{"id":"305315ab-920a-4958-95e0-d919d0fa9941","arxiv_id":"2501.05567","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A YOLO-based object detector with an extra regression head estimates distances to maritime objects from monocular USV images, achieving mean errors of roughly 15 to 50 meters depending on range.","lead":"This paper adds a distance-estimation head to a YOLO object detector, letting a single camera on an unmanned boat guess how far away other boats and buoys are. The method is cheap and runs in real time, but the reported accuracy is approximate and the dataset is not yet public.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table II reports per-interval distance errors for the proposed YOLOv7 all ≥15.1 m, yet Table III claims an overall MDE of 14.9 m; since any aggregate mean must be a weighted average of interval means, this is arithmetically impossible and the central claim is unsupported.","rationale":"The reader's weakest assumption concerns GPS/chart-derived labels and camera sensitivity, which are valid transferability concerns. However, the reader's rationale also notes 'inconsistent reported errors across tables.' My stress test isolates this inconsistency as the most load-bearing issue because it directly contradicts the central empirical claim of a 14.9 m mean distance error. Table II's interval means are all at or above 15.1 m, so no weighted average of them can be 14.9 m. This is not a subtle statistical issue; it is an arithmetic impossibility under the reported evaluation protocol. If the aggregate MDE is wrong, the comparison with Depth Anything V2 and triangulation baselines in Table III is also compromised. The paper may have a legitimate explanation, such as different test splits or a typo, but without that explanation the current manuscript does not support its headline result. I therefore recommend rejecting the current version, subject to the authors providing a corrected or clarified evaluation. I do not claim the underlying approach is invalid; the concern is about the verifiability of the reported performance. The proposed concrete check would settle the matter quickly.","tokens_in":12548,"tokens_out":7771,"duration_ms":66677,"concrete_test":"Request the per-interval object counts used for Table II and recompute the overall MDE as a weighted average of the interval means using any nonnegative weights. If the minimum possible overall MDE exceeds 14.9 m, the reported value in Table III is inconsistent and the central claim cannot stand. Alternatively, re-run the evaluation on the same test split so that Table II and Table III are generated from identical detections; if the aggregate MDE still falls below every interval mean, the aggregation code contains an error.","verdict_should_be":"REJECT","load_bearing_attack":"Section V-B, Table II reports weighted distance errors by distance interval and object type. For the proposed YOLOv7, the minimum interval error is 15.1 m (buoys, 0-100 m) and the boat errors range from 18.2 m to 74.6 m across intervals, with averages of 47.5 m (boats) and 63.8 m (buoys). Section V-C, Table III then reports 'Proposed YOLOv7' with an MDE of 14.9 m and 12.0% outliers. Because any overall mean error must be a weighted average of the interval-specific mean errors, an overall MDE of 14.9 m is impossible when every interval mean is at least 15.1 m. The paper does not explain this discrepancy, for example by stating that the two tables are computed on disjoint test subsets or with different association criteria. The same issue affects YOLOv9-M: Table II's minimum interval error is 15.8 m, yet Table V reports 14.9 m with tracking. Unless the tables are computed on different sets with different protocols, the headline MDE is not supported. This is the most load-bearing concern because the paper's central claim rests on that specific number.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a monocular distance estimation method for USVs by appending a distance regression head to YOLOv7 and YOLOv9 object detectors. Training and evaluation labels are derived from GPS/chart data for static objects, with an additional 5,000 images labeled by human annotators. The method is evaluated on a 1,000-image maritime dataset and compared against a triangulation baseline and Depth Anything V2, reporting a mean distance error of 14.9 m at 45.2 FPS. The authors also present ablation studies on the distance loss gain and a tracking-based smoothing module.","tokens_in":12801,"tokens_out":6047,"duration_ms":52598,"significance":"If the claimed accuracy holds, the approach is a practical, low-cost alternative for approximate distance awareness on USVs, introducing no runtime overhead and requiring no additional sensors. The paper contributes a dataset and a simple architectural modification that could be of interest to the maritime computer vision community. However, the headline accuracy number is currently unsupported by the paper's own tables, and the comparison protocol has gaps that must be addressed before the central claim can be accepted.","major_comments":[{"comment":"The reported