model effects. Annotations were redone sample-by-sample under one eligibility rule (annotate any visible, class-identifiable portion of a target; box only the visible extent), then categorized box-by-box as added, removed, tightened, or class-corrected, with three reviewers voting on each changed box. The protocol then trains YOLOv8 on either version and grades it against either version, including re-scoring identical OOD predictions with corrected class IDs. This design is what lets the authors attribute score changes to the labels r

What would settle it

Have an independent team re-annotate a random sample of the SULAND images under the same rules; if their boxes agree with SULAND v1 as often as with SULAND v2 (measured, say, as precision/recall of each version against the independent labels), the improvement is mostly policy, not accuracy. A minimal version the authors could run themselves: report inter-annotator agreement for the v2 labeling process, since only majority-vote consensus is described.

Watch

Extended reading notes

Core claim

The central discovery is that annotation quality, not just architecture, determines what the benchmark appears to say. The audit catalogued seven defect classes in SULAND v1 — missing and false labels, mislocalized boxes, inconsistent partial-visibility criteria, non-representative artifacts, blur-degraded boxes, and a reversed class-ID mapping between the IID and OOD splits. SULAND v2 keeps every image and every train/val/test split intact while replacing the labels: 3,134 boxes added and 1,930 removed in the IID training split alone, most retained boxes tightened, and the OOD class mapping corrected. The decisive evidence is cross-version evaluation that re-grades the same predictions: wit

Load-bearing premise

The load-bearing premise is that the authors' manual re-annotation is closer to the true target locations than SULAND v1's labels; the eligibility rule ('a visible and class-identifiable portion could be discerned') is a judgment call, and no inter-annotator agreement or independent ground-truth check is reported, so if that premise fails the reported gains partly measure a change in labeling policy rather than label accuracy.

Editorial extensions

If this is right

  • Re-grading identical OOD predictions after fixing only the class-ID mapping raises mean YOLOv8 OOD mAP@50 from 10.1% to 35.0%, so the original dataset's headline OOD degradation was largely a labeling artifact — prior comparisons built on those numbers need re-reading.
  • Detector rankings change with the corrected labels: YOLOv12-Small leads IID (0.908 mAP@50) while RF-DETR-Large leads OOD (0.799), with Faster R-CNN variants close behind in OOD — no single architecture tops both settings.
  • Every evaluated configuration loses accuracy on the OOD split; the gap is much larger for YOLO-family detectors than for RF-DETR and Faster R-CNN, so IID accuracy alone is not evidence of deployment robustness in this domain.
  • Within a detector family, increasing model scale does not consistently improve OOD retention, so parameter count is a poor predictor of domain-shift behavior.
  • A tiered operational workflow — a fast detector for initial candidate generation paired with a more OOD-robust detector or a human reviewer — follows from the accuracy–speed tradeoff the benchmark exposes.

Reading between the lines

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

  • Editorial: if relabeling changes scores by this much on SULAND, other small-domain remote-sensing benchmarks built by similar manual pipelines may carry label-error effects large enough to alter published rankings; a v1-vs-v2-style re-grading check is a cheap way to find out before trusting any domain-shift conclusion.
  • Editorial: the class-ID result offers a general sanity test for OOD benchmarks — re-grade fixed predictions under the swapped class mapping; if the gap largely vanishes, the 'domain shift' was partly a bookkeeping error.
  • Editorial: the paper's central assumption could be tested directly by an independent re-annotation of a random image sample; if an independent team agrees with v2 no more often than with v1, part of the reported improvement is a different annotation policy rather than greater accuracy.
  • Editorial: because the paper does not ablate pretraining, whether RF-DETR's and Faster R-CNN's OOD advantage comes from architecture or from the choice of pretrained backbone remains open; swapping backbones while fixing the detection head would separate these.
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 presents SULAND v2, a re-annotated version of the SULAND RGB surface-landmine dataset. The authors audit SULAND v1, categorize seven annotation defect types, manually re-annotate all 33,771 images, and release corrected labels while keeping the images and IID/OOD splits unchanged. They quantify annotation changes and run YOLOv8 cross-version training/evaluation to isolate training-label and evaluation-label effects. They report that v2 annotations raise YOLOv8 IID test mAP@50 by 14.6–19.6 percentage points, and that correcting the inverted OOD class-ID convention raises mean OOD mAP@50 by about 25 percentage points. They then benchmark 35 detector configurations across nine families on v2, reporting YOLOv12-S as best IID (0.908 mAP@50) and RF-DETR-L as best OOD (0.799 mAP@50), and they use the results to argue that high IID accuracy does not imply operational readiness.

