{"id":"19028822-ccec-4be9-aba8-34fe02e9c126","arxiv_id":"2505.19186","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A yoga posture correction system that combines LSTM pose recognition with BiLSTM angle forecasting to flag deviations in real time on edge hardware.","lead":"PosePilot is an edge-deployed system that recognizes six yoga poses and issues per-frame correction feedback by comparing joint angles with a BiLSTM forecast of a correct pose trajectory. It adds a new 336-video yoga dataset and reports 97.52% recognition accuracy with low forecasting error on a Raspberry Pi 4.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The correction claim rests on an unverified premise: the in-house training videos are treated as correct executions, and the only reported evaluation is forecasting MSE, not error-detection accuracy. Expert validation of the dataset and a detection-level evaluation are needed.","rationale":"The reader identified the same weakest assumption. My stress-test pass agrees and sharpens it: the paper's only quantitative correction result is forecasting MSE, which cannot distinguish 'learned correct trajectory' from 'learned average amateur trajectory.' The strongest claim in the abstract — instant corrective feedback ensuring precise posture adjustments — would be true only if the training distribution itself is correct posture. Nothing in the manuscript establishes that; Section 3.1 explicitly critiques previous datasets for lacking professional ground truth but does not claim expert validation for PosePilot's own data. I therefore cannot find a separate, more load-bearing concern; the real-time/latency issue and arbitrary 1.5-SD threshold are secondary because even a perfectly real-time system would be correcting toward an unvalidated target. The proposed test — expert rating, retraining, and detection-level precision/recall — directly settles whether the forecast is an ideal trajectory or just a dataset average. Since the reader's conditional verdict already captures this gap, I recommend no change to the verdict, though the paper would be substantially strengthened by adding such an evaluation.","tokens_in":8596,"tokens_out":4591,"duration_ms":44119,"concrete_test":"Recruit two certified yoga instructors to independently rate each of the 336 training videos (or a stratified random subset) for correctness of the six asanas and their sub-poses; retain only clips both raters mark as correct. Retrain the BiLSTM correction model on this expert-validated subset. Then create a held-out test set with expert-labeled correct and deliberately incorrect performances and compute per-angle detection precision/recall at the 1.5-SD threshold, plus the flag rate on expert-validated correct performances. If the forecast trajectory shifts materially after validation, or if precision/recall is near chance, the correction claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PosePilot's feedback mechanism (Sec. 3.3) flags any of nine joint angles that deviates more than 1.5 standard deviations from a BiLSTM one-step forecast of the next angle vector. For this to constitute 'precise posture corrections' rather than statistical anomaly detection, the forecast must approximate the correct/ideal trajectory. That requires training clips to be expert-verified as correct executions. Section 3.1 says the dataset is 'high-quality' and criticizes prior datasets for using non-professional practitioners, but it never states that the 14 participants' performances were checked by a yoga expert; the only expert mention (Sec. 3.3) is for selecting the nine angles, not for validating ground truth.\n\nThe reported evaluation in Sec. 4.2 also does not test the correction claim: average MSE 0.00138 on a held-out split measures how well the BiLSTM predicts the next angle within the same dataset distribution, not whether a flagged deviation corresponds to a real error. There is no precision/recall, no expert-labeled error set, and the 'deliberately incorrect asanas' experiment is qualitative (one figure). Low MSE is consistent with a model that has simply memorized the average amateur trajectory. If the training data contains imperfect poses, feedback steers users toward that average rather than a verified safe posture, invalidating the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"PosePilot proposes an edge-AI system for yoga pose recognition and corrective feedback. It introduces an in-house multi-angle video dataset of six asanas from 14 participants, extracts 680 joint angles via MediaPipe, uses a key-frame selection heuristic and an LSTM with attention for classification (reported accuracy 97.52%, F1 0.99), and a two-layer BiLSTM with attention that forecasts the next nine-angle vector; deviations above 1.5 standard deviations are flagged as errors and displayed as corrective feedback. The authors also report deployment on a Raspberry Pi 4 with INT8 quantization. The main claimed contribution is 'instant corrective feedback at every stage of a movement,' and the paper argues that prior systems only address final-pose correction rather than sub-pose feedback.","tokens_in":8898,"tokens_out":4731,"duration_ms":41675,"significance":"If the correction claim were fully validated, PosePilot would be a useful contribution: it addresses an underexplored problem (sub-pose correction, not just final-pose