headline MDE of 14.9 m for 'Proposed YOLOv7' in Table III is arithmetically inconsistent with Table II. Table II lists interval-specific weighted distance errors for YOLOv7 with a minimum of 15.1 m (buoys, 0-100 m) and average errors of 47.5 m (boats) and 63.8 m (buoys). Since any overall mean must be a weighted average of interval means, no weighted average of values all ≥15.1 m can equal 14.9 m. The same problem affects Table V: YOLOv9-M without tracking has MDE 18.4 m, yet Table II's minimum interval error for YOLOv9-m is 15.8 m, so the reported MDE of 14.9 m with tracking is also impossible unless the two tables are computed on different test subsets or with different association criteria. The paper does not state such a difference. Because the central claim of Section V-C rests on this number, the authors must clarify the computation protocol or correct the tables.","section":"Section V-B, Table II; Section V-C, Table III; Section V-E, Table V"},{"comment":"The use of the two label sources is ambiguous. Section III-E says the 1,000-image evaluation dataset uses GPS/chart-derived distances for static objects, and adds the parenthetical remark that 'for our proposed method, we use manually labelled data.' Section III-F then introduces 5,000 additional images with human-annotated 'pseudo-distances.' It is never stated which labels train the distance head, which labels are used in Tables II-V, and whether the 5,000 human-labeled images are included in training, validation, or excluded. Table VI evaluates human labels against chart data but does not resolve this. If the model is trained or evaluated on human pseudo-labels while the baselines are compared on chart-derived ground truth, the comparison in Table III is not apples-to-apples. Please specify the exact training/evaluation label source and dataset split for every table.","section":"Section III-E, III-F, V-B, V-F"},{"comment":"The Depth Anything V2 baseline comparison is not interpretable as presented. Depth Anything V2 is a relative/affine-invariant monocular depth model; it does not directly output metric distances unless a metric variant or an alignment step (e.g., median or least-squares scaling to ground-truth depth) is applied. The paper states that 'dense distance values are projected into the bounding box by taking the median distance within a box' but gives no description of how absolute scale in meters was recovered. Without this, the MDE values of 43.5, 27.1, and 19.2 m cannot be reproduced or meaningfully compared with the proposed method's metric outputs. Please describe the alignment procedure or replace the baseline with a metric-depth method.","section":"Section V-C, Table III"},{"comment":"The final model configuration is underspecified. Table IV shows that the distance loss gain strongly affects results: MAE ranges from 19.7 m at gain 0.001 to 9.8 m at gain 0.1, while mAP@0.5 drops from 89.5% to 72.1%. Table III reports 14.9 m MDE for 'Proposed YOLOv7,' but the paper does not state which distance loss gain (if any) was used for the models in Tables I-III, nor does it reconcile the MAE values in Table IV with the MDE values elsewhere. Without this information, the reported performance is not reproducible.","section":"Section V-D, Table IV"}],"minor_comments":[{"comment":"There is an incomplete sentence: 'This formulation allows us' is followed directly by a new paragraph; the sentence should be completed or removed.","section":"Section III-A"},{"comment":"The text refers to 'NOAH' when the intended organization is presumably NOAA; please correct the typo.","section":"Section III-E"},{"comment":"The error metric is referred to as 'weighted distance error,' 'MAE,' and 'MDE' across the paper; please unify the terminology and define the relationship between these quantities.","section":"Tables III, IV, V"},{"comment":"The caption 'Dist. (×100)' is ambiguous; please clarify in the text that the intervals are 0-100 m, 100-200 m, etc.","section":"Table II"},{"comment":"The statement that 'Depth Anything v2 struggles with distances beyond approx. 60m' is not supported by any distance-binned result in Table III; either add supporting evidence or soften the claim.","section":"Section V-C"}],"recommendation":"major_revision","confidential_remarks":"The identical MDE and outlier values (14.9 m, 12.0%) in Tables III and V suggest a possible transcription error; please ask the authors to verify against raw experimental logs. Additionally, the manuscript does not state a train/validation/test split, which is needed to assess whether the reported results generalize. The arithmetic inconsistency between Tables II and III should be resolved before external review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nThe paper is a classic extension: add a distance regression head to YOLOv7/YOLOv9 and train on a new maritime dataset where labels come from GPS and NOAA chart positions. The genuinely useful piece is the dataset—1,000 images with bounding boxes and measured distances, plus 5,000 with pseudo-distances from human annotators. The normalization study (linear vs. log) and the real-time benchmarks on an Orin are also decent engineering. If the method works, it gives USVs approximate single-camera distance estimates at zero runtime cost, which is worth having.