{"id":"a6392732-8697-44c3-af25-4e0015e62af5","arxiv_id":"2412.05871","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MID is a new OBB-annotated ship detection dataset focused on dense occlusion and interaction scenarios, with baseline results across ten YOLO-family detectors.","lead":"This paper introduces MID, a shore-based maritime dataset of 5,673 images with 135,884 ship instances labeled with oriented bounding boxes, covering dense traffic, occlusions, and varied weather. It also reports detection baselines for ten YOLO variants to show how current models perform on these hard scenes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never accounts for the selection of 5,673 annotated images from 15,050 extracted frames; without that accounting, the dataset's diversity statistics may not support the claimed real-world distribution.","rationale":"The reader's weakest assumption is exactly the unexplained gap between 15,050 extracted frames and 5,673 annotated images. This is the most load-bearing flaw because nearly every downstream claim about the dataset's value as a benchmark depends on the final image set being representative of the claimed real-world scenarios. The paper provides no selection procedure, so the reported diversity statistics could reflect curator preference rather than environmental frequency. This is addressable in revision by documenting the filtering/subsampling procedure and providing evidence that the retained images are representative. The concern does not overturn the dataset's potential value, but it does justify the conditional verdict: the dataset should not be fully accepted as a benchmark until the selection step is clarified and verified.","tokens_in":12454,"tokens_out":2832,"duration_ms":30307,"concrete_test":"Release the frame-to-image mapping or selection script that identifies which of the 15,050 extracted frames became the 5,673 annotated images. Then independently annotate a stratified random sample of the 15,050 frames (e.g., 500 frames drawn proportionally across the 43 videos) using the same OBB annotation guidelines, and compare per-image instance density, occlusion proportion, and weather/background distributions against the published 5,673-image statistics. If the selected subset differs by more than roughly 20% relative in these quantities, the representative-diversity claim fails; if the distributions are statistically equivalent, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.A states that one frame was extracted every 176 frames (about every 6 seconds) from 43 video segments, producing 350 images per video and 15,050 original images total. The abstract and dataset description, however, report only 5,673 images. Nowhere in the manuscript is the reduction from 15,050 to 5,673 described: no filtering criteria, no random subsampling procedure, no exclusion policy for blurry or empty frames, and no discussion of whether the retained images were deliberately curated. This is load-bearing because the central claim that MID reflects busy-port real-world distributions rests on the representativeness of the final image set. The paper's own language suggests curation rather than random sampling: Section III.B.3 says 'we collected as much occluded data as possible,' and the abstract mentions 'manually supplemented annotations.' If the 5,673 images were selected to emphasize dense, occluded, or easily annotatable scenes, then the reported statistics (23.95 instances per image on average, 16% occlusion rate, weather and scale distributions) may substantially overstate how often those conditions occur in the original monitoring stream. The diversity and applicability claims would then be true of a curated subset, not of the shore-based maritime environment as encountered in practice. Because the selection step is completely unspecified, the dataset's advertised composition cannot be independently validated or reproduced.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces MID, a shore-based optical maritime dataset with 5,673 images and 135,884 oriented-bounding-box (OBB) annotations derived from 43 video segments recorded by port surveillance cameras. The paper documents dataset organization, analyzes diversity in terms of weather, scale, aspect ratio, background, occlusion, and collision scenarios, and reports detection baselines for ten YOLO-family variants with OBB heads. The central claim is that MID fills a gap in existing ship datasets by providing dense, occluded, multi-scale real-world maritime interactions and therefore supports detection, tracking, and trajectory-prediction research.","tokens_in":12696,"tokens_out":6063,"duration_ms":55780,"significance":"If the dataset is released as described, MID is a potentially valuable community resource: shore-based OBB video-derived data with temporal ordering