{"id":"4ba84a76-0999-4d69-a3e8-274f506409bc","arxiv_id":"2507.02519","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"IMASHRIMP combines ViTPose keypoint estimation, ResNet-50 discriminators, and SVM regression to measure 23 white-shrimp traits from RGBD images, reporting a mean error of 0.07 cm.","lead":"IMASHRIMP is an automated imaging pipeline that measures 23 body traits of white shrimp from photos, replacing slow manual measurements used in breeding programs. If its reported accuracy holds, it could accelerate genetic selection in shrimp farming.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.07 cm MAE claim rests on an unstated evaluation protocol: §5.4 never specifies a train/test split for the SVM regressors, and with 12 images per animal a random split could leak the same individuals into both fitting and testing.","rationale":"The reader's weakest assumption correctly identifies that manual measurements are used as ground truth for both fitting and evaluating the regression, and that systematic bias would make 0.07 cm an agreement-with-humans rather than physical accuracy. I agree with that concern, but I think the more immediately decisive issue is the protocol gap: no train/test split is reported for the SVMs, and the 12-images-per-animal acquisition design makes image-level leakage plausible. The two issues reinforce each other: even a perfect split cannot correct for a biased reference instrument, and even an unbiased reference cannot rescue an evaluation that may fit and test on the same animals. I did not find a separate objection to the pose-estimation or discriminator claims; the mAP and rostrum-error inconsistencies noted by the reader (97.94% versus 93.12% mAP; 3.64% versus 4.64% rostrum error) are real but look like reporting inconsistencies rather than threats to the core pipeline. A reproduced leave-one-animal-out regression evaluation would settle whether the 0.07 cm result is robust. The reader's condition of releasing data or benchmarking on held-out industrial data remains necessary; I would sharpen the condition to require an explicit split protocol for the regression module, ideally by animal ID. Because the concern is addressable by re-evaluation and does not by itself show the method cannot work, the CONDITIONAL verdict is the right level; no change to the final verdict is needed, but the condition should be made explicit.","tokens_in":18019,"tokens_out":8828,"duration_ms":91561,"concrete_test":"Using the released ImaShrimp-public repository, enumerate the animal IDs and reproduce the size-regression evaluation with a leave-one-animal-out 5-fold split: fit the SVM regressors on 80% of the 1030 animals and evaluate on the held-out 20%, then repeat across folds. Compare the mean held-out MAE with Table 4's 0.07 ± 0.1 cm and with the ruler-scaling column (0.13 ± 0.19 cm). If the held-out MAE is not clearly below the ruler-scaling baseline (for example, if the 95% confidence intervals overlap or the mean exceeds 0.10 cm), the headline advantage is attributable to evaluation setup rather than to the regression itself. If animal IDs or a split script are absent from the repository, request the exact split from the authors; the absence of a reproducible split is itself a negative finding for the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the pixel-to-centimetre regression MAE of 0.07 ± 0.1 cm (abstract; Table 4). For this number to support the claimed advantage over ruler scaling, the SVM regressors must be evaluated on individuals and images not used to fit them. The manuscript does not establish this. Section 5.4 describes fitting the SVMs using the manual measurements and annotated key points, but it gives no data split, no cross-validation, and no statement that the 1236 test images (or the 1030 animals behind them) were excluded from regression training. Section 5.1 states that each animal contributes 12 images (three views at four angles); if the split is by image rather than by animal, the same specimen appears on both sides of the split, and the regressor can exploit within-individual correlation to lower the apparent error. This is compounded by the use of the same manual measurements as both regression target and evaluation reference: even with a correct split, the 0.07 cm figure measures agreement with the human labelers, not true physical accuracy, so a systematic bias in the manual labels would be inherited by the system. These are internal protocol gaps, not disagreements with an external consensus. Until the split is specified or the repository is inspected, the 0.07 versus 0.13 comparison