Significance. If the v2 labels are accepted, this is a valuable and potentially high-impact benchmark correction for a scarce, safety-critical domain. The cross-version experimental design is a genuine strength: it cleanly separates training-supervision effects from evaluation-ground-truth effects. The OOD class-ID correction is objective, externally checkable, and well supported by the ~25 pp OOD gain. The release of audit records and code is exemplary. The main risk is that the v2 annotation set itself is not independently validated; the headline IID gains and resulting detector rankings depend on the unverified premise that the manual re-annotation is closer to true target locations. The paper is transparent about this in §VII-D, and the concern is addressable.

major comments (3)
  1. [§IV-A, §VII-D, Tables IV–V] The benchmark-reliability claim rests on the assumption that v2 annotations are closer to true target positions than v1. The eligibility criterion ("a visible and class-identifiable portion") is subjective, and the three-reviewer majority vote was applied to crops of already-made corrections, not to independent blind re-annotation. This protocol can catch inconsistencies but not systematic annotator bias. Section VII-D concedes that residual errors may remain and asks for inter-annotator agreement in future versions. Without such a check, the reported IID gains (14.6–19.6 pp, Table IV) and cross-version deltas (Table V) partly measure an annotation-policy shift rather than label accuracy. Please add an independent validation: have multiple annotators re-annotate a random subset (e.g., 200–300 images), report box-level and class-level agreement, and evaluate detectors against each annotat
  2. [§V-B, Tables IV–VI] All experiments use a single fixed seed (42), so every mAP, precision, recall, and ranking is a point estimate with no error bars. Given the relatively small OOD split (4,436 images) and the fact that some ranking differences are small (e.g., YOLOv12-S 0.908 vs YOLO11-L 0.906 IID mAP@50), stochastic variation could alter several conclusions. Please run at least 3–5 seeds for the main comparisons (or bootstrap the test set), and report mean ± std. This is necessary to support the benchmark's ranking claims.
  3. [§VI-B, §VII-D] The OOD conclusions are based on a single Italy→USA split. The paper explicitly acknowledges this limitation in §VII-D, but the claim that "RF-DETR-L yields the strongest OOD performance" and the broader robustness ranking are conditioned on one transfer direction with concurrent changes in terrain, illumination, viewpoint, and target appearance. A single split cannot establish whether the observed architecture-dependent ordering is a stable property or an artifact of this particular shift. At minimum, report per-sequence OOD results for the ten US folders and discuss variance across sequences; ideally add a second independent OOD domain.
minor comments (5)
  1. [Table III] The four correction scenarios are described as disjoint, but the arithmetic is not transparent. For example, OOD val adds 693, removes 624, and class-fixes 3,162; do the class-fixed boxes overlap with the tightened or added/removed categories? Please provide a Venn-style breakdown or a closed-form relation Inst_v2 = Inst_v1 + Added − Removed, with class-fix as a relabeling of matched boxes.
  2. [§V-B] "MuSGD optimizer (a Muon–SGD hybrid)" is unusual for Ultralytics YOLO training; please verify the exact optimizer name and configuration. Also clarify why Faster R-CNN is limited to 50 epochs even though the convergence analysis suggests earlier peak.
  3. [§V-C/§VI-A] Precision and recall are reported at the F1-optimal threshold selected on the IID validation set. This threshold may not be appropriate for OOD. At least state this limitation, or report the operating-point dependence for the main OOD comparisons.
  4. [Figure 2] t-SNE parameters (perplexity, iterations, seed) are not given, and t-SNE can exaggerate global separation. The qualitative conclusion is fine, but add the parameters and consider a quantitative divergence measure (e.g., Fréchet distance).
  5. [§IX] The URLs contain spaces ("SULAND v2 RGB Surface Landmine Dataset"); use percent-encoding or underscores for reliable access.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: benchmark results are empirical, and the v2 label self-reference is a validity caveat, not a definitional reduction.

