{"id":"84be4b1f-5c7d-4702-96c3-32afb4ee28ca","arxiv_id":"2504.21166","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A sliding-window Laban Movement Analysis pipeline from 3D pose and mesh estimates reports 99%+ accuracy on AIST++ dance styles, though the cross-validation split and a non-temporal baseline are not clearly documented.","lead":"This paper builds a dance style classifier from 3D pose and body shape features based on Laban Movement Analysis, adding a sliding window so the features capture movement over time. It reports over 99% accuracy on AIST++, but the evaluation split is not fully specified and out-of-distribution accuracy drops sharply.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 99% accuracy claim depends on an unspecified train/test split unit; if windows, not videos, are the split unit, near-duplicate windows from the same video leak across folds and inflate accuracy.","rationale":"The reader's weakest assumption identifies precisely the load-bearing issue: the cross-validation split unit is unspecified, and a window-level split would leak near-duplicate temporal samples into both training and test folds. I agree that this is the main threat to the central claim. The paper has genuine strengths: a well-structured feature-extraction pipeline, an attempt at interpretability, and an out-of-distribution evaluation that shows lower, more plausible generalization numbers. Those strengths do not, however, resolve the split-unit ambiguity, and the numerical inconsistency between the abstract's 99.18% and Section 4's reported accuracy values (RF 99.68%, SVM 99.07%) reinforces the need for clarification. The concern is addressable by rerunning with a video-level split and reporting per-fold results, so a conditional verdict rather than outright rejection is appropriate. My read does not change the reader's verdict.","tokens_in":7556,"tokens_out":5776,"duration_ms":63773,"concrete_test":"Ask the authors to specify and, ideally, release the exact fold assignment, then rerun the full pipeline with a video-level split: assign each of the 600 source videos entirely to one of three folds, extract all sliding-window instances, train on two folds, test on the held-out videos, and report accuracy and macro recall per fold for both SVM and Random Forest. Also report the number of windows per fold and state whether any test window shares source frames with a training window. If the video-level accuracy is materially lower than the current 99% figures, the reported result is inflated by temporal overlap between training and test windows.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim — 99.18% highest accuracy and that temporal context significantly improves performance — rests entirely on the evaluation protocol described in Section 3.4. The paper states only that 3-fold cross-validation is used to split the dataset into training, validation, and testing sets, and it never says whether the split unit is the source video or the individual sliding-window feature vector. This is not a minor detail. With a 55-frame sliding window on 60-fps video, consecutive windows overlap by 54 frames, and all windows from a single video share the same dancer, choreography, camera, background, and lighting. If windows are randomly assigned to folds, near-duplicate samples can appear in both training and test sets, allowing the classifier to memorize video identity rather than generalize across dance styles. That would produce exactly the kind of near-perfect accuracies reported in Table 1. The honest out-of-distribution evaluation in Section 4.2, where per-style accuracy falls to between 10% and 82%, is consistent with such memorization and does not rule it out. A secondary inconsistency adds to the concern: the abstract's 99.18% does not match the accuracy values stated in Section 4 (RF 99.68%, SVM 99.07%), but it does equal the average SVM recall in Table 1. Thus the headline number itself needs clarification, and the evaluation protocol must be specified before the strongest claim can be accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline for dance style recognition on the AIST++ dataset. It combines 3D pose estimation, monocular scene reconstruction for floor estimation, SMPL mesh fitting, and Laban Movement Analysis (LMA) feature extraction with a sliding window to incorporate temporal context. A 55-feature vector is used to train support vector machines and random forests, and SHAP is applied for model interpretation. The authors report near-perfect in-distribution accuracy (the abstract states 99.18%; Section 4 reports RF 99.68% and SVM 99.07%) and also present an out-of-distribution evaluation on online videos with substantially lower per-style accuracy.","tokens_in":7866,"tokens_out":5301,"duration_ms":52291,"significance":"If the evaluation were sound, the paper would be a useful contribution: it offers an explicit, feature-based, explainable alternative to end-to-end deep learning for dance style classification, and it provides a practical recipe for extracting floor-aware 3D body features from monocular video. The use of SHAP to connect LMA descriptors to individual dance styles is a valuable interpretive step, and the out-of-distribution evaluation, despite its limitations, is