{"id":"2a3316b8-373d-4fcb-aa13-d5922a343508","arxiv_id":"2504.21154","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A Laban Movement Analysis feature set with temporal windows classifies emotions in contemporary dance videos at 96.85% accuracy on a public dataset, though the evaluation may leak between training and test windows.","lead":"This paper builds a set of 54 movement features based on Laban Movement Analysis and trains classifiers to recognize 12 emotions in contemporary dance videos. It reports up to 96.85% accuracy, but the evaluation setup may inflate results through overlapping time windows.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 96.85% accuracy likely reflects overlapping-window leakage, not generalization; Section 3.4 gives no performance-level or temporal split.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern that I would raise: ungrouped 3-fold cross-validation on overlapping sliding windows almost certainly leaks near-duplicate frames across train and test folds. The paper's Section 3.4 describes the cross-validation as training, testing, and validation sets but provides no grouping by dancer or performance, and Section 4 repeatedly attributes the result to the sliding window without acknowledging the leakage risk. The 'unseen movement sequences' in Figure 2 are not shown to be excluded from training. Given the strong per-class metrics (several above 98%), the absence of error bars, and the lack of released code or splits, the reported accuracy cannot be trusted as a generalization estimate until the experiment is re-run with performance-level or temporal-disjoint folds. This is not an internal inconsistency in the feature derivation; it is a correctness risk in the evaluation of the central claim. The feature descriptors and SHAP interpretation remain potentially valuable, so the appropriate action is to require a corrected evaluation rather than to accept the current numbers. My recommendation therefore matches the reader's REJECT verdict; I set verdict_should_be to UNCHANGED to indicate that my stress-test does not alter that conclusion.","tokens_in":7059,"tokens_out":2359,"duration_ms":29728,"concrete_test":"Ask the authors to re-run the full RF and SVM pipelines using GroupKFold (or Leave-One-Performance-Out), where all windows extracted from one dance performance video are assigned to the same fold, and additionally ensure no training window shares any frame with a test window (e.g., by discarding a margin of at least window_size frames at performance boundaries or by using non-overlapping windows). Report the resulting accuracy, precision, and recall with the same hyperparameter tuning done inside each fold. If RF accuracy drops materially from 96.85% (for example, by more than 10 percentage points), the original result is inflated by window leakage and the central performance claim should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed LMA descriptor set with Random Forest reaches 96.85% accuracy on the Dance Motion Capture dataset. That claim rests on the evaluation protocol in Section 3.4, which says only that 3-fold cross-validation is used with sliding windows of 25 frames. The paper never states that windows are grouped by dancer or by performance, nor that a temporal gap prevents train and test windows from overlapping. Since the features are computed per window and windows are generated by a sliding process, adjacent windows share up to 24 of 25 frames. If the fold split is random at the window level, then near-identical windows from the same performance appear in both training and testing, so the classifier can succeed by recognizing almost the same frames it already saw. This would make the 96.85% figure a measure of memorization of overlapping windows, not of generalization to unseen dance performances. The same issue applies to the comparison in Table 2: if the baseline numbers were obtained under a different protocol, the comparison is not controlled. No code, data splits, or seeds are provided, so the reader cannot verify whether the folds are disjoint at the performance level. The SHAP analysis and feature recipe may be useful, but the headline quantitative result is unsupported unless the evaluation is re-run with proper group splits.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an LMA-based feature extraction framework for emotion recognition from 3D body keypoints in contemporary dance videos. The authors improve existing LMA descriptors with temporal dynamics through a sliding-window approach, introduce new descriptors for the Body, Effort, Space, and Shape components, and train SVM and Random Forest classifiers. They report a highest accuracy of 96.85% (Random Forest) on the Dance Motion Capture dataset with 12 emotions, outperforming several prior methods, and they also use SHAP to analyze feature importance. The main claim is that the proposed descriptors plus a sliding-window scheme yield state-of-the-art emotion recognition accuracy on this dataset.","tokens_in":7206,"tokens_out":4815,"duration_ms":45703,"significance":"If the reported accuracy were obtained under a sound protocol, the paper would be a useful contribution to affective computing and dance analysis: it introduces several novel LMA-inspired descriptors and provides an interpretability analysis via SHAP, which is relatively rare in this area. However, the current evaluation protocol appears to suffer from data leakage due to overlapping windows in cross-validation, which undermines the headline quantitative claim. The comparison with