full rationale

No step in this paper reduces to its own input by the paper's own equations or by a load-bearing self-citation. The headline quantities are empirical measurements: models are fine-tuned on fixed IID training splits and scored on held-out IID test and OOD validation splits using standard mAP metrics under a common protocol (Sections V-B, VI-A). Table V explicitly separates training-label effects from evaluation-label effects through cross-version v1->v2 / v2->v1 re-grading, so the same-version v2->v2 gains are not forced by construction. The OOD class-ID correction is an externally verifiable label convention, and the reported ~25 percentage-point effect is computed by re-scoring identical predictions against corrected labels (Section IV-D), not by fitting a parameter and then renaming it a prediction. The principal self-referential element is that the authors manually created the SULAND v2 annotations and then used those same annotations as the reference ground truth. Section IV-A's eligibility rule ('a visible and class-identifiable portion of the object could be discerned') is subjective, and Section VII-D concedes that 'some residual omissions, localization inconsistencies, or class-assignment errors may remain' and asks for future quantitative inter-annotator assessment. This is a validity and assumption threat for the benchmark-reliability claim, but it is not a circular derivation: no equation, fitted value, or self-citation makes the benchmark conclusions equivalent to the inputs. Existing self-citations ([12]-[14], [24], [25], [38]) are contextual literature references and are not load-bearing for the central empirical findings.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central numerical results depend on three hand-set components: the subjective annotation eligibility rule, the per-model F1 threshold, and the unified training schedule. These are not hidden, but they are choices that shape the benchmark. The main domain assumptions are that the authors' copy of SULAND v1 is representative, their manual labels are ground truth, and the Italy→USA split is a meaningful OOD test.

free parameters (3)
  • Annotation visibility/eligibility criterion
    Defined as 'visible and class-identifiable portion'; applied manually across all frames and used to add/remove boxes in v2, so the ground-truth change depends on this hand-set criterion.
  • F1-optimal confidence threshold = per-configuration threshold on IID val
    Precision/recall numbers are reported at the threshold maximizing F1 on IID validation; this is chosen from data and shifts OOD comparisons.
  • Training schedule and best-epoch selection = 100 epochs (50 for Faster R-CNN); best val mAP@50-95
    Schedule and checkpoint selection are hand-chosen to ensure convergence; adequate, but influences the reported results.
assumptions (3)
  • domain assumption The downloaded SULAND v1 release matches the official dataset closely enough to draw conclusions about the original benchmark.
    §III-B reports folder counts differ from the original paper (45 vs 47 IID folders, 10 vs 11 OOD sequences); all statistics use the authors' copy.
  • domain assumption Manual visual re-annotation is a more accurate ground truth than v1 annotations.
    No independent physical truthing or inter-annotator agreement metric is provided; §VII-D admits residual errors.
  • domain assumption Italy→USA transfer with additional scene variation is a valid operational OOD test.
    t-SNE separation of ImageNet features is qualitative (§VII-D); visual differences may be entangled with annotation-policy changes and class-prior shift.
invented entities (1)
  • SULAND v2 annotation set independent evidence
    purpose: Corrected ground truth for the existing SULAND images/splits; benchmark labels used for all experiments
    It is a publicly released dataset artifact with promised audit records; unlike a theoretical entity it can be inspected and reused, so it carries independent evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SULAND v2: A Refined RGB Dataset and Deep Learning Object Detection Benchmark for UAV/UGV-Based SUrface LANDmine Detection Under Domain Shift." pith.science (2026). https://pith.science/paper/RWCA4EAK

