{"id":"754bb9e3-cb24-4872-b807-8d8e2151d356","arxiv_id":"2607.26104","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Full-body images and multi-task regression improve CNN-based BMI/weight/height estimation on a new self-reported in-the-wild dataset, though no code, data, or trivial baseline is provided.","lead":"The paper trains deep networks to estimate a person's height, weight, and BMI from a single social-media photo, reporting that full-body images and multi-task learning give the best results. It introduces a 6,105-image in-the-wild dataset, but that dataset and the code are not yet available, and no trivial baseline is reported.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Height MAE of 0.08 m may equal trivial mean-predictor baseline, so the claimed superiority of MTR for height is unsupported.","rationale":"The reader's weakest assumption concerns label noise from self-reported reddit/imgur data. That is a valid concern, but the trivial-baseline issue is more directly tied to the central claim as stated: even if the labels were perfectly accurate, the reported height MAE may not exceed what a constant predictor achieves. Height is one of the three target variables and an explicit part of the claimed multi-task advantage, so if the height results are vacuous, the central claim is weakened substantially. The paper provides no error bars, no significance tests, and no comparison to a mean predictor, so the claimed MTR superiority over STR (and over Jiang et al.'s GPR, which also achieves 0.09 m on height) could be within noise. This does not contradict the reader's verdict of REJECT; it strengthens it by identifying a specific, testable failure mode that is independent of label quality. The concrete test can be executed from the authors' reported distribution and would settle the concern without additional data collection.","tokens_in":18558,"tokens_out":4516,"duration_ms":48597,"concrete_test":"Reproduce the height MAE of a trivial baseline that always predicts the training-set mean height on the same test splits (ITU-BMI B, ITU-BMI U, ITU-BMI F, and full ITU-BMI) used for Table II. If the baseline MAE is ≤0.08 m (or within 0.01 m of the reported MTR result), then the height results do not establish that the network estimates height beyond the prior, and the 'full-body is better' claim for height collapses. Additionally, report per-height-interval MAE to check whether errors concentrate near the mean, and run a paired significance test (e.g., bootstrap or paired t-test) between MTR and STR height predictions on the same test images.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim—MTR consistently achieves lower MAE than STR and prior methods across all four datasets (Table II)—rests on all three target variables, including height. The best height MAE is 0.08 m on ITU-BMI B. However, the dataset's height distribution (Figure 2a) is narrow: heights range 1.40–2.20 m with a peak at 1.63–1.73 m; the standard deviation is likely ~0.10–0.12 m. For a Gaussian variable with standard deviation σ, the optimal constant predictor (the mean) achieves MAE = σ·√(2/π) ≈ 0.80σ. If σ ≈ 0.10 m, the trivial baseline MAE is ≈0.08 m—the same as the reported MTR result. The paper does not report a trivial baseline, error bars, or significance tests, so the height column of Table II does not demonstrate that the model estimates height beyond predicting the dataset mean. This is reinforced by the near-identical height MAE across very different input modalities (0.08–0.11 m in Table III) and across backbones (Table V), suggesting height predictions are insensitive to image content. This concern is at least as load-bearing as label noise: even with perfect labels, the claimed height-estimation advantage could vanish against a mean predictor, and the small STR-vs-MTR differences (e.g., 0.09 vs 0.08) could be within noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ITU-BMI, a dataset of 6,105 social-media images labeled with self-reported height, weight, and BMI, and evaluates deep CNN architectures (ResNet-50, DenseNet-121, VGG-16) for regression and classification of these attributes. It compares single-task and multi-task learning on RGB images and on fused GAD/GAM modality images, reporting that multi-task regression (MTR) consistently achieves lower MAE than single-task regression and earlier methods, and that full-body images outperform face-only or upper-body inputs. The manuscript also reports classification accuracy and AUC for BMI, weight, and height.","tokens_in":18878,"tokens_out":6790,"duration_ms":69317,"significance":"If the empirical claims were well supported, the dataset and the multi-task design would be a useful benchmark for in-the-wild body-metric estimation, and the multimodal fusion study would inform practical deployment. The paper's assets are the relatively large in-the-wild image collection, the attempt to combine multiple input modalities, and the systematic comparison of STR/MTR/STC/MTC across several backbones. However, the paper does not yet provide the evidence needed to