an honest attempt to measure generalization. However, the central empirical claims are currently not verifiable because the cross-validation split is unspecified and the reported metrics are inconsistent.","major_comments":[{"comment":"The 3-fold cross-validation split is not defined at the correct unit. The paper must state whether the split is performed on the 600 source videos or on individual 55-frame sliding windows. With a 55-frame window at 60 fps, consecutive windows overlap by 54 frames, and all windows from the same video are highly correlated. If windows are randomly assigned to folds, the training and test sets will contain near-duplicate samples, allowing the classifier to memorize video-specific characteristics and inflate accuracy. This issue is load-bearing for the 99%+ accuracy claims. The experiments should use a video-level grouped split (e.g., GroupKFold), report variance across folds, and ideally report per-video majority-vote accuracy.","section":"Section 3.4"},{"comment":"The headline 'highest classification accuracy of 99.18%' in the abstract does not match the accuracy values stated in Section 4 (RF 99.68%, SVM 99.07%), but it does equal the average recall for SVM in Table 1. The paper must clarify whether the abstract reports accuracy or recall, and Table 1 should include an explicit accuracy row so the reader can verify the claimed metric. This inconsistency undermines the paper's central numerical claim.","section":"Abstract and Table 1"},{"comment":"The composition of the 55-feature vector is never specified. Equations (1)-(5) describe several quantities, but the mapping from these equations to the exact 55 input features is absent. For example, Equation (1) is a boolean inequality, not a scalar feature, and it is unclear how many joints, time offsets, and aggregations are used. Without a complete enumeration of the feature set, the method is not reproducible and the SHAP analyses cannot be interpreted. Provide the full list of the 55 features, including formulas for any remaining descriptors (spatial dispersion, volume, path, curvature, jerkiness, etc.).","section":"Section 3.3"},{"comment":"The claim that 'the addition of temporal context significantly improves dance style recognition performance' is not supported by a non-temporal baseline. The comparison in Fig. 2 is between different sliding-window sizes; it does not compare against the same feature set extracted without a sliding window (e.g., window size 1) or against existing methods on AIST++. Add a no-temporal-context baseline and compare with published AIST++ results to substantiate the contribution.","section":"Section 4, Fig. 2"},{"comment":"The out-of-distribution evaluation is described only qualitatively. The paper reports per-style accuracy values from 10.21% to 81.89% but does not state the number of test videos per style, the composition of the test set, or whether the same sliding-window protocol is used. Given the near-perfect in-distribution results, the low out-of-distribution performance for many styles suggests the model may be overfitting to AIST++ conditions. Report the evaluation protocol in sufficient detail so the robustness claims can be assessed.","section":"Section 4.2"}],"minor_comments":[{"comment":"The phrase 'in the context of emotion classification for contemporary dance' appears to be a leftover from a different paper; it should read 'dance style recognition'.","section":"Section 3.5"},{"comment":"The typo 'A verage' should be corrected to 'Average'.","section":"Table 1"},{"comment":"The figure caption and surrounding text would benefit from explicit axis labels and a description of the baseline to which the windowed results are compared.","section":"Section 4, Fig. 2"},{"comment":"The joint weights α_j are attributed to the OpenMMLab repository [12]; this is not a standard citable source for domain-specific joint importance. Please cite the original literature that defines these weights.","section":"Section 3.3, Ref [12]"},{"comment":"The dataset is described as the AIST Dance Video Database, but the paper consistently refers to AIST++. These are different resources (AIST++ provides 3D motion annotations); please clarify which dataset version is used and cite the corresponding paper.","section":"Section 3.1"},{"comment":"The kinetic-energy formula lacks a mass term; if the intended quantity is scaled kinetic energy (with mass absorbed into α_j), this should be stated explicitly.","section":"Equation (4)"}],"recommendation":"major_revision","confidential_remarks":"The evaluation-protocol ambiguity is the central technical issue: without a video-level split, the near-perfect accuracy figures are not trustworthy. This is correctable by re-running experiments with a grouped split and reporting fold variance, so I do not recommend rejection at this stage. The paper would also benefit from a clear enumeration of the 55 features and a comparison against existing AIST++ baselines; without those, the contribution is difficult to position in the literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a feature-engineering pipeline for dance style recognition on AIST++: 3D pose estimation, mesh fitting, floor-plane estimation, then LMA-style features over a sliding window, fed to SVM/RF, with SHAP analysis and an honest out-of-distribution test. The genuinely new pieces are the sliding-window temporal descriptors (initiation, spatial dispersion) and the floor-aware body modeling. That part is reasonable, and the OOD test is more candid than most papers in this area.