in §6.4 is not a verified generalization result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents IMASHRIMP, a pipeline for automatic morphological analysis of white shrimp (Penaeus vannamei) from single RGB-D images. The system combines two ResNet-50-based binary discriminators (lateral/dorsal view and rostrum integrity), a ViTPose-derived keypoint estimator predicting 23-keypoint virtual skeletons for lateral and dorsal views, and per-trait SVM regressors that convert pixel-based distances into centimeters. The authors report a human-AI two-factor authentication that reduces view-classification human error from 0.97% to 0% and rostrum-integrity error from 12.46% to 3.64%, a pose-estimation mAP of 93.12% (or 97.94% in the abstract), and a pixel-to-centimeter conversion MAE of 0.07 ± 0.1 cm with regression versus 0.13 ± 0.19 cm with ruler-based scaling. The code is publicly available on GitHub.","tokens_in":18281,"tokens_out":3611,"duration_ms":38869,"significance":"If the reported performance is reproducible and generalizes, IMASHRIMP would be a useful contribution to aquaculture phenotyping: it addresses a real operational need in genetic breeding programs, uses a relatively large annotated dataset (12,367 images, 1,030 individuals) spanning 23 morphological traits, and provides a concrete baseline comparison against ruler-based scaling. The two-factor human-AI authentication idea is simple and practical, and the availability of code is a strength. However, the central quantitative claim for the regression module rests on an incompletely specified evaluation protocol, and several internal numerical inconsistencies must be resolved before the results can be taken at face value.","major_comments":[{"comment":"The SVM regression evaluation protocol is not specified. The text does not state whether the 1,236 test images, or the 1,030 specimens behind them, were held out from fitting the SVM regressors, nor does it describe the train/test split or any cross-validation. Because each animal contributes 12 images (Section 5.1), an image-level split could place the same individual on both sides of the split, allowing within-individual correlation to inflate the reported advantage of regression (0.07 ± 0.1 cm) over ruler scaling (0.13 ± 0.19 cm). Please specify the split, preferably by individual, and report results on data never used in regression fitting.","section":"§5.4, Table 4"},{"comment":"The abstract reports a pose estimation mAP of 97.94%, but Table 2 and the conclusion report a general mAP of 93.12% on the 1,236-image test set. The value 97.94% does not appear in Table 2 or elsewhere in the results section. This discrepancy affects the paper's headline claim and must be resolved by identifying the correct number and reporting it consistently.","section":"Abstract vs. §6.2, Conclusion"},{"comment":"The rostrum-discriminator results are internally inconsistent: the text states that the two-factor system made 45 failures on the 1,236 test images, corresponding to a percentage error of 3.64% (45/1236), but the text reports 'a percentage error of 4.64%.' The abstract and conclusion report 3.64%, so the 4.64% figure appears to be a typographical error; it should be corrected.","section":"§6.1.2"},{"comment":"The manual measurements are used both as the target for fitting the SVM regressors and as the reference for evaluating them. Even with a clean train/test split by individual, the reported MAE therefore measures agreement with the human measurement process rather than true physical accuracy; any systematic bias in the manual labels would be inherited by the system. The paper should state this limitation explicitly and, if possible, validate on an independent physical reference (for example, repeated caliper measurements by a second operator or a known calibration object).","section":"§5.1, §5.4"}],"minor_comments":[{"comment":"There are several typographical issues: 'Penaeus vannamei}' contains a stray brace, 'whiteshrimp' is missing a space, and 'The code are available' should be 'The code is available.'","section":"Abstract"},{"comment":"The third bullet contains an editorial note to the authors ('Not only lenght, but also height and widht. Maybe is better usedorsal? To be consistent with the rest of the paper') that should be removed before publication.","section":"§1, list of contributions"},{"comment":"There are numerous typographical and grammatical errors, including 'c entimetres', 'abilinear', the duplicated phrase 'Shrimp selective breeding.', 'virtual virtual skeleton', and 'lenght'/'widht'. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"Table 2 reports 'General' mAP as 93.12, while the conclusion text also gives 93.12 for the test set; please state explicitly that this is the overall mAP on the 1,236 test images and reconcile it with the abstract's 97.94%.","section":"§6.2, Table 2"},{"comment":"The text says 'Of the 12367 ground truth labels, human researchers made 106 errors (0.85%),' but Table 1 reports 12 human errors (0.97%) on the 1,236-image test subset. Please clarify that the 106 errors are on the full annotated set, and that Table 1's percentages refer to the test subset.","section":"§6.1.1"},{"comment":"The row labeled 'total' in Table 4 has MAE 0.51 ± 0.76 cm with regression, while the 'General' row has MAE 0.07 ± 0.1 cm; this could confuse readers. Clarify that 'total' is the total-length morphological variable and 'General' is the mean over all traits.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a practical problem and provides a substantial dataset, but the current version has several internal numerical inconsistencies and an underspecified evaluation protocol for the central regression claim. The issues are fixable, but they require careful re-analysis or explicit clarification. The authors may also wish to consider whether the novelty claim about being the first to apply pose estimation to shrimp is adequately supported by the cited literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2507.02519. IMASHRIMP is a serious application paper: it adapts ViTPose to a 23-keypoint shrimp skeleton (lateral and dorsal), adds ResNet-50 discriminators for view and rostrum integrity, and an SVM to convert pixel distances to centimetres. They annotate over 12k images of 1030 animals and report end-to-end MAE of 0.07 ± 0.1 cm versus 0.13 ± 0.19 cm for ruler-based scaling. This is the first transformer-based pose estimation for shrimp, and the skeleton definition ties directly to the traits used in their BLUP breeding program. The engineering contribution is real and the dataset is substantial.\n\nThe soft spots are also real. The abstract says mAP 97.94%, but Table 2 says 93.12% for the general test set. The rostrum error is reported as 3.64% in Table 1 and the abstract, while section 6.1.2 says 45 failures out of 1236, which is 3.64%, not the stated 4.64%. Minor but sloppy.\n\nMore importantly, the stress-test concern lands. Section 5.4 describes fitting the SVM regressors on manual measurements and annotated keypoints but never states how the data was split. Each animal has 12 images (three views at four angles). If the split is by image rather than by animal, the same individual appears on both sides of the split, and the regressor can exploit within-individual correlation to lower the apparent error. The 0.07 versus 0.13 comparison in Section 6.4 is then not a verified generalization result. Combined with the manual measurements being both the regression target and the evaluation reference, the 0.07 cm figure only demonstrates agreement with the human labelling process, not true physical accuracy.\n\nI do not think this is fraud or a fatal flaw. The fix is straightforward: re-run with an animal-level split and report per-trait numbers. But until that is done, I would be cautious about quoting the 0.07 cm claim. There is also an apparent editorial leftover in the contributions section (\"Maybe is better usedorsal?\") and the abstract mentions a GitHub link that seems to be a placeholder in the full text. These are the kinds of things a desk editor would want cleaned up.\n\nWho is this for? People working on aquaculture phenotyping and applied pose estimation. It deserves a serious referee: the method is coherent, the task is important, and the open issues are addressable in revision. I would send it to peer review but ask for the split protocol and the numeric discrepancies to be resolved, and ideally a release of at least the evaluation data or a benchmark comparison. My own verdict is conditional, with the regression protocol as the gating issue.","headline":"Useful engineering work on shrimp phenotyping, but the headline 0.07 cm MAE is unverifiable as written because the SVM regression train/test protocol is never specified.","tokens_in":18960,"tokens_out":2729,"would_cite":false,"duration_ms":28766,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A pose-estimation system predicts 23 key points on a shrimp's body from a single RGBD image and regresses those pixel distances to centimeters, achieving an average error of 0.07 ± 0.1 cm across all morphological traits.","keywords":["Shrimp size estimation","Penaeus vannamei","Genetic