support the headline claims: there is no trivial baseline, no error bars or significance tests, and no label-verification protocol. The contribution is therefore currently suggestive rather than established.","major_comments":[{"comment":"The reported height MAE of 0.08 m is not interpretable without a mean-predictor baseline. Given the height distribution in Fig. 2(a) (1.40–2.20 m, peak 1.63–1.73 m), a model that always predicts the training mean can achieve MAE ≈ 0.8σ, which is ≈ 0.08 m when σ ≈ 0.10 m. All height entries in Tables II, III, and V fall in this range, and the near-identical values across input modalities and backbones are consistent with the model predicting the marginal mean rather than image-specific height. The authors should report null baselines for each target on each dataset/split, with confidence intervals, and demonstrate that the reported improvements are statistically significant.","section":"VI-B, Table II; Fig. 2(a)"},{"comment":"The ground-truth height and weight labels are self-reported values scraped from reddit and imgur. The text states that uploaded images come with exact measurements, but no verification protocol is described: there is no manual audit, no exclusion of motivational or deliberately mislabeled posts, and no error model for self-reported measurements. Since every MAE and accuracy value in Tables II–VIII is computed against these labels, systematic bias or random label noise directly affects all conclusions. The authors should provide a label-quality audit on a sample, cross-checks where the same user appears in multiple posts, and a sensitivity analysis under plausible label-noise levels.","section":"IV; all regression/classification tables"},{"comment":"All results come from a single random 80/20 split, with no repeated splits, cross-validation, or significance tests. Hyperparameters are described as empirically found without a separate validation set. The smallest reported differences — for example, STR vs MTR height of 0.09 vs 0.08 on ITU-BMI B, or the backbone differences in Table V — are within the likely sampling noise of a 537-image test set. The authors should report means and standard deviations across multiple splits and perform paired significance tests for STR vs MTR and for backbone comparisons.","section":"VI-A; Tables II–V"},{"comment":"The description of ITU-BMI B is ambiguous. The text says the 6,105 images are randomly divided into 80/20 train/test, and then says the 537 full-body images are \"taken from test set only\" and define ITU-BMI B. If this is literal, no model was trained on ITU-BMI B, making the B row in Table II not a valid evaluation of a model trained on that subset. If the intended meaning is simply that B is a subset of the full corpus, this should be stated and the overlap between ITU-BMI, ITU-BMI B, ITU-BMI U, and ITU-BMI F clarified. The small B-subset size also reinforces the need for confidence intervals.","section":"VI-A; Table II"},{"comment":"The text states that \"multi-task classification (MTC) has performed better than single-task classification (STC) in all experiments,\" but Table VII shows RGB BMI accuracy of 62.32 for STC versus 62.08 for MTC, and several other MTC-vs-STC differences are under one percentage point. This contradicts the stated claim. The authors should correct the claim, report whether any of the differences are statistically significant, and qualify the multi-task advantage accordingly.","section":"VI-C, Table VII"}],"minor_comments":[{"comment":"Subsections V-D and V-E are empty headings (\"Transformers for BMI Classification and Regression\" and \"Transformers for MTC and MTR of BMI, Weight and Height\"). Either the content is missing or these headings should be removed. As written, the manuscript is incomplete.","section":"V-D, V-E"},{"comment":"The conclusion says the dataset is \"labelled into five categories including under-weight, normal, over-weight, and obese,\" but four categories are listed and used. Please correct the count.","section":"VIII"},{"comment":"The text reports \"the heighest correlation is found between height and BMI which is 0.80, and between weight and BMI is 0.58.\" These values are surprising for a population where BMI is computed from weight and height, and they are not accompanied by a definition of the correlation or a table. Please clarify how these correlations were computed and correct the typo.","section":"VI-C"},{"comment":"The paper says experiments include \"RGB and gray scale image representations,\" but the reported modality results are RGB, GAD, and GAM. No gray-scale-only results appear in any table. Please either report grayscale results or revise the text.","section":"V-C"},{"comment":"The phrase \"significantly better\" and \"significantly larger error\" is used repeatedly, but no significance tests are reported anywhere. Please replace these terms with quantitative claims or add statistical tests.","section":"VI-B, Table II"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an unfinished draft: it has empty transformer subsections, a self-contradictory