prior work is also not controlled. The strengths are the descriptor design and the explainability analysis, but the central numerical result needs to be re-established.","major_comments":[{"comment":"The cross-validation protocol is not described in enough detail to rule out data leakage. The paper uses 3-fold cross-validation on 25-frame sliding windows, but it does not state that the folds are split by dancer or performance, nor does it impose a temporal gap between train and test windows. Since consecutive windows overlap by up to 24 frames, a random window-level split would place near-duplicate examples in both training and test folds, so the reported 96.85% accuracy could largely reflect memorization of overlapping windows rather than generalization. The authors must specify whether folds are grouped by performance or dancer, or use a temporal split, and report results under that protocol.","section":"§3.4"},{"comment":"The comparison with prior methods is not controlled. The baseline accuracies are taken from published papers ([3], [7], [27]) that may use different data splits, window sizes, and preprocessing. Without re-implementing the baselines under the same protocol, the claim that the proposed method outperforms all existing methods is not established. Additionally, no standard deviations or confidence intervals are reported for any of the accuracies, making it impossible to judge whether the differences are statistically significant.","section":"Table 2"},{"comment":"The description of hyperparameter tuning is ambiguous. The text says '3 fold cross-validation to divide the dataset into 3 sets including training, testing, and validation' and that the best hyperparameters were selected using 'validation accuracy.' If the same fold is used for both model selection and final evaluation, the test accuracy is optimistically biased. A nested cross-validation or a separate hold-out set is needed to avoid leakage from hyperparameter tuning into the reported test accuracy.","section":"§3.4"},{"comment":"The caption claims predictions for 'unseen movement sequences,' but the paper does not specify how these sequences were defined or whether they were excluded from training. This claim is therefore unverifiable. The authors should describe how these sequences were selected and confirm that they were not used in any training or hyperparameter tuning.","section":"Figure 2"}],"minor_comments":[{"comment":"The sentence 'train multiple classifiers' should be 'train multiple classifiers' (grammar) and 'with highest accuracy' should be 'with the highest accuracy.'","section":"Abstract"},{"comment":"The notation is incomplete. Please define P_j(t), Δt, and w, and clarify whether the inequality is evaluated for each frame and aggregated (e.g., as a count) or over the whole window.","section":"§3.3, Eq. (1)"},{"comment":"The denominator ∥P_j(T) - P_j(t_1)∥ can be zero for a stationary joint, leading to division by zero; a small epsilon or fallback should be specified.","section":"§3.3, Eq. (2)"},{"comment":"The 'spatial dispersion' descriptor is described only textually; a precise formula would help reproducibility.","section":"§3.3"},{"comment":"The definition of joint weights α_j is deferred to reference [13] without explaining how they were set; this is important for reproducibility of Eqs. (3)-(5).","section":"§3.3"},{"comment":"'3 fold cross-validation' should be '3-fold cross-validation' and the description of the three sets is confusing; clarify whether a nested CV was used.","section":"§3.4"},{"comment":"Indicate in the caption whether baseline numbers are re-computed by the authors or taken as reported in the original papers.","section":"Table 2"},{"comment":"The plot shows accuracy versus window size but lacks error bars; error bars over cross-validation folds should be added.","section":"Fig. 3"},{"comment":"The statement 'It is evident that the proposed method outperforms all other methods' is too strong without statistical significance testing; report fold-level variance and perhaps a McNemar test.","section":"§4"},{"comment":"The SHAP analysis is presented as 'in-depth explanation,' but the paper only discusses a few top features; consider discussing potential collinearity among features that may affect SHAP values.","section":"§3.5"}],"recommendation":"major_revision","confidential_remarks":"The paper's current evaluation appears to suffer from data leakage due to overlapping sliding windows and an unspecified fold grouping. If the authors cannot provide a proper grouped split or temporal split and re-run the experiments, the paper should not be accepted. The comparison with baselines from other papers is also not controlled and needs to be redone under the same protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the paper has a genuinely useful feature recipe, but the headline 96.85% figure is not trustworthy as reported because the evaluation protocol allows overlapping sliding windows to leak across folds. The reader's stress-test is right.\n\nWhat is actually new: the authors take LMA descriptors from Larboulette & Gibet, extend them with temporal windows, and add a few new statistics (initiation, spatial dispersion, body volume). That is a modest but legitimate extension within an established program. The SHAP analysis is a positive addition—it gives practitioners a sense of which features matter (body volume, effort time) and the figures are interpretable.