@misc{pith2026260728996,
  author       = {Pith},
  title        = {Pith review of: SULAND v2: A Refined RGB Dataset and Deep Learning Object Detection Benchmark for UAV/UGV-Based SUrface LANDmine Detection Under Domain Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RWCA4EAK}},
  note         = {Machine review of arXiv:2607.28996}
}
read the original abstract

RGB imagery offers a practical, low-cost option for Unmanned Aerial/Ground Vehicle (UAV/UGV) survey support in surface-landmine detection, but object detectors remain underexplored in this safety-critical domain. Limited cross-architecture benchmarking and insufficient out-of-distribution (OOD) analysis obscure whether detectors generalize across deployment conditions. This challenge is amplified by the scarcity of public RGB landmine datasets, making SULAND a key benchmark for PFM-1 and PMA-2 detection. However, inspection reveals missing/false annotations, localization errors, inconsistent visibility criteria, visual artifacts, temporal labeling inconsistencies, and an inverted OOD class-ID convention in SULAND. We present SULAND_v2, a refined RGB surface-landmine dataset and benchmark. Preserving original images and splits, we manually revise annotations to ensure completeness, precise localization, label validity, and class consistency. SULAND_v2 contains 33,771 images and 12,433 bounding boxes. We benchmark 35 detector configurations across nine families. Annotation refinement improves YOLOv8 in-distribution (IID) test mAP@50 by 14.6-19.6 percentage points, while fixing the OOD class-ID convention increases mean YOLOv8 OOD mAP@50 by ~25 percentage points. On SULAND_v2, YOLOv12-Small achieves the highest IID mAP@50 (0.908), while RF-DETR-Large yields the strongest OOD performance (0.799 mAP@50, 0.675 recall). Our results demonstrate that high IID accuracy does not guarantee operational readiness. SULAND_v2 provides a reliable benchmark for evaluating domain-shift robustness in RGB-based mine-action survey support.

Figures

Figures reproduced from arXiv: 2607.28996 by the authors.

Figure 1
Figure 1. Representative target class examples segmented from [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. t-SNE projection of ImageNet-pretrained features for [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Additional folder-level details are provided in the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (7 more)
Figure 3
Figure 3. Figure 3: Representative cropped examples of annotation defects identified during the SULAND [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]
Figure 4
Figure 4. Figure 4: Representative examples of class-ID mismatch in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Representative annotation corrections from SULAND [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: IID and OOD detection performance for all models trained on SULAND [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Accuracy–speed tradeoff for detectors trained on SULAND [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Per-class precision and recall for PFM-1 ( [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Representative IID and OOD detections for the best OOD configuration from each detector family, selected by [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 2 canonical work pages

  1. [1]

    Mine action,

    United Nations Mine Action Service (UNMAS), “Mine action,” https://unmas.org/en, 2026, accessed: 2026-05-01

    2026

  2. [2]

    Detection and clearance,

    Geneva International Centre for Humanitarian Demining (GICHD), “Detection and clearance,” https://www.gichd.org/ our-response/operations-management/detection-and-clearance/, 2026, accessed: 2026-05-01

    2026

  3. [3]

    Landmine monitor 2025,

    International Campaign to Ban Landmines, “Landmine monitor 2025,” https://the-monitor.org/reports/landmine-monitor-2025, Dec. 2025

    2025

  4. [4]

    Ukraine: Surveying and clearing mines,

    The HALO Trust, “Ukraine: Surveying and clearing mines,” https://www.halotrust.org/where-we-work/ europe-and-caucasus/ukraine/, 2026, accessed: 2026-05-01

    2026

  5. [5]

    Geneva International Centre for Humanitarian Demining (GICHD), “Survey,” https://www.gichd.org/our-response/ operations-management/survey/, 2026, accessed: 2026-05-01