statement about the ITU-BMI B split, and a classification claim contradicted by its own table. The dataset — the primary claimed contribution — is not yet publicly available. A revised version with null baselines, multi-split statistics, a label audit, and completed methods sections could be reconsidered; without these, the central empirical claims are not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is the dataset, not the method. A public in-the-wild full-body image collection with height, weight, and BMI labels would fill a genuine gap; the cited alternatives are either face-only or not public. The paper deserves credit for assembling 6,105 images with claimed label diversity and for recognizing that full-body crops should beat face-only crops. The multi-task regression recipe itself is standard hard-parameter-sharing transfer learning, and the authors do not oversell it as more than that.\n\nThat said, the empirical claims are not supported as stated. The stress-test concern about height lands. Heights in this dataset cluster around 1.63–1.73 m with total range 1.4–2.2 m. With a standard deviation near 0.1 m, a constant predictor at the mean already achieves MAE ≈ 0.08 m—exactly the best reported height error. No trivial baseline is reported, no error bars, no significance tests, and a single 80/20 split. So the height column of Table II does not demonstrate that the model extracts any height information from images. The near-identical height MAE across modalities and backbones reinforces this. The MTR-versus-STR advantage for BMI and weight may be real, but the differences are small (e.g., 3.73 vs 4.29 BMI MAE) and could be noise.\n\nOther soft spots are proportionate but real. Labels are self-reported from reddit/imgur posts; the paper asserts the images carry \"exact measurements\" but offers no verification protocol or error model. Sections V-D and V-E are empty headings titled \"Transformers\"—this is either missing content or placeholder text. Table I marks the dataset as available, but the paper only says it \"will soon\" be released; no code or data link is given. Those are not fatal to the dataset idea, but they are fatal to the paper's current verifiability.\n\nWho is this for? A reader working on soft biometrics or body attribute estimation would want to know this dataset exists and could be useful once released. As a paper, it is not ready in its current form. The right outcome is major revision with (1) a trivial baseline and error bars, (2) a released dataset, (3) filler sections removed, and (4) the height claims toned down or dropped. It deserves serious peer review rather than desk rejection because the dataset, if real, is a valuable resource and the failures are omissions, not fabricated results.","headline":"New dataset idea worth tracking, but the empirical claims, especially on height, aren't supported by the numbers as reported.","tokens_in":19407,"tokens_out":1323,"would_cite":false,"duration_ms":16752,"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 a person's weight, height, and body mass index can be estimated from a single unconstrained photo using multi-task deep learning, with the best full-body results reaching a mean absolute error of 3.32 BMI points.","keywords":["BMI estimation","weight estimation","height estimation","multi-task learning","in-the-wild images","deep convolutional networks","human body analysis","dataset collection"],"falsifier":"Collect a subset of the dataset images, have the same people measured in a controlled setting (or ask them to re-submit measurements), and compare the labels; if self-report error exceeds the reported MAE margins, the benchmark numbers are not reliable. Alternatively, test the trained model on a separate corpus of images with medically verified height/weight and observe whether MAE degrades substantially.","tokens_in":18410,"feed_emoji":"📏","tokens_out":7395,"duration_ms":74131,"temperature":0.7,"pith_summary":"The paper sets out to show that a person's weight, height, and Body Mass Index can be estimated from a single casual photograph taken in uncontrolled conditions, using deep multi-task learning rather than controlled lab setups or face-only images. To do this, the authors assemble a new dataset of 6,105 social-media photos with self-reported height and weight labels, covering diverse poses, backgrounds, and body types. Their experiments indicate that joint regression of all three quantities beats separate single-task models, and that whole-body images are more informative than face-only or upper-body crops. If correct, the work would make mass-scale height/weight/BMI screening from everyday imagery feasible for health monitoring, self-assessment, and person re-identification. The main empirical claim is a multi-task regression reaching a mean absolute error of 3.32 BMI points (12.60 kg, 0.08 m) on full-body images.","feed_headline":"One photo estimates BMI within 3.3 points","feed_subtitle":"Joint multi-task network outperforms face-only models on 6,105 in-the-wild social photos, enabling mass health