classification), reports a new multi-angle yoga dataset, and includes an edge-deployment study with a code repository link. The recognition component is reasonably evaluated with 10-fold cross-validation and reports a high F1 score. The paper's strength is the integration of angle-based LSTM recognition, forecasting-based correction, and edge quantization. However, the central correction claim is not yet established by the evidence presented: the ideal-pose ground truth is internal to the training set, and no detection-level metrics are reported. The significance is therefore conditional on the additional validation requested below.","major_comments":[{"comment":"The correction model's core premise is that the BiLSTM forecast represents the correct/ideal trajectory, but the paper never establishes that the in-house training videos are expert-validated correct executions. Section 3.1 criticizes prior datasets for using non-professional practitioners, yet it only states that the 14 participants were recorded, not that their performances were checked by a yoga expert; the only expert involvement reported (Section 3.3) is in selecting the nine angles for correction. If the training clips contain imperfect or non-representative executions, the 1.5-standard-deviation flags steer users toward the dataset's average movement rather than a verified safe posture. Please state explicitly whether expert validation was performed; if not, either add it or re-frame the system as deviation-from-learned-pattern feedback and provide external validation.","section":"Section 3.1 and 3.3"},{"comment":"The evaluation of the correction model does not test the correction claim. The only quantitative metric is forecasting MSE of 0.00138 on a held-out split (Table 2), which measures the model's ability to predict the next angle within the same dataset distribution; it does not measure whether a flagged deviation corresponds to a real postural error. There is no precision/recall against expert-labeled error frames, no comparison with a baseline forecaster (e.g., persistence or linear extrapolation), and the 'deliberately incorrect asanas' experiment is presented only as a qualitative figure without detection accuracy. The MSE is also not interpretable without the angle scale or a baseline. Please add a detection-level evaluation with expert-labeled errors, report precision/recall or F1 for the error flag, and compare the forecaster against a simple baseline.","section":"Section 4.2"},{"comment":"The key-frame hyperparameter k is selected on test data: the text states 'During testing, we found that k=10 yielded optimal results.' This makes the reported recognition accuracy potentially optimistic and the selection procedure non-standard. In addition, the description of data augmentation does not clarify whether augmentation is applied before or after the train/test split; if before, augmented copies of test sequences can leak into training. Please use a validation split or nested cross-validation for hyperparameter selection and describe the augmentation/split order explicitly.","section":"Section 3.2"},{"comment":"The real-time claim is only partially supported: the pose-correction model alone runs at 6.42 FPS after INT8 quantization, but feature extraction on video frames is excluded and is stated to need further optimization. The end-to-end latency (camera capture + MediaPipe keypoint extraction + angle computation + recognition + correction + GUI rendering) is not reported. Since the abstract promises 'instant corrective feedback at every stage of a movement,' please provide an end-to-end frame-rate or latency measurement on the deployed hardware, or soften the real-time claim accordingly.","section":"Section 4.3"}],"minor_comments":[{"comment":"In Table 1, the first two rows both list 'Yadav et al. [27]' as the dataset, which appears to be a copy-paste error; the second row should cite the Infinity Yoga Tutor entry.","section":"Table 1"},{"comment":"Equation (2a) uses the symbol p both as the interval length and in the summation bounds; please clarify the window size (e.g., p=5 for t-2 to t+2) and define the index range unambiguously.","section":"Equations 2a-2c"},{"comment":"The correction target range is described as 'within one standard deviation from the predicted value,' but it is not stated whether the standard deviation is computed per angle, per pose, or across the dataset; please specify.","section":"Section 3.3"},{"comment":"The reported F1 score of 0.99 and accuracy of 97.52% are not accompanied by confidence intervals or per-class breakdowns; please include these to support the recognition claim.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's strongest contributions are the new dataset and the edge deployment study; the correction mechanism is essentially a statistical anomaly detector whose instructional validity is unproven. The reader's concern about circularity is well-founded: the 'ideal pose' is defined by the model's own training distribution. If the authors add expert validation of the dataset and a detection-level evaluation with baselines, the paper could become acceptable; without these, the central 'personalized correction' claim remains unsupported. I do not see a citation or novelty problem; the main gap is evidential."