{"id":"b97e24de-1df8-4017-9800-10d76b4f2b35","arxiv_id":"2506.00154","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Adding an instance segmentation head to YOLOv11 improves detection of small, occluded marsh deer in UAV imagery compared with detection-only YOLOv11 and RT-DETR-l.","lead":"This study compares three deep learning models for spotting marsh deer in drone images, where the deer are tiny and often hidden by vegetation. It reports that a YOLOv11 model with an added segmentation head detects deer better than the detection-only version or RT-DETR.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed superiority of YOLOv11n-seg is not established: the test set is undefined, its labels may include 54 model-discovered deer, and the overlapping-specimen advantage rests on a single counting-matrix cell without sample size or error bars.","rationale":"The reader's weakest assumption correctly identifies the undefined test split as a central problem. I partially agree, and I would sharpen the concern in two ways. First, the test labels may not be independent of the models: Section III-A explicitly says the network discovered 54 deer that were then added to the final ground-truth count and segmentation masks, and Section IV-C reports these same 54 as a claimed success. If any of those model-found deer are in the test set, the reported metrics are inflated in a way that directly touches the central claim. Second, the specific evidence for YOLOv11n-seg's superiority is remarkably thin: one diagonal cell in the counting matrices, with no indication of how many images contain two deer and no error bars. This makes the conclusion vulnerable to small-sample noise, which is especially important because the qualitative difference in Figure 11 could easily be driven by a few images. These problems are addressable through data release and a clean re-evaluation, so the appropriate verdict remains CONDITIONAL rather than ACCEPT or REJECT. My concrete test targets both the circular-label concern and the small-sample concern in a single rerun.","tokens_in":9213,"tokens_out":3732,"duration_ms":40528,"concrete_test":"Ask the authors to publish the exact train/validation/test image IDs and a provenance flag for each deer annotation (expert-only vs. model-found-then-confirmed), then recompute Figures 9-11 on the test subset that excludes all images containing model-found deer, using thresholds fixed before inspecting that subset. Record the number of two-deer test images in the bottom-right cell and a bootstrap confidence interval for the difference between models. If the YOLOv11n-seg advantage vanishes, or the cell contains fewer than roughly 10 images, the superiority conclusion is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim: YOLOv11n-seg is superior because the segmentation masks help in overlapping specimens. For this to hold, the comparison must be on an independent, well-described test set and the metric delta must survive basic uncertainty. Both conditions fail. Section III-B-3 defines only training and validation sets; no test set composition or split is given. Section IV-A reports \"test set\" metrics with per-model confidence thresholds (0.255, 0.174, 0.164) selected on the validation set, but since the test set is not defined, independence cannot be verified. Section III-A states that during the work the network found 54 new deer later confirmed by experts, and these were folded into the final count of 285 deer and their masks; Section IV-C reports the same 54 as an increase. If any such model-found deer appear in test labels, the evaluation is contaminated by the detector being evaluated. Finally, the only evidence for the overlapping-specimen advantage is the bottom-right cell of the counting matrices in Figures 9-11, with no number of images in that cell, no confidence interval, and no repeated runs. With 262 deer images total, that cell could contain a handful of images, making the headline difference indistinguishable from noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper compares three object detectors—YOLOv11n-det, YOLOv11n-seg (with an instance segmentation head), and RT-DETR-l—for detecting marsh deer in high-resolution (1920x1920) UAV imagery from the Pantano Project in the Paraná Delta. The authors train on a dataset that includes fine-grained segmentation masks, a contribution they claim is novel for wildlife monitoring, and they report that YOLOv11n-seg achieves superior detection performance, particularly for overlapping specimens. The paper also reports that the network discovered 54 deer missed by initial human labeling, increasing the count from 231 to 285. The central conclusion is that