is scarce, and the internal statistics are arithmetically consistent (the instance counts in Tables II and III sum to 135,884, and 135,884/5,673 = 23.95). The paper also provides reproducible-looking OBB conversion tooling and a public release plan. The significance is conditional, however, on resolving the undocumented gap between the 15,050 extracted frames and the 5,673 released images, and on clarifying the definitions behind the scale and occlusion statistics.","major_comments":[{"comment":"The paper states in Section III.A that 43 video segments, each sampled at one frame per 176 frames, yield 350 images per video and 15,050 original images, yet the released dataset and all subsequent statistics use 5,673 images. No filtering, exclusion, or subsampling procedure is described anywhere in the manuscript. This is load-bearing because the dataset's claims to reflect real-world shore-based maritime distributions rest on the representativeness of the final image set; the phrase in Section III.B.3, 'we collected as much occluded data as possible,' and the abstract's mention of 'manually supplemented annotations' suggest curation rather than random sampling. Please specify exactly how 5,673 images were obtained from 15,050 frames, report any exclusion criteria (e.g., empty frames, blur, annotation difficulty), and state whether selection was randomized and, if so, with what seed or protocol.","section":"III.A and IV.B"},{"comment":"The reported recall of 0.917 for YOLOv10s-obb head is a striking outlier: every other model in the table has recall between 0.688 and 0.722, including YOLOv11s-obb, which has the same mAP50 of 84.9. No explanation or experimental note accompanies this value, and it is unlikely to be correct as reported. Because the baseline comparison is one of the paper's central evaluation claims, please verify the YOLOv10s result, report corrected numbers, and, ideally, include variance over multiple runs or seeds.","section":"Table V"},{"comment":"The scale categories in Table II are labeled only as pixel thresholds (e.g., 'Tiny Instances ≤ 16 pixels'), without specifying whether the threshold refers to bounding-box area, long side, short side, or some other quantity. The resulting distribution—65,748 tiny instances and 66,984 extra-large instances, with almost no instances in the intermediate bins—is surprising for a dataset advertised as multi-scale and needs explanation. If the threshold is on area, 16 square pixels is far below a plausible annotatable ship size; if it is on side length, the units are unspecified. Please define the scale measure and discuss the apparent bimodality, which may also indicate that the 'tiny' and 'extra-large' bins are not measuring what the text implies.","section":"Table II and Section IV.B"},{"comment":"The paper's central occlusion contribution lacks a formal definition. Section III.B.3 states that the authors 'annotate both the visible parts of the hull and the obscured sections at different visibility ratios,' but each object has a single OBB; it is unclear how a single box can encode both visible and occluded portions or how the occlusion percentage in Table III is computed. Section IV.F says 16% of the dataset contains occlusion, but Table III reports instance counts, not image counts. Please define the occlusion ratio, describe the annotation protocol for occluded targets (e.g., is the box drawn around the full extent or only the visible part?), and report occlusion statistics at both image and instance levels.","section":"III.B.3 and IV.F"}],"minor_comments":[{"comment":"The weather categories 'silty,' 'fuzzy,' and 'color-distorted' are non-standard and are not defined; please replace them with standard meteorological or visual categories or provide quantitative criteria.","section":"IV.A"},{"comment":"The sentence '16% of the dataset contains varying levels of occlusion' should say '16% of instances' if it refers to Table III, or should be recomputed at image level.","section":"IV.F"},{"comment":"The column 'Time Dimension Year' is confusing: the checkmark for Ours appears to refer to the video/frame ID naming convention rather than to temporal annotations; please clarify what is being compared.","section":"Table VI"},{"comment":"The baseline experiments evaluate models trained and tested on MID only; a cross-dataset evaluation (e.g., fine-tune on MID and test on HRSID, HRSC2016, or SeaShip) would substantiate the claim that MID improves generalization to real-world complex scenes.","section":"V"},{"comment":"Annotation quality is stated to be ensured by four experienced annotators over three months, but no inter-annotator agreement, quality-control, or re-check procedure is described; a brief protocol statement would be valuable for dataset reliability.","section":"III.B"},{"comment":"Reference [30] for