    2026

  6. [6]

    A guide to non-technical survey,

    ——, “A guide to non-technical survey,” https: //www.gichd.org/publications-resources/publications/ a-guide-to-non-technical-survey/, 2026, accessed: 2026-05-01

    2026

  7. [7]

    Unmanned aerial systems for photogrammetry and remote sensing: A review,

    I. Colomina and P. Molina, “Unmanned aerial systems for photogrammetry and remote sensing: A review,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 92, pp. 79– 97, 2014. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S0924271614000501

    2014

  8. [8]

    Uav for 3d mapping applications: a review,

    F. Nex and F. Remondino, “Uav for 3d mapping applications: a review,”Applied Geomatics, vol. 6, no. 1, pp. 1–15, Mar 2014. [Online]. Available: https://doi.org/10.1007/s12518-013-0120-x

    doi:10.1007/s12518-013-0120-x 2014

Show all 54 references
  1. [9]

    Detection and identification of remnant pfm-1 ‘butterfly mines’ with a uav-based thermal-imaging protocol,

    A. Nikulin, T. S. De Smet, J. Baur, W. D. Frazer, and J. C. Abramowitz, “Detection and identification of remnant pfm-1 ‘butterfly mines’ with a uav-based thermal-imaging protocol,” Remote Sensing, vol. 10, no. 11, 2018. [Online]. Available: https://www.mdpi.com/2072-4292/10/11/1672

    2018

  2. [10]

    Uav for landmine detection using sdr-based gpr technology,

    M. R. P. Cerquera, J. D. C. Monta ˜no, and I. Mondrag ´on, “Uav for landmine detection using sdr-based gpr technology,” inRobots Operating in Hazardous Environments, H. Canbolat, Ed. London: IntechOpen, 2017, ch. 2. [Online]. Available: https://doi.org/10.5772/intechopen.69738

    2017 doi

  3. [11]

    A comprehensive review on landmine detection using deep learning techniques in 5G environment: open issues and challenges,

    A. Barnawi, I. Budhiraja, K. Kumar, N. Kumar, B. Alzahrani, A. Almansour, and A. Noor, “A comprehensive review on landmine detection using deep learning techniques in 5G environment: open issues and challenges,”Neural Computing and Applications 2022 34:24, vol. 34, no. 24, pp....

    2022 doi

  4. [12]

    Viability of Substituting Handheld Metal Detectors with an Airborne Metal Detection System for Landmine and Unexploded Ordnance Detection,

    S. Lekhak, E. J. Ientilucci, and A. W. Brinkley, “Viability of Substituting Handheld Metal Detectors with an Airborne Metal Detection System for Landmine and Unexploded Ordnance Detection,”Remote Sensing 2024, Vol. 16,, vol. 16, no. 24, 12

    2024

  5. [13]

    A comparative evaluation of uav-based remote sensing and geophysical techniques for landmine detection on a seeded minefield,

    J. Baur, S. Lekhak, G. Steinberg, A. Nikulin, T. de Smet, A. Brinkley, E. J. Ientilucci, F. Nitsche, H. Myers, J. Elliott, T. Bauch, N. Raqueno, and J. Frucci, “A comparative evaluation of uav-based remote sensing and geophysical techniques for landmine detection on a seeded m...

    2026

  6. [14]

    Human- in-the-loop signature bootstrapping for uav hyperspectral pfm-1 mine detection,

    S. Lekhak, P. R. Pulakurthi, and E. J. Ientilucci, “Human- in-the-loop signature bootstrapping for uav hyperspectral pfm-1 mine detection,” 2026. [Online]. Available: https: //arxiv.org/abs/2607.25310

    2026 arXiv

  7. [15]

    How to implement drones and machine learning to reduce time, costs, and dangers associated with landmine detection,