YOLOv11n-seg is superior because the segmentation masks help in overlapping-specimen scenarios.","tokens_in":9395,"tokens_out":3465,"duration_ms":32636,"significance":"If the central claim holds, the paper would provide a useful practical result for UAV-based wildlife monitoring: that adding an instance segmentation head to a small YOLO model improves detection of small, occluded animals, and that the resulting model can assist human labelers in finding missed individuals. The dataset with precise segmentation masks is a potentially valuable resource for the community, and the comparison of a transformer-based detector (RT-DETR) with a lightweight fully convolutional detector at large input sizes is of independent interest. However, the significance is substantially tempered by evaluation weaknesses: the test set is not defined, the evaluation may be contaminated by model-discovered deer in the ground truth, and the headline advantage for overlapping specimens rests on a single counting-matrix cell without uncertainty quantification. The paper does not provide code or data availability, which limits reproducibility.","major_comments":[{"comment":"The test set is never defined. Section III-B-3 specifies only the training set (160 deer images, plus other classes) and a validation set (52 deer images, etc.), but Section IV-A reports metrics on a \"test set\" whose composition, size, and relationship to the training/validation splits are not given. Without a clear, independent test set, the reported superiority of YOLOv11n-seg over the other models cannot be verified or reproduced.","section":"Sections III-B-3 and IV-A"},{"comment":"The \"finding of unlabeled deer\" introduces a circularity risk. The manuscript states that the network found 54 deer that were later confirmed by experts and that the final count rose from 231 to 285, implying these model-proposed detections were added to the ground-truth labels. If any of these 54 deer appear in the test set used for evaluation—and the text does not rule this out—then the detector under test has contributed to its own ground truth, making the reported metrics partially circular. The authors must clearly state whether any of the 54 deer are present in the test-set labels and, if so, re-run the evaluation after removing them.","section":"Sections III-A and IV-C"},{"comment":"The claim that YOLOv11n-seg is superior in overlapping-specimen cases rests on a single cell of the counting matrices: the bottom-right entries in Figures 9-11, where the model is said to correctly predict two animals in close proximity. No sample size is reported for that cell, no confidence interval is given, and there are no repeated training runs. With 262 deer images total and presumably few multi-deer images, the observed difference could easily be due to chance. Quantitative evidence—such as the number of images in that cell, a confidence interval, or a statistical test—is required to support this load-bearing claim.","section":"Section IV-B and Figures 9-11"},{"comment":"Each model is evaluated at its own validation-tuned confidence threshold (0.255 for RT-DETR-l, 0.174 for YOLOv11n-det, 0.164 for YOLOv11n-seg), and the reported precision, recall, F1-score, and mAP@50 values are single point estimates without error bars, variance, or repeated runs. Threshold tuning on the validation set followed by reporting on an undefined test set can inflate apparent differences between models. The comparison should include threshold-independent metrics (e.g., precision-recall curves) or variance estimates to establish that the observed differences are not artifacts of threshold selection or training stochasticity.","section":"Section IV-A"}],"minor_comments":[{"comment":"There are frequent typographical spacing errors in \"UA V\" (e.g., in the abstract and introduction), and \"sligthly\" appears in the conclusions; these should be corrected.","section":"Throughout"},{"comment":"The citation [22] is given as the source for DETR, but the reference listed is \"An end-to-end transformer model for 3D object detection\" by Misra et al., which is not the DETR paper. The correct reference is Carion et al., \"End-to-End Object Detection with Transformers,\" ECCV 2020. This should be fixed, as the current citation does not support the text.","section":"Section II"},{"comment":"The phrase \"achieved an mAP@10 of 0.950\" in the discussion of the authors' previous work is likely a typo; typical metrics are mAP@50 or mAP@0.5. Please clarify the intended metric.","section":"Section II"},{"comment":"Figure 1 is described as showing a sample image from the test set, but since the test set is not defined anywhere in the paper, the figure caption is ambiguous. Once the test set is properly described, the caption should reference it explicitly.","section":"Figure 1"},{"comment":"The paper does not state whether the dataset, code, or trained models will be made publicly available. Given that one of the contributions is the dataset with segmentation masks, a data availability statement would strengthen the reproducibility of the work.