YOLO is incomplete (missing co-authors and publication details), and there are occasional formatting inconsistencies in the reference list (e.g., incomplete venue names).","section":"References"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The main unresolved issue is not technical in a narrow sense but is central to the dataset's validity—the undocumented reduction from 15,050 extracted frames to 5,673 released images makes the reported diversity statistics impossible to verify. This is fixable with a clear protocol description, and the dataset itself appears potentially useful, so I do not recommend rejection. Please also ensure the dataset release includes a license and a privacy or ethics statement, since the images come from operational port surveillance cameras."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the MID dataset paper. Bottom line: it's a potentially useful resource for shore-based ship detection with oriented boxes, but the paper as written has one load-bearing gap that needs fixing before the dataset's advertised diversity claims can be trusted.\n\nWhat's genuinely new: MID is, as far as I can tell, the first shore-based optical ship dataset with OBB annotations (SeaShip uses HBB; HRSC2016 and ShipRSImageNet are remote sensing). The temporal naming (video ID + frame ID) is a small but real addition for tracking work. The annotation effort is substantial — 135,884 instances across 5,673 images, with per-image instance counts up to 81, and the statistics in Tables II and III are internally consistent. The baseline evaluation of ten YOLO variants is a useful service to the community, even if it's just a reference point.\n\nThe soft spots:\n\n1. The 15,050 → 5,673 gap. The paper says 43 videos × 350 frames = 15,050 images, then the dataset has 5,673. No filtering or selection procedure is described anywhere. This matters because the paper's claim to represent 'busy-port real-world distributions' rests on the final image set. If the authors curated frames to emphasize dense, occluded scenes — and their own language in Section III.B.3 suggests they did ('we collected as much occluded data as possible') — then the reported 23.95 instances/image average and 16% occlusion rate may not reflect the actual monitoring stream. This needs a paragraph describing exactly how the 5,673 were chosen.\n\n2. Table V anomaly. YOLOv10s-obb head shows recall 0.917 with precision 0.933, while every other model sits around 0.70-0.72 recall. That looks like a typo (maybe mAP or precision value copied), but as printed it undermines confidence in the table.\n\n3. The 'fills the gap' claim is asserted, not demonstrated. No cross-dataset experiment (e.g., train on MID, test on SeaShip or vice versa) to show MID is harder or more realistic. That's the kind of evidence that would make the dataset's value concrete.\n\n4. Minor: baselines are single-run without variance; the dataset link isn't confirmed live; the occlusion degree thresholds (10%, 20%, 50%, 90%) are stated without justification.\n\nNone of these are fatal. The unexplained frame reduction is the one that must be addressed; the rest are fixable in revision. The dataset itself deserves a serious referee — if the selection process is clarified, this could be a solid benchmark paper for the maritime vision community.\n\nI'd bring it to reading group once the selection question is resolved, and I'd cite it if I worked on shore-based maritime detection. Recommend peer review, conditional on the authors accounting for the frame selection.\n\nCheers.","headline":"Potentially valuable shore-based OBB ship dataset, but the unexplained 15,050-to-5,673 frame selection gap must be addressed before the diversity claims are credible.","tokens_in":13242,"tokens_out":2801,"would_cite":false,"duration_ms":26754,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MID, a new shore-based dataset with 135,884 oriented-box ship instances, gives busy-port detection a realistic occlusion-heavy benchmark that satellite and SAR datasets lack.","keywords":["oriented bounding box","ship detection","optical shore-based dataset","dense occlusion","small object detection","maritime situational awareness","video-derived dataset","YOLO baselines"],"falsifier":"Annotate all 15,050 extracted frames, or a random sample of the 9,377 frames that MID leaves out, and compare their instance counts, occlusion rates, weather conditions, and video coverage with MID's 5,673 images; a systematic mismatch would show that MID's diversity statistics describe the chosen subset, not the captured navigation scenes.","tokens_in":12236,"feed_emoji":"🚢","tokens_out":7723,"duration_ms":70328,"temperature":0.7,"pith_summary":"The paper introduces MID, a shore-based optical dataset of 5,673 images annotated with oriented bounding boxes, containing 135,884 