    J. Baur, G. Steinberg, A. Nikulin, K. Chiu, and T. de Smet, “How to implement drones and machine learning to reduce time, costs, and dangers associated with landmine detection,” 2021. [Online]. Available: https://commons.lib.jmu.edu/cisr-journal/ vol25/iss1/29

    2021

  8. [16]

    Deep Learning-Based Real-Time Detection of Surface Landmines Using Optical Imaging,

    E. Vivoli, M. Bertini, and L. Capineri, “Deep Learning-Based Real-Time Detection of Surface Landmines Using Optical Imaging,”Remote Sensing 2024, Vol. 16,, vol. 16, no. 4, 2

    2024

  9. [17]

    Joint fusion and detection via deep learning in uav-borne multispectral sensing of scatterable landmine,

    Z. Qiu, H. Guo, J. Hu, H. Jiang, and C. Luo, “Joint fusion and detection via deep learning in uav-borne multispectral sensing of scatterable landmine,”Sensors, vol. 23, no. 12, 2023. [Online]. Available: https://www.mdpi.com/1424-8220/23/12/5693

    2023

  10. [18]

    Available: https://www.mdpi.com/2072-4292/ 16/4/677

    [Online]. Available: https://www.mdpi.com/2072-4292/ 16/4/677

  11. [19]

    Domain adaptation in remote sensing image classification: A survey,

    J. Peng, Y . Huang, W. Sun, N. Chen, Y . Ning, and Q. Du, “Domain adaptation in remote sensing image classification: A survey,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 15, pp. 9842–9859, 2022

    2022

  12. [20]

    Deep learning based domain adaptation methods in remote sensing: A comprehensive survey,

    S. Lyu, Q. Zhao, Z. Zhou, M. Li, Y . Zhou, D. Yao, G. Cheng, H. Zhou, and Z. Shi, “Deep learning based domain adaptation methods in remote sensing: A comprehensive survey,”arXiv preprint arXiv:2510.15615, 2025

    2025

  13. [21]

    Benchmarking object de- tectors under real-world distribution shifts in satellite imagery,

    S. A. Al-Emadi, Y . Yang, and F. Ofli, “Benchmarking object de- tectors under real-world distribution shifts in satellite imagery,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 8299–8309

    2025

  14. [22]

    Domain adaptation for object detection based on domain-aware prompting in re- mote sensing imagery,

    X. Zhang, X. Yao, G. Cheng, and J. Han, “Domain adaptation for object detection based on domain-aware prompting in re- mote sensing imagery,”IEEE Transactions on Geoscience and Remote Sensing, vol. 64, pp. 1–16, 2026

    2026

  15. [23]

    A survey of landmine detection using hyperspectral imaging,

    I. Makki, R. Younes, C. Francis, T. Bianchi, and M. Zucchetti, “A survey of landmine detection using hyperspectral imaging,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 124, pp. 40–53, 2 2017

    2017

  16. [24]

    Identifying label errors in ob- ject detection datasets by loss inspection,

    M. Schubert, T. Riedlinger, K. Kahl, D. Kr ¨oll, S. Schoenen, S. ˇSegvi´c, and M. Rottmann, “Identifying label errors in ob- ject detection datasets by loss inspection,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2024, pp. 4582–4591

    2024

  17. [25]

    Benchmarking Deep Learning and Statistical Target Detection Methods for PFM-1 Landmine Detection in UA V Hyperspectral Imagery,

    S. Lekhak, P. R. Pulakurthi, R. Bhatta, and E. J. Ientilucci, “Benchmarking Deep Learning and Statistical Target Detection Methods for PFM-1 Landmine Detection in UA V Hyperspectral Imagery,” 2 2026. [Online]. Available: http://arxiv.org/abs/2602. 10434

    2026

  18. [26]

    A UA V-Based VNIR Hyperspectral Benchmark Dataset for Landmine and UXO Detection,

    S. Lekhak, E. J. Ientilucci, J. Baur, and S. Ghosh, “A UA V-Based VNIR Hyperspectral Benchmark Dataset for Landmine and UXO Detection,” 2 2026. [Online]. Available: http://arxiv.org/abs/2510.02700

    2026

  19. [27]

    Faster r-cnn: Towards real-time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” in Advances in Neural Information Processing Systems, C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, Eds., 20 vol. 28. Curran Associates, Inc...