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important applied problem and introduces a potentially useful dataset, but the evaluation methodology must be substantially improved before the central claim can be accepted. The undefined test set and the possible contamination from model-discovered deer are serious issues that require more than minor editing. The comparison also lacks uncertainty quantification. I recommend major revision with a request for a clear description of the test set, an explicit statement on whether the 54 model-found deer are in the evaluation labels, and a re-analysis that addresses the points above. Additionally, the DETR citation should be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the dataset is the real contribution; the headline comparison is not yet supported because the test set is never defined, the thresholds are tuned per model on validation, and the 54 'found' deer may have been folded into the evaluation labels. The paper is a useful applied study, but the central claim needs more work.\n\nWhat's actually new: a UAV marsh deer dataset with fine-grained segmentation masks, and a three-way comparison of YOLOv11n-det, YOLOv11n-seg, and RT-DETR-l. Adding a segmentation head is a known trick from Mask R-CNN, but testing it on tiny occluded animals in aerial imagery is a sensible experiment. The paper also gives a clear account of flight parameters, annotation protocol, and training setup (1920x1920, batch size, inference times). The observation that RT-DETR loses its real-time advantage at high resolution is a useful practical detail.\n\nThe soft spots are real and they hit the main conclusion. Section III-B-3 defines training and validation sets only; Section IV-A reports 'test set' metrics without saying what the test set is. That's a load-bearing gap. Each model also gets its own confidence threshold tuned on validation; that's not fatal by itself, but with an undefined test set you can't tell whether the comparison is fair. The 54 deer the network found and experts confirmed were added to the final count of 285. If any of those appear in the test labels, the detector is being scored against its own output. The paper never clarifies. Finally, the overlapping-specimen advantage rests entirely on the bottom-right cell of the counting matrices, with no sample size, confidence interval, or repeated runs. With 262 deer images, that cell could be a handful of images.\n\nThe math and citation pattern look fine. The authors cite the relevant work, including their own prior paper, which is reasonable. No invented entities. This is an honest applied paper that overreaches in its conclusion.\n\nWho this is for: conservation practitioners and people building low-cost UAV monitoring systems. A serious referee should send this back for major revision: define the test split explicitly, separate any model-assisted discoveries from training labels, add uncertainty via repeated runs or bootstrap, and report the sample size behind the overlap claim. Ideally release the dataset. With those fixes, it could be a solid applied contribution.","headline":"The dataset is the real contribution; the superiority claim for YOLOv11n-seg is not yet supported because the test set is unspecified and the 54 'found' deer may contaminate the labels.","tokens_in":9999,"tokens_out":3287,"would_cite":false,"duration_ms":28860,"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":"This paper claims that adding an instance segmentation head to a small YOLOv11 detector improves detection of small, occluded marsh deer in UAV imagery compared with detection-only YOLOv11 and the larger RT-DETR-l.","keywords":["marsh deer","Blastocerus dichotomus","UAV wildlife monitoring","YOLOv11","RT-DETR","instance segmentation","small object detection","aerial census"],"falsifier":"Re-run the three models on a documented test set built from transects held out before any labeling, with the 54 expert-confirmed ‘new’ deer either excluded from all training data or explicitly marked as a separate evaluation, and compare per-image counting accuracy on images containing two deer that overlap. If YOLOv11n-seg’s edge over YOLOv11n-det disappears under that clean split, the claimed advantage is an artifact of the unspecified test set or of the models having