\n\nThe problem is the evaluation. The stress-test note is right: Table II gives per-interval weighted errors for YOLOv7 that bottom out at 15.1 m (buoys, 0–100 m) and average 47.5 m (boats) and 63.8 m (buoys). Table III then reports a 14.9 m overall MDE for the same model. A mean over the intervals cannot be lower than the smallest interval mean unless the test sets are different. The paper never says they are. That means the headline number is either wrong or computed on a different protocol, and the central claim in Section V-C is unsupported as written.\n\nThere are other soft spots: no error bars or repeated-run variance; the 5,000 human-labeled images are analyzed but never used in training, which is fine but under-explained; and the Depth Anything V2 comparison inherits the unfair median-in-box projection and only runs at 10 FPS or slower. The camera-FOV sensitivity is honestly stated in the limitations.\n\nSo where does this land? The dataset and the application are enough to justify a serious referee. The arithmetic inconsistency is fixable—either retract the 14.9 m number or define the disjoint test set. I'd send it to review with a request for clarification, not desk-reject it. But as it stands, I wouldn't cite the reported accuracy numbers.\n\nBest.","headline":"Practical maritime dataset plus a plausible distance head, but the headline MDE contradicts the per-interval table and is unsupported.","tokens_in":13322,"tokens_out":2866,"would_cite":false,"duration_ms":26736,"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":"Adding a distance regression head to a YOLO object detector lets a single camera estimate approximate distances to boats and buoys at real-time speed.","keywords":["unmanned surface vehicles","monocular distance estimation","object detection","YOLO","maritime dataset","distance regression head","GPS chart ground truth","real-time perception"],"falsifier":"Measure distances to moving boats with a laser rangefinder while running the model on the same frames; if the 14.9-meter mean error and 12 percent outlier rate do not reproduce for moving targets, the GPS-and-chart labelling assumption is the weak link and the central claim fails for the objects that matter most in collision avoidance.","tokens_in":12341,"feed_emoji":"🚤","tokens_out":6726,"duration_ms":60883,"temperature":0.7,"pith_summary":"This paper tries to establish that an unmanned surface vehicle can get approximate distances to nearby boats, buoys, and other waterborne objects from a single ordinary camera, by adding a distance prediction to an object detector instead of using LiDAR, radar, stereo vision, or depth cameras. The central claim is that a YOLO detector with one extra regression output per anchor, trained on images labelled with bounding boxes and distance measurements, produces mean distance errors around 14.9 meters at 45.2 frames per second on an embedded platform. The authors also claim this beats geometric triangulation and the Depth Anything V2 monocular depth model on their maritime data, while maintaining real-time speed. If correct, the method gives small boats a cheap proximity-alerting capability and makes any video with distance labels usable as training data.","feed_headline":"YOLO reads boat distances from one camera at 45 FPS","feed_subtitle":"Mean distance error of 14.9 m, beating triangulation and depth-estimation baselines on maritime data.","key_machinery":"The distance regression head is the carrying mechanism: one extra output neuron per anchor added to the YOLOv7 and YOLOv9 detector heads, trained with L1 loss on distances linearly normalized to [0,1] by d/d_max and inverted at inference. The head reuses the detector's existing visual features at no extra runtime cost, letting the network learn the size-to-distance and appearance-to-distance mapping directly from supervised labels; temporal consistency is then obtained by feeding SORT-tracked detections through a running average. The paper reports that only the basic linear normalization converged, while logarithmic and negative-range variants produced unstable training.","core_discovery":"The paper's central claim is that supervised object detection can double as an approximate range sensor. By appending a distance regression head to YOLOv7 and YOLOv9 detectors and training with an L1 loss on linearly normalized distances, the network learns to read distance from visual cues such as apparent object size, category, and scene context. On their collected maritime dataset, the modified YOLOv7 reaches a mean distance error of 14.9 meters with 12 percent outliers at 45.2 FPS, and the authors report that this is the lowest error and smallest outlier fraction among triangulation, Depth Anything V2 in three sizes, and the proposed method. The same architecture also supports a demonstration assistance system that projects detections with distances onto a 2D plane for operator alerts.","pith_inferences":["Beyond