    2015

  20. [28]

    Automated Airborne Ordinance Detection Using Data Fusion of Magnetometer and Ground Penetrating Radar,

    K. Kuru, D. Ansell, D. Jones, B. Watkinson, J. M. Pinder, R. Hamila, and C. Tinker-Mill, “Automated Airborne Ordinance Detection Using Data Fusion of Magnetometer and Ground Penetrating Radar,” pp. 1–6, 11 2025. [Online]. Available: https://ieeexplore.ieee.org/abstract/documen...

    2025

  21. [29]

    Dota: A large-scale dataset for object detection in aerial images,

    G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang, “Dota: A large-scale dataset for object detection in aerial images,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3974–3983

    2018

  22. [30]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 779–788

    2016

  23. [31]

    Vision meets drones: A challenge,

    P. Zhu, L. Wen, X. Bian, H. Ling, and Q. Hu, “Vision meets drones: A challenge,” 2018. [Online]. Available: https://arxiv.org/abs/1804.07437

    2018 arXiv

  24. [32]

    The unmanned aerial vehicle bench- mark: Object detection and tracking,

    D. Du, Y . Qi, H. Yu, Y . Yang, K. Duan, G. Li, W. Zhang, Q. Huang, and Q. Tian, “The unmanned aerial vehicle bench- mark: Object detection and tracking,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 370–386

    2018

  25. [33]

    Applying deep learning to automate uav-based detection of scatterable landmines,

    J. Baur, G. Steinberg, A. Nikulin, K. Chiu, and T. S. de Smet, “Applying deep learning to automate uav-based detection of scatterable landmines,”Remote Sensing, vol. 12, no. 5,

  26. [34]

    Multi-scale object detection in remote sensing imagery with convolutional neural networks,

    Z. Deng, H. Sun, S. Zhou, J. Zhao, L. Lei, and H. Zou, “Multi-scale object detection in remote sensing imagery with convolutional neural networks,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 145, pp. 3–22, 2018, deep Learning RS Data. [Online]. Available: https://...

    2018

  27. [35]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” inEuropean conference on computer vision. Springer, 2020, pp. 213–229

    2020

  28. [36]

    Sparse r- cnn: End-to-end object detection with learnable proposals,

    P. Sun, R. Zhang, Y . Jiang, T. Kong, C. Xu, W. Zhan, M. Tomizuka, L. Li, Z. Yuan, C. Wanget al., “Sparse r- cnn: End-to-end object detection with learnable proposals,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 14 454–14 463

    2021

  29. [37]

    Comparing surface landmine object detection models on a new drone flyby dataset,

    N. Agrawal-Chung and Z. Moin, “Comparing surface landmine object detection models on a new drone flyby dataset,” 2024. [Online]. Available: https://arxiv.org/abs/2410.19807

    2024 arXiv

  30. [38]

    Uncertainty Quantification in Surface Landmines and UXO Classification using MC Dropout,

    S. Lekhak, E. J. Ientilucci, D. Dera, and S. Ghosh, “Uncertainty Quantification in Surface Landmines and UXO Classification using MC Dropout,” pp. 1177–1181, 11 2025. [Online]. Available: https://ieeexplore.ieee.org/document/11243138

    2025

  31. [39]

    AMLID: An Adaptive Multispectral Landmine Identification Dataset for Drone-Based Detection,

    J. E. Gallagher and E. J. Oughton, “AMLID: An Adaptive Multispectral Landmine Identification Dataset for Drone-Based Detection,” 12 2025. [Online]. Available: http://arxiv.org/abs/ 2512.18738

    2025

  32. [40]