\n\nWhere it falls down: Section 3.4 says only that 3-fold cross-validation is used with 25-frame sliding windows. The paper never states that folds are split by dancer or performance. Adjacent windows share up to 24 of 25 frames, so window-level random splits place near-duplicate frames in both train and test. That inflates accuracy well above true generalization. There are no error bars and no released code or data splits, so the result is impossible to verify. Table 2 also compares against accuracies from other papers that may have used different protocols—not a controlled comparison. The authors do note the dataset is small (5 dancers, 12 emotions, fixed camera), which limits external validity even with a clean split.\n\nThe feature recipe and the SHAP findings could be salvaged with a proper group-wise split (e.g., leave one dancer or performance out) and reported variance. As is, the central claim is unsupported.\n\nWho this is for: people working on movement-based affect analysis or dance training tools. It deserves a serious referee because the method is clearly described and the idea has merit, but acceptance should require a re-evaluated protocol. I would not cite the current version.","headline":"Useful LMA feature recipe for dance emotion recognition, but the headline 96.85% accuracy is unsupported as reported because overlapping sliding windows almost certainly leak across folds.","tokens_in":7806,"tokens_out":1980,"would_cite":false,"duration_ms":19372,"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":"A 54-feature Laban descriptor set reads 12 dance emotions at 96.85% accuracy.","keywords":["emotion recognition","contemporary dance","Laban Movement Analysis","3D body pose","sliding window features","random forest","SVM","SHAP explainability"],"falsifier":"Split the Dance Motion Capture data by whole performance or by dancer so no clip from a performance appears in both training and test folds, retrain the same descriptors and classifiers, and compare the resulting accuracy; if it falls well below 96.85% (or toward chance), the high figure came from near-duplicate frames appearing in both training and testing, not from genuine emotion recognition.","tokens_in":6784,"feed_emoji":"💃","tokens_out":8848,"duration_ms":85271,"temperature":0.7,"pith_summary":"The paper aims to show that emotion recognition in contemporary dance can be handled by handcrafted, explainable movement descriptors rather than opaque deep models. It builds a 54-feature descriptor set from the four components of Laban Movement Analysis—body, effort, shape, and space—and adds temporal dynamics through sliding windows over 3D joint keypoints. On the Dance Motion Capture dataset, with five dancers expressing 12 emotions, the Random Forest classifier reaches 96.85% accuracy and the SVM reaches 93.90%, outperforming the compared baselines. The authors also use SHAP, a game-theory-based feature attribution method, to show which descriptors drive predictions, arguing that the same features can support performance analysis, dance training, and human–computer interaction.","feed_headline":"Laban motion features read 12 dance emotions at 96.85% accuracy","feed_subtitle":"Spatial and energetic dance descriptors, plus sliding-window timing, outperform earlier classifiers and stay explainable.","key_machinery":"The load-bearing object is the 54-component LMA descriptor vector, computed per 25-frame sliding window from 3D body keypoints. Body descriptors include joint distances and angles, plus an initiation descriptor that flags which joint starts a movement by comparing displacement over a window against a data-driven threshold $\\tau$. Effort descriptors cover Space (windowed path relative to total displacement), Weight (kinetic energy $\\sum \\frac{1}{2} \\alpha_j v_j^2$), and Time (mean acceleration over the window). Shape is represented by body volume from a convex hull, and Space is represented by spatial dispersion and trajectory curvature. Joint weights $\\alpha_j$ emphasize the extremities. The resulting vector is fed to Random Forest and SVM classifiers, and SHAP attributes each feature's contribution to individual predictions.","core_discovery":"On the paper's own terms, the central claim is that adding temporal dynamics to Laban Movement Analysis descriptors yields a compact, interpretable representation of expressive movement that classifies emotional state more accurately than earlier LMA-based methods. The Random Forest on the 25-frame sliding-window descriptor vector reaches 96.85% accuracy across 12 emotion classes; the SVM reaches 93.90%. Individual emotion classes stay roughly balanced, unlike earlier methods where some emotions were recognized much better than others. SHAP analysis identifies body volume and effort time as the strongest contributors, with high values pushing predictions toward expansive, high-energy emotions such as anger or excitement, and low values toward contracted, low-energy states such as sadness or tiredness.","pith_inferences":["A clean split by entire performance or by dancer would settle whether 96.85% is generalization or overlap memorization; the paper does not report such a split.","If dancer-independent accuracy stays high, the same descriptors may transfer to other expressive-motion tasks such as actor performance, physical therapy, or interactive robots, but that transfer is not established here.","The dominance of body volume and effort time suggests a tiny descriptor subset may carry most of the signal; trimming the 54 features and re-measuring would test whether the extra machinery is needed.","The initiation threshold is computed from the full sequence's standard