the paper: because the model learns a size-to-distance mapping, feeding it the camera's field of view or focal length as an extra input could make the same regression head transfer across cameras without retraining.","Beyond the paper: the human pseudo-labels are reliable only under about 100 meters, so a practical pipeline could train on large volumes of human-labelled close-range images and reserve scarce GPS or chart labels for calibration of farther ranges.","Beyond the paper: the same detector-plus-one-regression-output recipe could be applied to other object categories on other platforms, such as vehicles or UAVs, wherever approximate range warnings rather than precise metrology are the goal."],"forward_implications":["A single monocular camera plus a standard YOLO detector can act as a coarse proximity sensor for maritime collision alerts, with no extra sensor hardware and no added runtime cost.","Distance accuracy degrades with range: on buoys, weighted error grows from roughly 15 meters at 0 to 100 meters to about 130 meters at 400 to 500 meters, so the method is an alerting layer rather than precise ranging.","Temporal smoothing with SORT and a running average lowers mean error from 18.4 meters to 14.9 meters and the outlier rate from 22.5 percent to 12.0 percent, making video sequences noticeably more reliable than single frames.","The distance-loss gain is a tunable trade-off: a gain of 0.001 keeps detection mAP at 89.5 percent with distance MAE of 19.7 meters, while a gain of 0.1 improves MAE to 9.8 meters but drops mAP to 72.1 percent."],"supporting_citations":[{"why":"Supplies the YOLOv7 and YOLOv9 detector architectures whose heads receive the extra distance output.","marker":"[71]"},{"why":"Serves as the monocular depth-estimation baseline that the proposed method is compared against.","marker":"[6]"},{"why":"Provides the SORT tracking used with a running average to smooth distance estimates across frames.","marker":"[72]"},{"why":"Precedent for supervised object-specific distance estimation from monocular images, which this paper adapts to the maritime domain.","marker":"[68]"},{"why":"Documents the poor stereo accuracy beyond about 100 meters that motivates a monocular alternative.","marker":"[60]"},{"why":"Underpins the single-view-geometry triangulation baseline compared in the experiments.","marker":"[62]"},{"why":"Frames the sensor cost and complexity problems in unmanned surface vehicles that the proposed camera-only method addresses.","marker":"[1]"}],"fun_headline_variants":["YOLO reads boat distances from one camera","Boat distance via YOLO: 14.9 m mean error","Single-camera range sensing for USVs with YOLO","YOLO branch estimates distance on water from images","Detect and range: YOLO regression for maritime ops"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training and evaluation distances come from matching the boat's GPS position and heading to chart positions of static objects, so the reported accuracy depends on those chart associations being correct and on the deployment camera matching the training camera's field of view.","fun_headline_variants_meta":{"raw":{"variants":["YOLO reads boat distances from one camera","Boat distance via YOLO: 14.9 m mean error","Single-camera range sensing for USVs with YOLO","YOLO branch estimates distance on water from images","Detect and range: YOLO regression for maritime ops"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00035,"raw_usage":{"total_tokens":1862,"prompt_tokens":848,"completion_tokens":1014,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":930}},"tokens_in":464,"tokens_out":1014,"duration_ms":9720,"temperature":1.0,"reasoning_tokens":930,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:12:06.204985+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure distances to moving boats with a laser rangefinder while running the model on the same frames; if the 14.9-meter mean error and 12 percent outlier rate do not reproduce for moving targets, the GPS-and-chart labelling assumption is the weak link and the central claim fails for the objects that matter most in collision avoidance.","supporting_citations":[{"cited_title":"Simple online and realtime tracking,","cited_arxiv_id":null,"evidence_quote":"Provides the SORT tracking used with a running average to smooth distance estimates across frames."},{"cited_title":"Supervised object-specific distance estimation from monocular images for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Precedent for supervised object-specific distance estimation from monocular images, which this paper adapts to the maritime domain."},{"cited_title":"Know your limits: Accuracy of long range stereoscopic object measurements in practice,","cited_arxiv_id":null,"evidence_quote":"Documents the poor stereo accuracy beyond about 100 meters that motivates a monocular alternative."},{"cited_title":"Real-time vehicle distance estimation using single view geometry,","cited_arxiv_id":null,"evidence_quote":"Underpins the single-view-geometry triangulation baseline compared in the experiments."}],"review_version":1}