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before you see it cited as evidence that per-frame yoga correction works. The recognition half is a reasonably solid engineering effort: joint angles from Mediapipe, an LSTM classifier with key-frame selection, 97.52% accuracy on their own 336-clip dataset, and they report cross-validation and out-of-dataset generalization to Yadav et al. That part is credible, and the new four-angle video dataset is a real resource, assuming it actually gets released. The paper also deserves credit for attempting per-frame temporal correction, which prior systems largely skip, and for shipping quantized models that run on a Raspberry Pi 4. The correction claim is where it gets soft, and the softness is structural, not cosmetic. The BiLSTM forecaster is trained on videos that are assumed to be correct, but the paper never states that the 14 participants were expert-validated. Section 3.1 criticizes earlier datasets for using non-professionals and then describes its own participants with no mention of expert checking. So the 'ideal pose' is just the average trajectory of the training set. That could be fine if the participants were verified, but we are not told. The evaluation also misses the point: the only quantitative metric is MSE 0.00138 on a forecasting task, which measures fit to the dataset, not whether flagged deviations are actual errors. There is no precision/recall, no expert-labeled error set, and the deliberate-error test is one qualitative figure. Low MSE is consistent with memorizing amateur averages. I also note the real-time claim is undercut by their own numbers: pose correction inference is 6.42 FPS, which is not real-time for video, though they acknowledge feature extraction needs optimization. The k=10 key-frame hyperparameter is selected on the test set, and the correction model has no baseline forecaster, so even the forecasting result lacks context. These are fixable: get expert labels for the training clips, add a detection-level evaluation with deliberately corrupted poses, and report correction latency end-to-end. But as is, the central correction claim is not established. It is an interesting prototype, not a validated correction system. I would send it to peer review because the dataset and the recognition pipeline are useful, and the correction idea is worth scrutiny, but I would not cite the correction result until the evaluation is redone. Give it to a referee who will push on ground truth. My reading group verdict is maybe; we would discuss the dataset more than the conclusion.","headline":"Plausible recognition pipeline plus a genuinely new dataset, but the pose-correction claim rests on forecasting MSE and an unverified assumption that the training videos are correct executions.","tokens_in":9414,"tokens_out":949,"would_cite":false,"duration_ms":11790,"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":"PosePilot treats posture correction as continuously forecasting the next joint-angle vector, flagging deviations at every step of a yoga pose.","keywords":["yoga pose correction","pose recognition","BiLSTM","multi-head attention","edge AI","joint angles","corrective feedback","temporal forecasting"],"falsifier":"Have yoga instructors independently annotate the frames where deliberately incorrect asanas deviate from correct form, then compare their labels with the model's 1.5-standard-deviation flags; if the flags and expert labels do not overlap substantially, the correction claim is not supported.","tokens_in":8417,"feed_emoji":"🧘","tokens_out":6320,"duration_ms":60284,"temperature":0.7,"pith_summary":"The paper sets out to show that automated posture correction can be recast as trajectory forecasting rather than final-pose classification. Using Yoga as the test case, a two-layer bidirectional LSTM learns the joint-angle history of correctly performed asanas and predicts the next angle vector, so any angle that strays more than 1.5 standard deviations from the forecast is flagged with a direction to adjust. This gives feedback continuously through a movement rather than only at the end, which the paper identifies as the missing piece in prior correction systems. If the claim holds, users can receive personalized, on-device coaching without internet or cloud processing.","feed_headline":"Forecasts the next joint angle to correct yoga form live","feed_subtitle":"Per-instant angle forecasts catch errors mid-pose and run locally on a Raspberry Pi-class device.","key_machinery":"The load-bearing object is the forecasted joint-angle vector $\\hat{p}_t$ for nine body angles: left and right shoulders, elbows, hips, knees, and the neck. The two-layer BiLSTM consumes standardized sequences of these angles and predicts the next vector; the error module compares the practitioner's actual angles with $\\hat{p}_t$ and flags any angle exceeding 1.5 standard deviations, using the vector difference to show the direction of the needed adjustment. This converts correction into a next-frame forecasting task, which is what enables per-instant feedback rather than a single end-of-pose verdict.","core_discovery":"On the paper's own terms, the central claim is that pose correction becomes a sequence-forecasting problem: a two-layer BiLSTM with multi-head attention is trained on the temporal pattern of nine joint angles sampled from correctly performed asanas, and at each time step it forecasts the next joint-angle