seen the newly discovered deer during training.","tokens_in":8971,"feed_emoji":"🦌","tokens_out":9658,"duration_ms":80946,"temperature":0.7,"pith_summary":"This paper tries to establish that adding an instance segmentation head to a lightweight YOLOv11 detector improves its ability to find marsh deer in high-resolution drone imagery, particularly when deer occupy only a few pixels or are partly hidden by vegetation. The motivation is practical: counting endangered deer from aerial surveys is slow and error-prone when done by human reviewers, and small, occluded animals are exactly where detection-only models stumble. The authors compare three models on a common data set, a detection-only YOLOv11n, a YOLOv11n with a segmentation head, and the transformer-based RT-DETR-l, and conclude that the segmentation-equipped model is superior, at the cost of a modest increase in inference time. If the claim holds, conservation programs get a concrete recipe: collect segmentation masks alongside bounding boxes, and small single-stage models will handle the occluded-animal cases that matter for accurate population estimates.","feed_headline":"Adding segmentation masks sharpens drone deer detection","feed_subtitle":"A tiny YOLO with a segmentation head beats detection-only rivals on small, overlapping deer in drone images.","key_machinery":"The mechanism that carries the argument is the instance segmentation head added to the YOLOv11 nano backbone, making the YOLOv11n-seg variant. Following the YOLACT design, this head predicts prototype masks shared across detections together with per-instance mask coefficients, and the final mask for each object is a linear combination of prototypes weighted by those coefficients, so each animal must be localized by its silhouette rather than only by a box. Training on images resized to 1920×1920 keeps the few-pixel deer specimens recognizable, and the paper argues that having to predict exact boundaries is what lets the model distinguish overlapping animals that detection-only heads merge. The comparison is completed by per-model confidence thresholds (0.164 for YOLOv11n-seg, 0.174 for YOLOv11n-det, 0.255 for RT-DETR-l) tuned on the validation set, with mAP@50 as the headline metric.","core_discovery":"The central claim is that fine-grained segmentation masks are more than a decorative output: as a training signal, they make the detector itself better. The authors report that YOLOv11n-seg, the smallest of the three models, achieves the best balance of precision, recall, and F1-score at its tuned confidence threshold, and is the only model that reliably separates cases where two animals appear close together in one image. They conclude that “the model YOLOv11n-seg is superior, showing greater potential in situations where there is overlapping specimens since it takes advantage of training considering the segmentation masks.” RT-DETR-l, the transformer-based rival, was competitive on accuracy despite having no segmentation head, but its inference time grew to 3.1–3.2 seconds per image at 1920×1920, erasing its real-time advantage. The paper also reports that the models found 54 marsh deer that expert human labelers had missed, an increase of over 23% in the total count, and presents the mask-annotated image set as a resource it believes is unique among wildlife monitoring datasets.","pith_inferences":["If the segmentation head acts as a general regularizer for small targets, the same mask-equipped YOLO recipe may transfer to other aerial wildlife census species with only a modest labeling investment; this is a testable extrapolation the paper does not make.","The 23% increase from model-found deer, combined with the re-labeling workflow described in Section III-A, leaves open whether those newly found deer were folded into training data, which would create a feedback loop that flatters the final reported counts.","A cheaper route to the same overlapping-specimen capability might be to keep a detection-only model but add an auxiliary ‘two animals here’ loss; the paper’s data alone cannot distinguish mask supervision from silhouette supervision more generally.","For operational surveys, the relevant comparison is counting error per transect at a fixed budget of human verification time rather than mAP; the paper reports counting matrices for up to two deer per image that begin to answer this, but does not compute a population-density estimate from the corrected counts."],"forward_implications":["Marsh deer census from drone imagery becomes more reliable with a segmentation-augmented detector, since the only model that told apart two deer in close proximity was the one trained on masks.","RT-DETR-l cannot keep its