Frame-to-Panorama Localization and Context-Aware Sampling for Scene-Specific Ship Detection in a Smart Marina Testbed
Ignat Romanov, Andreas Hadjipieris, Neofytos Dimitriou
cs.CV · cs.AI · cs.CG
Quick verdict
Pragmatic metadata-free PTZ data-curation pipeline cuts 40,718 frames to 220; grouped-CV YOLO26 reaches AP50 94.78% but sampling gains lack a matched baseline.
Summary
This paper presents an end-to-end data-curation pipeline for adapting ship detectors to a fixed smart-marina PTZ camera whose historical footage lacks reliable pan, tilt, and zoom metadata. The authors manually stitch a reference panorama in Photoshop, then localize 40,718 daylight frames onto it using SuperPoint keypoints, LightGlue matching, and USAC-MAGSAC homography estimation, yielding projected frame centers and coverage-based zoom proxies. Frames are enriched with hourly weather observations and solar-state labels, stratified into horizontal bins, and reduced via Kennard-Stone diversity sampling to 2,402 images. A second stage crops a YOLO11-n-detected horizon region into overlapping tiles, extracts ATSS-SwinL-DyHead embeddings, and retains frames through BIC-selected GMM clusters, reaching 220 images for annotation — a 99.5% reduction. The annotated set (4,017 boxes) fine-tunes a YOLO26-m detector under sequence-grouped five-fold outer cross-validation (182 sequences), achieving AP50 94.78%±0.51% and AP50-95 75.10%±1.73%. Localization is benchmarked against SIFT+FLANN as 20-bin classification (87.36% vs 12.34% accuracy), though a macro F1 of 62.83% exposes weakness on rare view bins. Main limitations: no matched-budget baseline showing the curated subset outperforms random sampling, single-camera single-scene daylight scope, manual stitching hindering reproducibility, no code or data release, and a small evaluation set (44 test images per fold).
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A SuperPoint+LightGlue homography pipeline localizes metadata-less PTZ frames onto a manually stitched reference panorama at 87.36% accuracy on 20-bin spatial classification, versus 12.34% for SIFT+FLANN.
- Combining Kennard-Stone diversity sampling with GMM-based horizon-context filtering reduces 40,718 daylight frames to 220 images for annotation, a 99.5% reduction, while preserving camera-view and environmental diversity.
- A YOLO26-m detector fine-tuned on the 220-image subset achieves 94.78%±0.51% AP50 and 75.10%±1.73% AP50-95 under sequence-grouped five-fold outer cross-validation with no temporal leakage.
- Smart-marina infrastructure video streams can be converted into compact, scene-specific training sets that substantially reduce annotation effort.
Strengths
Leakage-aware detection evaluation protocol
Section II-D groups 182 acquisition sequences across five outer folds (44 test images each, 158/18 train/val), evaluates every image exactly once, fixes seed 42 and hyperparameters, and uses internal validation only for checkpoint selection — a genuinely leakage-resistant design; Table II reports per-fold results plus SD and a 95% t-interval (94.15–95.41% AP50).
Concrete, auditable data-reduction chain
The 99.5% reduction is decomposed into reproducible stages with intermediate counts: 40,718 → 2,402 (94.1%, Kennard-Stone, II-B) → 220 (GMM horizon-context filtering, II-C), yielding 4,017 boxes over 208 vessel images plus 12 negatives.
Principled two-stage sampling design
Honest failure-mode reporting
Weaknesses
No matched-budget sampling control
Headline localization accuracy masks per-bin weakness
Narrow scope, no external validation
Reproducibility gaps
Evaluation distribution equals curation distribution
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper addresses a narrow but genuinely operational problem: historical PTZ footage from the CMMI MDigi-I Smart Marina testbed (Ayia Napa, Cyprus; camera ~100 m above water; one frame per minute; 64,000 frames recorded between 1 August 2025 and 10 January 2026, of which 40,718 daylight frames are used) carries no reliable pan/tilt/zoom metadata, so training-set construction cannot be stratified by camera pose — especially since up to 90% of footage comes from the default park position. Section I lists four contributions: demonstrating the testbed as an AI development platform, a frame-to-panorama localization pipeline (SuperPoint + LightGlue + USAC-MAGSAC), a two-stage sampling methodology achieving a 99.5% data reduction, and a fine-tuned NMS-free YOLO26-m detector reaching 94.78% ± 0.51% AP50 under grouped cross-validation. Claims are consistently scoped to scene-specific adaptation — appropriately modest — though the abstract's headline numbers could be misread by casual readers as general maritime detection performance.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How much did this pipeline reduce the maritime video dataset?
It reduced 40,718 daylight candidate frames from a smart-marina PTZ camera to just 220 annotated images — a 99.5% reduction. This happened in two steps: Kennard-Stone diversity sampling cut the set to 2,402 images (94.1%), then GMM-based filtering of horizon tiles kept only frames likely to contain distant vessels, reaching 220 images with 4,017 labeled ship bounding boxes.
What detection accuracy did the fine-tuned YOLO26-m model achieve?
Under a sequence-grouped five-fold cross-validation (which prevents temporally adjacent frames from leaking between training and test sets), the model reached a mean AP50 of 94.78% ± 0.51% and a stricter AP50–95 of 75.10% ± 1.73%, with mean precision of 94.95% and recall of 89.68% for ship detection in the marina scene.
Is the code or dataset from this paper publicly available?
No. The paper contains no code or data availability statement. The footage comes from the private CMMI MDigi-I Smart Marina testbed in Cyprus, and the reference panorama was stitched manually in Adobe Photoshop, both of which limit independent reproduction.
Why did the authors need feature matching instead of using the camera's pan/tilt/zoom metadata?
The historical PTZ recordings simply lacked reliable pan, tilt, zoom, weather, and lighting metadata. So the authors recovered the camera view by matching each frame to a manually built reference panorama using SuperPoint keypoints, LightGlue matching, and homography estimation — turning the video archive into a spatially indexed dataset without camera telemetry.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–6 are referenced with captions but could not be visually analyzed from the extracted text (panorama example, coverage heatmap, GMM clusters, annotation sample, localization cases, live prediction).
- The abstract was truncated mid-number in the supplied feed; the full body text contains the complete results, so impact is minimal.
- No supplementary material or per-bin localization breakdown is available in the text, so claims about rare-bin behavior rest on the aggregate macro/weighted figures in Table I.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.