vector. Any of the user's nine angles that deviates by more than 1.5 standard deviations from that forecast is reported as an error requiring adjustment, with the direction of correction coming from the vector difference. Because the forecast is produced at every frame, the feedback applies to intermediate sub-poses, not just the final posture. The recognition module classifies the ongoing asana and selects one of six per-pose correction models. On an in-house dataset of 336 videos (14 participants, six asanas, four camera angles), the paper reports 97.52% recognition accuracy and average correction-model MSE of 0.00138; quantized to 8-bit integers on a Raspberry Pi 4, recognition runs at 330.65 FPS and correction at 6.42 FPS, with reported accuracy losses of about 1.0-1.3%.","pith_inferences":["An implicit test the paper leaves open is whether the 1.5-standard-deviation flagging rule agrees with expert judgment: comparing system flags on deliberately incorrect asanas with annotations from yoga instructors would settle whether the threshold marks the same mistakes a human coach would.","A corollary not developed in the paper is that errors in the recognition stage would route the user to the wrong correction model, so end-to-end correctness depends on recognition confidence and would benefit from an explicit uncertainty estimate.","The statistical notion of 'correct' here is the center of the training distribution; with a larger, expert-validated, body-diverse dataset, the learned trajectory could become a coaching-grounded standard rather than only the average of the recorded movements."],"forward_implications":["Users receive correction signals throughout a movement: each frame that deviates from the forecasted angle trajectory is flagged, so intermediate sub-poses are corrected rather than only the final pose.","Feedback is angle-specific and directional: the system identifies which of the nine monitored body angles is off and in which direction to adjust, instead of merely classifying the pose as incorrect.","The recognition-plus-correction loop is designed to run on a Raspberry Pi-class edge device at usable frame rates, which would make personalized coaching available offline and keep exercise video local.","Because the correction model is selected per recognized asana, the same architecture can be retrained for other sequence-based physical activities named in the paper, such as rehabilitation and sports coaching."],"supporting_citations":[{"why":"Supplies the per-frame keypoint extraction from which all 680 joint angles and the reduced 9-angle correction features are computed.","marker":"[15]"},{"why":"Provides the prior real-time yoga dataset and recognition pipeline used as the baseline and cross-dataset evaluation target.","marker":"[27]"},{"why":"The CARE system is the main contrast for the paper's claim: it corrects only the final pose, whereas PosePilot addresses intermediate sub-poses.","marker":"[7]"},{"why":"An existing explainable pose-correction system, used as the literature-table comparison point for correction capability.","marker":"[6]"},{"why":"Represents threshold-based real-time feedback systems that the paper argues cannot personalize corrections or suggest how to adjust.","marker":"[12]"}],"fun_headline_variants":["Forecast next joint angle to correct yoga form live","Edge AI predicts joint angles to fix yoga poses mid-move","Angle forecasting on edge corrects yoga posture in real time","BiLSTM with attention forecasts angles to spot yoga errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the in-house videos are faithful executions of each asana, yet the paper does not report expert verification of the 14 participants' performances; if those recordings contain imperfect form, the forecast guides users toward the dataset's average movement rather than a verified correct posture.","fun_headline_variants_meta":{"raw":{"variants":["Forecast next joint angle to correct yoga form live","Edge AI predicts joint angles to fix yoga poses mid-move","Angle forecasting on edge corrects yoga posture in real time","BiLSTM with attention forecasts angles to spot yoga errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000282,"raw_usage":{"total_tokens":1697,"prompt_tokens":1001,"completion_tokens":696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":629}},"tokens_in":617,"tokens_out":696,"duration_ms":4841,"temperature":1.0,"reasoning_tokens":629,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:18:59.256763+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have yoga instructors independently annotate the frames where deliberately incorrect asanas deviate from correct form, then compare their labels with the model's 1.5-standard-deviation flags; if the flags and expert labels do not overlap substantially, the correction claim is not supported.","supporting_citations":[{"cited_title":"Neural Computing and Applications31, 9349–9361 (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the prior real-time yoga dataset and recognition pipeline used as the baseline and cross-dataset evaluation target."},{"cited_title":"International Journal of Industrial Ergonomics99, 1035–40 (2024)","cited_arxiv_id":null,"evidence_quote":"Represents threshold-based real-time feedback systems that the paper argues cannot personalize corrections or suggest how to adjust."}],"review_version":1}