    Varifo- calnet: An iou-aware dense object detector,

    H. Zhang, Y . Wang, F. Dayoub, and N. Sunderhauf, “Varifo- calnet: An iou-aware dense object detector,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 8514–8523

    2021

  33. [41]

    Yolov8: A novel object detection algorithm with enhanced performance and robustness,

    R. Varghese and S. M., “Yolov8: A novel object detection algorithm with enhanced performance and robustness,” in2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS), 2024, pp. 1–6

    2024

  34. [42]

    Yolov11: An overview of the key architectural enhancements,

    R. Khanam and M. Hussain, “Yolov11: An overview of the key architectural enhancements,”arXiv preprint arXiv:2410.17725, 2024

    2024 arXiv

  35. [43]

    MineInsight: A Multi-Sensor Dataset for Humanitarian Demining Robotics in Off-Road Environments,

    M. Malizia, C. Hamesse, K. Hasselmann, G. De Cubber, N. Tsiogkas, E. Demeester, and R. Haelterman, “MineInsight: A Multi-Sensor Dataset for Humanitarian Demining Robotics in Off-Road Environments,”IEEE Robotics and Automation Let- ters, vol. 11, no. 2, pp. 1650–1657, 2026. [On...

    2026

  36. [44]

    Yolo26: key architectural enhancements and performance benchmarking for real-time object detection,

    R. Sapkota, R. H. Cheppally, A. Sharda, and M. Karkee, “Yolo26: key architectural enhancements and performance benchmarking for real-time object detection,”arXiv preprint arXiv:2509.25164, 2025

    2025

  37. [45]

    Yolo-world: Real-time open-vocabulary object detection,

    T. Cheng, L. Song, Y . Ge, W. Liu, X. Wang, and Y . Shan, “Yolo-world: Real-time open-vocabulary object detection,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16 901–16 911

    2024

  38. [46]

    Yolov12: Attention-centric real-time object detectors,

    Y . Tian, Q. Ye, and D. Doermann, “Yolov12: Attention-centric real-time object detectors,”Advances in neural information processing systems, vol. 38, pp. 78 433–78 457, 2026

    2026

  39. [47]

    D-fine: Redefine regression task in detrs as fine-grained distribution refinement,

    Y . Peng, H. Li, P. Wu, Y . Zhang, X. Sun, and F. Wu, “D-fine: Redefine regression task in detrs as fine-grained distribution refinement,” 2024

    2024

  40. [48]

    Rf-detr: neural architecture search for real-time de- tection transformers,

    I. Robinson, P. Robicheaux, M. Popov, D. Ramanan, and N. Peri, “Rf-detr: neural architecture search for real-time de- tection transformers,”arXiv preprint arXiv:2511.09554, 2025

    2025

  41. [49]

    Detrs beat yolos on real-time object detection,

    Y . Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y . Liu, and J. Chen, “Detrs beat yolos on real-time object detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 16 965–16 974

    2024

  42. [50]

    Label Studio: Data labeling software,

    M. Tkachenko, M. Malyuk, A. Holmanyuk, and N. Liubimov, “Label Studio: Data labeling software,” 2020-2025, open source software available from https://github.com/HumanSignal/label- studio. [Online]. Available: https://github.com/HumanSignal/ label-studio

    2020

  43. [51]

    Microsoft COCO: Com- mon objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ra- manan, P. Doll ´ar, and C. L. Zitnick, “Microsoft COCO: Com- mon objects in context,” inEuropean Conference on Computer Vision (ECCV), 2014, pp. 740–755. Sagar Lekhakreceived the B.E. degree in Electron- ics and Com...

    2014

  44. [52]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255

    2009

  45. [2020]

    Available: https://www.mdpi.com/2072-4292/ 12/5/859

    [Online]. Available: https://www.mdpi.com/2072-4292/ 12/5/859

  46. [2024]

    Available: https://www.mdpi.com/2072-4292/ 16/24/4732

    [Online]. Available: https://www.mdpi.com/2072-4292/ 16/24/4732

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.