assessment","Pose estimation","Computer vision","Deep learning","RGBD imaging","Morphological traits"],"falsifier":"Image a rigid object of known dimensions, or shrimp measured by an independent calibrated 3D scanner, with the same 30 cm zenithal camera and run the full pipeline. If the predicted traits deviate from the known values by more than the reported 0.07 ± 0.1 cm mean absolute error, the claim that the system measures true morphology rather than reproducing the manual labels is refuted.","tokens_in":17796,"feed_emoji":"🦐","tokens_out":10316,"duration_ms":99109,"temperature":0.7,"pith_summary":"IMASHRIMP is a computer-vision system that automatically measures 23 morphological traits of white shrimp (Penaeus vannamei) from single RGBD images, replacing manual measurements used in genetic selection programs. The system couples a pose-estimation network that locates 23 key points on the shrimp's body with a per-trait regression model that converts pixel distances into centimeters, reporting an average error of 0.07 ± 0.1 cm across all traits. Two verification classifiers check the human operator's labels for image view and rostrum integrity, reducing human error in those labels from 0.97% to 0% and from 12.46% to 3.64%, respectively. If the reported accuracy holds, breeders could evaluate far more shrimp per generation, faster and with fewer transcription mistakes.","feed_headline":"0.07 cm error measured across 23 shrimp traits from one image","feed_subtitle":"Pose-estimation system replaces manual calipers in shrimp breeding, cutting human error and speeding genetic selection.","key_machinery":"The load-bearing object is the virtual skeleton: 23 key points placed at the start and end of each morphological measurement, with one skeleton for the lateral view and one for the dorsal view. The pose estimator adapts the VitPose vision-transformer encoder/decoder to predict these points from RGBD images, using separate networks for lateral and dorsal views, while a modified ResNet-50 provides the two binary classifiers (view and rostrum integrity). The pixel-to-centimeter conversion is carried out by a Support Vector Regression model fitted separately for each of the 22 measurements, learned from paired pixel distances and manual centimeter measurements.","core_discovery":"The paper claims that measuring a shrimp can be reformulated as a keypoint-detection problem: a network predicts the coordinates of 23 anatomically defined points (a lateral skeleton or a dorsal skeleton), every trait is computed as a distance between two of those points, and a Support Vector Regression converts those pixel distances into centimeters. On a test set of 1236 images, the full system achieves a mean average precision of 93.12% for keypoint detection and a pixel-to-centimeter mean absolute error of 0.07 ± 0.10 cm, outperforming the ruler-scaling baseline of 0.13 ± 0.19 cm. The paper further claims this is the first application of pose estimation to shrimp, and that its two-factor authentication scheme—human annotation plus an AI classifier—cuts human error in view classification to zero and reduces rostrum-integrity errors from 12.46% to 3.64%.","pith_inferences":["Because the manual measurements recorded at capture time are the training and evaluation labels, the reported error measures agreement with that human process; an independent validation against a calibrated 3D scanner would show whether the system tracks true physical length.","The same pose-estimation-plus-regression pipeline should transfer to other farmed crustaceans or fish with defined anatomical landmarks, requiring only a modest annotated dataset.","A natural stress test is to quantify performance at the 40 cm and 60 cm camera distances that the paper shows only qualitatively, since the regression was fitted at the fixed 30 cm setup."],"forward_implications":["If the 0.07 cm average error holds in routine use, breeders can replace manual caliper measurements with automated image-based phenotyping for thousands of shrimp per selection cycle.","Because all 23 traits come from the same image, genetic selection indices based on multiple correlated morphological traits can be computed without additional photography.","The two-factor authentication scheme offers a template for any data-entry pipeline where a cheap classifier can catch human labeling mistakes.","The per-trait regression conversion beats ruler scaling, indicating that learned conversion is the better default when the camera setup is fixed and a 3D object is imaged in 2D."],"supporting_citations":[{"why":"Supplies the VitPose architecture that the