deviation, so an online version would need an adaptive estimate; comparing offline and online thresholds is a natural next experiment."],"forward_implications":["If the 96.85% figure holds, handcrafted LMA descriptors can match or beat deep-learning classifiers on emotion recognition from skeletal pose while needing far less data.","Because emotional state can be read from a 25-frame sliding window, the descriptor vector is usable for real-time emotion feedback in dance training and interactive performance systems.","The SHAP rankings give an actionable finding: body volume and effort time are the features that separate high-energy states like anger or excitement from low-energy states like sadness or tiredness.","The method's balanced per-class results indicate the descriptor set does not collapse distinct emotions into one another, a failure mode noted for earlier LMA features."],"supporting_citations":[{"why":"Supplies the Dance Motion Capture dataset of five dancers and 12 emotions on which all experiments are run.","marker":"[5]"},{"why":"Supplies the 3D body-pose estimator used to obtain the joint keypoints from which all descriptors are computed.","marker":"[23]"},{"why":"Provides the starting point for the Effort component descriptors (Space, Weight, Time) that the paper modifies with temporal dynamics.","marker":"[18]"},{"why":"Defines the joint weights used to emphasize extremities in the Space and Effort feature equations.","marker":"[13]"},{"why":"Supplies the convex-hull algorithm used to compute body volume for the Shape component.","marker":"[26]"},{"why":"Supplies the SHAP game-theory-based method used to explain feature contributions to model predictions.","marker":"[21]"},{"why":"Is the earlier LMA-based emotion classification baseline the paper explicitly compares against and improves on.","marker":"[3]"},{"why":"Provides comparative accuracy numbers on the same dataset for the decision-tree and random-forest baselines in Table 2.","marker":"[27]"},{"why":"Provides another comparative baseline for emotion classification in dance movements in Table 2.","marker":"[7]"}],"fun_headline_variants":["Dance emotion recognition hits 96.85% with Laban temporal features","Temporal Laban descriptors read 12 dance emotions at 96.85%","Explainable Laban analysis decodes dance emotions with 96.85% accuracy","Sliding-window Laban features push dance emotion accuracy to 96.85%","Laban movement analysis recognizes 12 dance emotions, SHAP explains why"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 3-fold cross-validation treats 25-frame windows as independent samples even though consecutive windows overlap, so the 96.85% accuracy may be measuring memorization of near-duplicate frames from the same performance rather than generalization to unseen dancers or performances.","fun_headline_variants_meta":{"raw":{"variants":["Dance emotion recognition hits 96.85% with Laban temporal features","Temporal Laban descriptors read 12 dance emotions at 96.85%","Explainable Laban analysis decodes dance emotions with 96.85% accuracy","Sliding-window Laban features push dance emotion accuracy to 96.85%","Laban movement analysis recognizes 12 dance emotions, SHAP explains why"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1180,"prompt_tokens":799,"completion_tokens":381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":415,"completion_tokens_details":{"reasoning_tokens":277}},"tokens_in":415,"tokens_out":381,"duration_ms":4141,"temperature":1.0,"reasoning_tokens":277,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:10:59.292034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Split the Dance Motion Capture data by whole performance or by dancer so no clip from a performance appears in both training and test folds, retrain the same descriptors and classifiers, and compare the resulting accuracy; if it falls well below 96.85% (or toward chance), the high figure came from near-duplicate frames appearing in both training and testing, not from genuine emotion recognition.","supporting_citations":[{"cited_title":"In: Computer Graphics Forum","cited_arxiv_id":null,"evidence_quote":"Is the earlier LMA-based emotion classification baseline the paper explicitly compares against and improves on."},{"cited_title":"In: Proceedings of the 2nd International Workshop on Movement and Computing","cited_arxiv_id":null,"evidence_quote":"Provides the starting point for the Effort component descriptors (Space, Weight, Time) that the paper modifies with temporal dynamics."},{"cited_title":"https: //github.com/open-mmlab/mmpose (2020)","cited_arxiv_id":null,"evidence_quote":"Defines the joint weights used to emphasize extremities in the Space and Effort feature equations."},{"cited_title":"Nature methods17(3), 261–272 (2020)","cited_arxiv_id":null,"evidence_quote":"Supplies the convex-hull algorithm used to compute body volume for the Shape component."},{"cited_title":"In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R","cited_arxiv_id":null,"evidence_quote":"Supplies the SHAP game-theory-based method used to explain feature contributions to model predictions."},{"cited_title":"IEEE Access8, 124928–124938 (2020)","cited_arxiv_id":null,"evidence_quote":"Provides comparative accuracy numbers on the same dataset for the decision-tree and random-forest baselines in Table 2."},{"cited_title":"Computer Animation and Virtual Worlds32(6), e1996 (2021)","cited_arxiv_id":null,"evidence_quote":"Provides another comparative baseline for emotion classification in dance movements in Table 2."}],"review_version":1}