ship instances extracted from 43 real navigation videos. It argues that existing ship datasets—chiefly satellite or SAR-based, horizontal-box, or single-task datasets such as HRSID, SSDD, NWPU-10, HRSC2016, ShipRSImageNet, and SeaShip—do not cover dense occlusion, small-target clustering, and complex ship interactions in busy ports. MID is designed to fill that gap by encoding eight weather conditions, wide scale and aspect-ratio variation, multiple viewpoints, graded occlusion levels, and crossing, overtaking, and head-on encounters. The paper evaluates ten YOLO-family detectors with oriented-box heads as baselines, intended to serve as a reference for future work on maritime situational awareness, tracking, and trajectory prediction.","feed_headline":"Ship dataset supplies 135,884 dense and occluded targets","feed_subtitle":"Shore-based OBB images from 43 real navigation videos give detectors a harder busy-port testbed.","key_machinery":"The central object is the oriented bounding box (OBB): a rotated rectangle specified by four corner coordinates, chosen over axis-aligned boxes because ships appear at arbitrary headings and dense pixel overlap makes horizontal boxes inaccurate. Equally load-bearing is the dataset's annotation and organization scheme—each image is tied to a video ID and frame ID extracted at one frame every 176 frames (roughly every 6 seconds) from 1920×1080 shore-mounted cameras, giving a temporal ordering that detection alone would not provide. Around this scheme, the paper builds a difficulty taxonomy (weather, scale, aspect ratio, occlusion degree, background, and collision type) that lets the dataset be sliced into focused test conditions, and it runs ten YOLO-family detectors with OBB heads under fixed training settings to supply reference numbers.","core_discovery":"MID is a video-derived optical dataset whose images carry a time dimension (video ID and frame ID) and point-based oriented bounding box annotations in the form of four corner coordinates. The dataset's defining claim is that dense occlusion and interaction-rich scenes, not just clean single-ship views, are the norm in real port monitoring: 16% of instances are at least slightly occluded, 837 instances are almost fully or fully occluded, and roughly half of all instances are tiny (at most 16 pixels) while the other half are extra-large (above 256 pixels). By including these cases alongside rain, fog, lens water droplets, overexposure, and multiple camera viewpoints, the authors aim to provide a harder and more realistic training and evaluation ground than existing datasets, one that supports both supervised and semi-supervised learning and downstream tasks such as tracking, trajectory extraction and prediction, and traffic information analysis. The baseline runs of ten YOLO variants with oriented-box heads are presented as first reference results on this benchmark.","pith_inferences":["Because the paper does not say how the 5,673 annotated images were selected from the 15,050 extracted frames, the dataset's diversity statistics implicitly assume that this subset represents the full video corpus; releasing the selection rule or all frames would let users test that assumption.","The occlusion labels include fully occluded instances, which are not visible; this opens a route to evaluating track-based re-identification or weakly supervised detection that the paper does not develop.","All data come from one port area during 10 days in March, so generalizations to other seasons, regions, or port layouts are plausible but untested; a straightforward check is training on MID and testing on a second port's footage.","The time dimension plus OBB annotations could support a unified detection-and-tracking benchmark with occlusion-conditioned metrics, a construction the paper leaves for future versions."],"forward_implications":["Detectors trained on MID should transfer better to crowded port and narrow-channel monitoring than models trained on satellite, SAR, or single-target datasets, because the training distribution includes occluded, tiny, and overlapping ships.","The video ID and frame ID naming makes MID usable for tracking and trajectory extraction without extra alignment, directly supporting speed estimation and ship counting.","The graded occlusion annotations let researchers measure how detection performance degrades as occlusion increases, and provide a test set for occlusion-aware detectors.","The extreme scale distribution—about half tiny and half extra-large instances—stresses multiscale detectors and makes the dataset a demanding benchmark for small-target detection.","The fixed training settings and ten baseline configurations provide a reproducible comparison point for future oriented-box detectors."],"supporting_citations":[{"why":"HRSID, the high-resolution SAR ship