real-time promise at 1920×1920 input resolution, so its fast convergence does not translate into practical large-scale surveys at the resolutions needed for small specimens.","The combined models raised the detected deer count by over 23% compared with the expert census, showing that AI-assisted review can catch animals human labelers missed.","Collecting precise segmentation masks is worth the annotation effort because the same masks improve the detection task itself, not just the visual output of the model."],"supporting_citations":[{"why":"Supplies the Pantano Project aerial survey imagery of marsh deer that forms the entire dataset.","marker":"[3]"},{"why":"Ultralytics YOLO repository providing the YOLOv11 implementations, default hyperparameters, and pretrained weights used for fine-tuning.","marker":"[5]"},{"why":"The RT-DETR paper that defines the transformer-based baseline architecture and its real-time performance claim at 640×640.","marker":"[6]"},{"why":"The authors' earlier deer detection study whose dataset and mAP@10 of 0.950 this work extends with segmentation masks.","marker":"[21]"},{"why":"Mask R-CNN, the evidence cited for the idea that adding a segmentation head improves the detection task itself.","marker":"[25]"},{"why":"YOLACT, which supplies the prototype-mask plus coefficients mechanism the YOLOv11n-seg head uses to produce instance masks.","marker":"[29]"}],"fun_headline_variants":["Tiny YOLO with segmentation head finds 54 hidden marsh deer","Segmentation masks boost drone deer detection accuracy","Smallest drone-detection model beats rivals on overlapping deer","Segmentation training elevates tiny drone deer detector","Drone deer counting gets 23% boost from mask-aware AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the test set used for the final metrics is a fair, independent sample of the flight imagery; the paper describes training and validation sets in detail but never defines the composition, size, or overlap of the test set, so the reported superiority of YOLOv11n-seg could in principle rest on an unverifiable split of the data.","fun_headline_variants_meta":{"raw":{"variants":["Tiny YOLO with segmentation head finds 54 hidden marsh deer","Segmentation masks boost drone deer detection accuracy","Smallest drone-detection model beats rivals on overlapping deer","Segmentation training elevates tiny drone deer detector","Drone deer counting gets 23% boost from mask-aware AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000485,"raw_usage":{"total_tokens":2350,"prompt_tokens":859,"completion_tokens":1491,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":1411}},"tokens_in":475,"tokens_out":1491,"duration_ms":12455,"temperature":1.0,"reasoning_tokens":1411,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:10:36.206762+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the three models on a documented test set built from transects held out before any labeling, with the 54 expert-confirmed ‘new’ deer either excluded from all training data or explicitly marked as a separate evaluation, and compare per-image counting accuracy on images containing two deer that overlap. If YOLOv11n-seg’s edge over YOLOv11n-det disappears under that clean split, the claimed advantage is an artifact of the unspecified test set or of the models having seen the newly discovered deer during training.","supporting_citations":[{"cited_title":"Proyecto Pantano,","cited_arxiv_id":null,"evidence_quote":"Supplies the Pantano Project aerial survey imagery of marsh deer that forms the entire dataset."},{"cited_title":"Ultralytics YOLO,","cited_arxiv_id":null,"evidence_quote":"Ultralytics YOLO repository providing the YOLOv11 implementations, default hyperparameters, and pretrained weights used for fine-tuning."},{"cited_title":"DETRs Beat YOLOs on Real-Time Object Detection,","cited_arxiv_id":null,"evidence_quote":"The RT-DETR paper that defines the transformer-based baseline architecture and its real-time performance claim at 640×640."},{"cited_title":"Efficient endangered deer species monitoring with uav aerial imagery and deep learning,","cited_arxiv_id":null,"evidence_quote":"The authors' earlier deer detection study whose dataset and mAP@10 of 0.950 this work extends with segmentation masks."},{"cited_title":"Mask R-CNN,","cited_arxiv_id":null,"evidence_quote":"Mask R-CNN, the evidence cited for the idea that adding a segmentation head improves the detection task itself."},{"cited_title":"YOLACT: Real-time instance segmentation,","cited_arxiv_id":null,"evidence_quote":"YOLACT, which supplies the prototype-mask plus coefficients mechanism the YOLOv11n-seg head uses to produce instance masks."}],"review_version":1}