\n\nThe problem is the headline number. The abstract's 99.18% is actually the average recall of the SVM in Table 1, not an accuracy; Section 4 reports RF 99.68% and SVM 99.07% accuracy. That mismatch needs fixing. More importantly, Section 3.4 says only that 3-fold cross-validation is used, without stating whether the split unit is the source video or the individual sliding window. With 55-frame windows from 60fps video, consecutive windows share 54 frames. If windows are randomly assigned to folds, near-duplicates appear in both training and test sets, letting the classifier memorize video identity rather than generalize across performances. That would fully explain the 99% figures. The OOD drop to 10–82% is consistent with that kind of memorization, though it is not proof of it.\n\nThe other soft spots are secondary: the non-temporal baseline that the sliding-window approach is claimed to beat is never defined; the 55-feature vector is under-specified, with the Space and Time equations having unclear indexing; and no variance across folds is reported, so we cannot tell whether the 99% figure is stable or lucky. These are all fixable in revision.\n\nI want to be clear: this is not a takedown. The pipeline is plausible, the SHAP feature attributions give useful per-style insight, and the authors are upfront about the OOD limitations. But the central accuracy claim is not yet supported as written. A serious referee should send it back for the evaluation protocol to be specified and the numbers reconciled. I would accept it for peer review with the expectation of major revision.","headline":"A plausible LMA feature pipeline whose headline 99% accuracy rests on an unspecified cross-validation split unit; the paper deserves peer review but needs major revision to clarify the evaluation protocol.","tokens_in":8370,"tokens_out":2091,"would_cite":false,"duration_ms":21536,"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":"The paper claims that adding temporal context to Laban Movement Analysis features, pulled from 3D pose and body-shape estimates, classifies ten street-dance styles at roughly 99% accuracy.","keywords":["dance style recognition","Laban Movement Analysis","3D pose estimation","temporal sliding window","explainable AI","SHAP","AIST++","human mesh reconstruction"],"falsifier":"Re-run the exact 55-frame window pipeline with a leave-one-video-out or video-level 3-fold split, so no window of the same source video appears in both training and test folds; if per-class accuracy falls far below the reported 99%, particularly for freestyle genres, the headline accuracy was inflated by window overlap.","tokens_in":7394,"feed_emoji":"🕺","tokens_out":4609,"duration_ms":45817,"temperature":0.7,"pith_summary":"This paper tries to show that dance style recognition stops being a black-box cross-frame problem once Laban Movement Analysis (LMA) features are given explicit temporal context. The authors build a pipeline that estimates 3D body pose and a fitted body mesh from single-view video, estimates the floor instead of assuming it, and derives 55 LMA descriptors covering Body, Effort, Shape, and Space. A sliding window over consecutive frames lets each descriptor carry short-term movement dynamics, and random forest / SVM classifiers then label ten street-dance genres. On the AIST++ single-dancer frontal-view subset they report near-perfect classification, with the highest average recall at 99.18% and random-forest accuracy at 99.68%. If the result holds under proper split protocols, it would make fine-grained style identification explainable, since the SHAP analysis points to effort time and body volume as the dominant cues.","feed_headline":"Time windows push dance-style recognition to 99%","feed_subtitle":"Laban features from 3D pose and body shape classify ten street-dance genres — and explain why.","key_machinery":"Laban Movement Analysis — a choreological language describing movement through Body, Effort, Shape, and Space — is the organizing framework, and the load-bearing mechanism is the sliding window. Each descriptor, including a new 'initiation' detector (joint displacement over a short window exceeding a data-driven threshold), is computed over 55-frame windows, converting single-frame poses into short movement phrases. Floor-aware modeling, combining monocular scene reconstruction, quantile-regression floor fitting, SMPL mesh fitting, and pose-independent anthropometry, supplies metric body and floor references. The result is a 55-dimensional descriptor vector whose classification power is attributed to temporal context, and SHAP is applied to attribute each model decision to individual descriptors.","core_discovery":"On the paper's own terms, the central claim is that LMA's four qualitative dimensions can be quantified with temporal dynamics and that these features alone, without deep learning on raw pixels, separate dance styles nearly perfectly. Concretely, sliding-window features of 55 frames