detection dataset, is the main comparison point for the claim that existing datasets lack dense occlusion and interaction coverage.","marker":"[18]"},{"why":"SSDD supplies the SAR-domain baseline that MID contrasts with, supporting the argument that optical shore-based data fills a gap.","marker":"[19]"},{"why":"NWPU-10 (object detection in remote sensing imagery) is cited with [21] as an existing dataset that does not address complex maritime interactions.","marker":"[20]"},{"why":"The HQ-ISNet companion to NWPU-10 extends instance segmentation in remote sensing and is used as a second marker for the comparison group.","marker":"[21]"},{"why":"HRSC2016 provides the oriented/polygon ship annotation baseline in remote sensing that MID's shore-based OBB annotations are compared against.","marker":"[22]"},{"why":"ShipRSImageNet sets the large-scale remote sensing ship detection and classification baseline that MID distinguishes itself from in scenario coverage.","marker":"[23]"},{"why":"SeaShip is the prior optical shore-based ship dataset whose scope MID extends with weather, occlusion, and interaction scenarios.","marker":"[25]"},{"why":"YOLO is the detector family whose variants, fitted with oriented-box heads, provide the baseline results on MID.","marker":"[30]"}],"fun_headline_variants":["New ship dataset targets dense occlusions with Oriented Boxes","Dense ship occlusion dataset: 135,884 OBB instances from real videos","Occlusion-heavy ship benchmark: 5,673 images, 135,884 boxes","Real navigation videos yield 135k oriented ship boxes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 5,673 annotated images fairly represent the 15,050 frames extracted from the 43 videos, yet the paper gives no selection or filtering procedure between the two sets.","fun_headline_variants_meta":{"raw":{"variants":["New ship dataset targets dense occlusions with Oriented Boxes","Dense ship occlusion dataset: 135,884 OBB instances from real videos","Occlusion-heavy ship benchmark: 5,673 images, 135,884 boxes","Real navigation videos yield 135k oriented ship boxes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000768,"raw_usage":{"total_tokens":3431,"prompt_tokens":997,"completion_tokens":2434,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":2355}},"tokens_in":613,"tokens_out":2434,"duration_ms":17564,"temperature":1.0,"reasoning_tokens":2355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:15:26.472038+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Annotate all 15,050 extracted frames, or a random sample of the 9,377 frames that MID leaves out, and compare their instance counts, occlusion rates, weather conditions, and video coverage with MID's 5,673 images; a systematic mismatch would show that MID's diversity statistics describe the chosen subset, not the captured navigation scenes.","supporting_citations":[{"cited_title":"Sar ship detection dataset (ssdd): Official release and comprehensive data analysis,","cited_arxiv_id":null,"evidence_quote":"SSDD supplies the SAR-domain baseline that MID contrasts with, supporting the argument that optical shore-based data fills a gap."},{"cited_title":"Object detection and instance segmentation in remote sensing imagery based on precise mask r-cnn,","cited_arxiv_id":null,"evidence_quote":"NWPU-10 (object detection in remote sensing imagery) is cited with [21] as an existing dataset that does not address complex maritime interactions."},{"cited_title":"Hq- isnet: High-quality instance segmentation for remote sensing imagery,","cited_arxiv_id":null,"evidence_quote":"The HQ-ISNet companion to NWPU-10 extends instance segmentation in remote sensing and is used as a second marker for the comparison group."},{"cited_title":"A high resolution optical satellite image dataset for ship recognition and some new baselines,","cited_arxiv_id":null,"evidence_quote":"HRSC2016 provides the oriented/polygon ship annotation baseline in remote sensing that MID's shore-based OBB annotations are compared against."},{"cited_title":"Shiprsimagenet: A large-scale fine-grained dataset for ship detection in high-resolution optical remote sensing images,","cited_arxiv_id":null,"evidence_quote":"ShipRSImageNet sets the large-scale remote sensing ship detection and classification baseline that MID distinguishes itself from in scenario coverage."},{"cited_title":"Seaships: A large-scale precisely annotated dataset for ship detection,","cited_arxiv_id":null,"evidence_quote":"SeaShip is the prior optical shore-based ship dataset whose scope MID extends with weather, occlusion, and interaction scenarios."},{"cited_title":"You only look once: Unified, real-time object detection,","cited_arxiv_id":null,"evidence_quote":"YOLO is the detector family whose variants, fitted with oriented-box heads, provide the baseline results on MID."}],"review_version":1}