screening.","key_machinery":"The central mechanism is hard-parameter-sharing multi-task learning: a shared convolutional backbone with three task-specific regression heads that jointly estimate normalized height, weight, and BMI. Because these quantities are physically correlated, the shared representation is meant to capture body shape and scale cues that benefit all three tasks, outperforming single-task regressors. The accompanying benchmark dataset — 6,105 unconstrained images with self-reported height/weight labels — provides the scale and diversity needed to train such networks, and the paper also experiments with pose-affinity maps, monocular depth maps, and foreground masks as auxiliary input modalities.","core_discovery":"The paper claims that a single everyday photograph of a person — taken in the wild, with varied pose, background, clothing, and without controlled camera geometry — contains enough visual information to estimate that person's weight, height, and Body Mass Index, and that doing the three estimations jointly (multi-task regression) is more accurate than estimating them separately. On the authors' newly collected dataset of 6,105 social-media images of people sharing weight-loss or weight-gain progress, the multi-task regression network using full-body images reaches a mean absolute error of 3.32 BMI points, 12.60 kg in weight, and 0.08 m in height, outperforming face-only and upper-body inputs","pith_inferences":["The reported accuracies are measured on a random split of the same social-media distribution the model was trained on; a truly independent test set (e.g., clinical intake photos) is needed before the numbers can be trusted for real-world deployment.","The labels are self-reported and the paper describes no verification; under-reported weight or over-reported height would bias both training and evaluation, so a noise-aware model or a verified subset could change results.","The manuscript contains placeholder sections on transformer architectures, suggesting that more modern attention-based models were contemplated but not evaluated; they might outperform the convolutional backbones tested.","The method's sensitivity to pose, clothing, and occlusion is not analyzed; a stratified error analysis by these factors would define the operational envelope of the approach."],"forward_implications":["If the authors' multi-task regression result holds, a single full-body photo can estimate BMI to within about 3.3 points (mean absolute error), weight to within 12.6 kg, and height to within 0.08 m — an accuracy range that makes automatic health screening from everyday photos feasible.","Jointly learning height, weight, and BMI in one network is consistently more accurate than learning them separately, so future systems for physical-attribute estimation should adopt multi-task architectures.","The new dataset of 6,105 in-the-wild images with self-reported height/weight labels — intended for public release — would give the research community a benchmark for body-based BMI estimation that previously did not exist.","Using full-body images instead of face-only or upper-body crops improves all three estimations, indicating that body shape carries the most relevant cues for these attributes.","Coarse BMI classification into WHO categories reaches about 65% accuracy with 81% area under the curve, which is enough for population-level trend monitoring but not for individual medical diagnosis."],"fun_headline_variants":["Single photo predicts BMI within 3.3 points","Deep net reads BMI from one wild photo","One selfie estimates your BMI, weight, height","Full-body images beat face-only for BMI prediction","Joint learning improves BMI from social photos"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim rests on the assumption that the self-reported heights and weights attached to the scraped social-media photos are accurate and correctly matched to the person in each image, yet no verification or error model is provided.","fun_headline_variants_meta":{"raw":{"variants":["Single photo predicts BMI within 3.3 points","Deep net reads BMI from one wild photo","One selfie estimates your BMI, weight, height","Full-body images beat face-only for BMI prediction","Joint learning improves BMI from social photos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1186,"prompt_tokens":825,"completion_tokens":361,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":291}},"tokens_in":569,"tokens_out":361,"duration_ms":4648,"temperature":1.0,"reasoning_tokens":291,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:22:17.799389+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a subset of the dataset images, have the same people measured in a controlled setting (or ask them to re-submit measurements), and compare the labels; if self-report error exceeds the reported MAE margins, the benchmark numbers are not reliable. Alternatively, test the trained model on a separate corpus of images with medically verified height/weight and observe whether MAE degrades substantially.","supporting_citations":[],"review_version":1}