shrimp pose estimator adapts for keypoint detection.","marker":"Xu et al. (2022)"},{"why":"Establishes the convolutional-pose-machines line that motivates modeling measurement as keypoint detection.","marker":"Wei et al. (2016)"},{"why":"Provides the ResNet-50 architecture used by the view and rostrum discriminators.","marker":"He et al. (2016)"},{"why":"Defines the Support Vector Regression used to convert pixel distances to centimeters.","marker":"Vapnik et al. (1996)"},{"why":"Shows that a regression model yields more precise fish-size conversion than simple scaling, the approach this paper follows.","marker":"Garcia et al. (2019)"},{"why":"IMAFISH is the ruler-scaling baseline to which the regression method is compared.","marker":"Navarro et al. (2016)"},{"why":"Defines the growth and morphological traits and their genetic parameters, establishing which measurements the system must produce.","marker":"Shin et al. (2023)"},{"why":"Supplies the trait definitions and genetic relevance of rostrum and cephalothorax measurements used in the skeleton design.","marker":"Martínez Soler et al. (2024)"},{"why":"The closest prior work, a 2D shrimp analysis with 2.1 cm MAE, which the paper's 3D pose-based approach claims to outperform.","marker":"Chirdchoo et al. (2024)"},{"why":"Introduces the vision transformer encoder that the pose network's encoding stage is built on.","marker":"Dosovitskiy et al. (2021)"}],"fun_headline_variants":["AI measures 23 shrimp traits with 0.07 cm error from one image","Pose estimation automates shrimp biometrics, cuts human error to zero","One photo yields 23 shrimp measurements accurate to 0.07 cm","First shrimp pose-estimation network: 93% precision, 0.07 cm error","Deep learning replaces calipers in shrimp breeding, error 0.07 cm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The manual measurements recorded during image capture are treated as ground truth for fitting and evaluating the regression, so the 0.07 cm error shows how well the system reproduces those human labels, not how close either is to true physical length.","fun_headline_variants_meta":{"raw":{"variants":["AI measures 23 shrimp traits with 0.07 cm error from one image","Pose estimation automates shrimp biometrics, cuts human error to zero","One photo yields 23 shrimp measurements accurate to 0.07 cm","First shrimp pose-estimation network: 93% precision, 0.07 cm error","Deep learning replaces calipers in shrimp breeding, error 0.07 cm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001014,"raw_usage":{"total_tokens":4321,"prompt_tokens":1024,"completion_tokens":3297,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":3194}},"tokens_in":640,"tokens_out":3297,"duration_ms":21804,"temperature":1.0,"reasoning_tokens":3194,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:27:44.925251+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Image a rigid object of known dimensions, or shrimp measured by an independent calibrated 3D scanner, with the same 30 cm zenithal camera and run the full pipeline. If the predicted traits deviate from the known values by more than the reported 0.07 ± 0.1 cm mean absolute error, the claim that the system measures true morphology rather than reproducing the manual labels is refuted.","supporting_citations":[{"cited_title":", author Ramakrishna, V","cited_arxiv_id":null,"evidence_quote":"Establishes the convolutional-pose-machines line that motivates modeling measurement as keypoint detection."},{"cited_title":", author Golowich, S.E","cited_arxiv_id":null,"evidence_quote":"Defines the Support Vector Regression used to convert pixel distances to centimeters."},{"cited_title":", author Prados, R","cited_arxiv_id":null,"evidence_quote":"Shows that a regression model yields more precise fish-size conversion than simple scaling, the approach this paper follows."},{"cited_title":", author Lee-Montero, I","cited_arxiv_id":null,"evidence_quote":"IMAFISH is the ruler-scaling baseline to which the regression method is compared."},{"cited_title":", author Shin, H.S","cited_arxiv_id":null,"evidence_quote":"Supplies the trait definitions and genetic relevance of rostrum and cephalothorax measurements used in the skeleton design."},{"cited_title":", author Mukviboonchai, S","cited_arxiv_id":null,"evidence_quote":"The closest prior work, a 2D shrimp analysis with 2.1 cm MAE, which the paper's 3D pose-based approach claims to outperform."},{"cited_title":", author Beyer, L","cited_arxiv_id":null,"evidence_quote":"Introduces the vision transformer encoder that the pose network's encoding stage is built on."}],"review_version":1}