feed SVM and Random Forest classifiers: RF reaches 99.68% accuracy and 99.64% macro F1, while SVM reaches 99.07% accuracy and 99.16% macro F1 across ten AIST++ street-dance genres. The paper further claims that window size matters, with most of the gain coming between 5 and 30 frames and larger windows still helping. Out-of-distribution videos show the descriptors transfer well to structured styles such as lock (81.89% accuracy) but poorly to freestyle genres such as break (10.21%), which the authors attribute to the structured, professional nature of the training corpus.","pith_inferences":["A video-level or dancer-level train/test split would test whether the near-perfect accuracy survives when overlapping windows from the same clip cannot leak across folds; if accuracy collapses, temporal overlap is doing the work.","The same sliding-window LMA descriptors, with floor-aware metric scaling, could be applied to clinical gait assessment or sports technique analysis, where interpretable temporal features are valued as much as accuracy.","The SHAP ranking suggests a compact descriptor subset (effort time, body volume, pelvis jerkiness, ankle kinetics) could reproduce most of the accuracy, reducing the 55-feature vector to a smaller explainable set."],"forward_implications":["Widening the temporal context from 5 to 30 frames produces most of the accuracy gain, so short movement phrases carry the discriminative signal.","Effort Time and Body Volume dominate SHAP attributions, suggesting that temporal energy and spatial extent are the style-separating cues.","Style-specific kinematics emerge: Krump and Pop are quick, Lock shows pause patterns, while Break, Waack, and Middle Hip Hop are less predictable.","Out-of-distribution transfer is uneven, with lock transferring at 81.89% accuracy and break at 10.21%, so structured styles generalize better than freestyle ones."],"supporting_citations":[{"why":"Supplies LMA itself, the four-component framework from which all descriptors are derived.","marker":"[18]"},{"why":"Supplies the AIST Dance Video Database, the 600 single-dancer frontal-view videos and ten style labels used in all experiments.","marker":"[29]"},{"why":"Neural Localizer Fields provide the 3D joint positions in absolute camera space that feed feature extraction.","marker":"[26]"},{"why":"MoGe reconstructs the monocular 3D scene whose point cloud is projected for floor estimation.","marker":"[32]"},{"why":"The SMPL body model is fitted to the estimated joints, enabling anthropometric measurements.","marker":"[20]"},{"why":"Pose-independent 3D anthropometry computes body measurements such as dancer height from the fitted mesh.","marker":"[8]"},{"why":"Supplies the joint weights (alpha_j) that emphasize extremities in the Effort component computations.","marker":"[12]"},{"why":"Provides the SHAP game-theoretic attribution used to explain which LMA features drive each style prediction.","marker":"[23]"}],"fun_headline_variants":["Temporal Laban features classify dance styles at 99.68%","Sliding-window Laban features hit 99.68% on street dance","99.68% dance-style accuracy from temporal Laban features","Temporal context boosts dance-style recognition to 99.68%","Sliding-window Laban features classify dance styles at 99.68%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The near-perfect accuracy assumes that the 3-fold cross-validation splits data so that all overlapping windows from the same video stay inside a single fold; if windows from one clip appear in both training and test folds, the reported numbers could measure temporal overlap rather than generalization to unseen performances.","fun_headline_variants_meta":{"raw":{"variants":["Temporal Laban features classify dance styles at 99.68%","Sliding-window Laban features hit 99.68% on street dance","99.68% dance-style accuracy from temporal Laban features","Temporal context boosts dance-style recognition to 99.68%","Sliding-window Laban features classify dance styles at 99.68%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000563,"raw_usage":{"total_tokens":2645,"prompt_tokens":892,"completion_tokens":1753,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":1656}},"tokens_in":508,"tokens_out":1753,"duration_ms":11873,"temperature":1.0,"reasoning_tokens":1656,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:11:43.918891+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the exact 55-frame window pipeline with a leave-one-video-out or video-level 3-fold split, so no window of the same source video appears in both training and test folds; if per-class accuracy falls far below the reported 99%, particularly for freestyle genres, the headline accuracy was inflated by window overlap.","supporting_citations":[{"cited_title":"In: Proceedings of the 20th International Society for Music Information Retrieval Conference, ISMIR 2019","cited_arxiv_id":null,"evidence_quote":"Supplies the AIST Dance Video Database, the 600 single-dancer frontal-view videos and ten style labels used in all experiments."},{"cited_title":"https://github.com/DavidBoja/ SMPL-Anthropometry (2024)","cited_arxiv_id":null,"evidence_quote":"Pose-independent 3D anthropometry computes body measurements such as dancer height from the fitted mesh."}],"review_version":1}