{"id":"205b62ed-1a98-48e5-b372-e5279ec31f07","arxiv_id":"2608.12145","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-module RGB-video rehabilitation pipeline is proposed, with a 96.45% squat-quality classifier and an adopted STARS predictor, but no end-to-end or clinical validation is performed.","lead":"This paper couples an exercise-quality classifier with a motion-prediction model to create a feedback pipeline for home rehabilitation from plain RGB video. It reports 96.45% mean-class accuracy on squat exercises, but the two parts were never tested together and the motion-prediction numbers come from prior papers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The joint-level feedback signal is computed from the motion predictor, but only evaluated on Vicon ground-truth skeletons; no evidence shows it survives noisy marker-free pose estimates, so the marker-free pipeline claim is untested.","rationale":"The reader's conditional verdict is sound, and I agree with its spirit: the pipeline is not yet validated as an autonomous telerehabilitation system. However, I see the most load-bearing concern slightly differently from the reader's primary choice. The paper's own Section VIII explicitly acknowledges that the JPE signal is not clinically validated and that no end-to-end evaluation has been done. That honesty is to its credit, but those limitations are declared future work. The deeper technical problem is that the motion-prediction module, which generates the joint-level feedback, has only ever been evaluated on high-accuracy Vicon skeletons. Since the pipeline is framed as operating on marker-free RGB video, the observed poses in Eq. 8 will come from MediaPipe or OpenPose, not from mocap. The classifier at least runs on skeletons extracted from PROZIS video, but the predictor has no such evidence. The qualitative demo uses STS-GCN on a ground-truth walking sequence, and the headline MPJPE numbers are taken from the STARS paper rather than measured here. If pose-estimation noise is comparable to or larger than the prediction error, the color-coded deviation signal is meaningless even if both modules individually achieve their published benchmarks. This is not an ad hominem or a demand for clinical validation; it is a request for evidence that the integrated marker-free pipeline can actually compute a meaningful JPE. The proposed transfer experiment would settle this directly. I therefore keep the reader's conditional verdict unchanged.","tokens_in":11087,"tokens_out":6556,"duration_ms":63018,"concrete_test":"Run MediaPipe/OpenPose on Human3.6M walking sequences, map the output joints to the predictor's 22-joint format, and compute JPE (Eq. 8) using predicted vs. estimated observed poses, then compare with JPE computed against ground-truth skeletons. If pose-estimation noise changes color-band assignments on a substantial fraction of frames (for example, more than 20% of frames), or if the per-joint noise floor at distal joints exceeds the system's yellow threshold, the marker-free claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an integrated telerehabilitation pipeline 'operating on marker-free RGB video' (Abstract, Section III). The joint-level feedback depends on Eq. 8, comparing predicted poses to observed poses from the chosen pose estimator. But every quantitative motion-prediction result is on Human3.6M Vicon mocap skeletons: Table II is quoted from Xu et al. [12] (Section VII.D), and the qualitative demo in Section VII.E uses STS-GCN on a ground-truth walking sequence. The PROZIS dataset is explicitly unusable for the graph predictor because it lacks kinematic-tree metadata (Section VI.B). Consequently, the predictor has never been run on the noisy MediaPipe/OpenPose skeletons that the pipeline specifies. Pose-estimation noise on RGB video is typically centimetre-scale, especially at distal joints, while the predictor's long-horizon error is 75.8 mm at 560 ms; if these are comparable, the color-coded overlays will flag estimation noise rather than movement deviations. This gap is load-bearing because the system's novel contribution is joint-level feedback from marker-free video, not just the isolated benchmarks. The clinical-validity issue identified by the reader is important, but it is a separate future-work item; the marker-free transfer problem undermines the technical feasibility claim even before clinical meaning is considered.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-module telerehabilitation pipeline that takes marker-free RGB video, extracts skeletons, and combines (i) a self-attentive BiLSTM with MMD-NCA metric learning for holistic correct/incorrect exercise classification and (ii) a graph-based motion predictor whose per-joint position error (JPE, Eq. 8) is color-coded into spatially localized joint-level feedback. The classifier is evaluated on squat sequences from the proprietary PROZIS dataset and reported at 96.45% mean-class accuracy; the motion-prediction module is represented by literature-reported STARS results on Human3.6M (75.8 mm MPJPE at 560 ms) and by a qualitative STS-GCN-based feedback demonstration on a walking sequence. The paper explicitly positions the work as component-level feasibility and lists as future work end-to-end integration, clinical validation of the JPE signal, and evaluation with users and therapists (Section VIII).","tokens_in":11368,"tokens_out":4776,"duration_ms":45580,"significance":"The architectural blueprint is reasonable and, if the component results were fully supported, would be a useful step toward automated, marker-free rehabilitation feedback. The paper is honest about its limitations and provides training/validation curves for the squat classifier, which is a concrete piece of evidence. However, the evidence currently falls short of the system-level claim: the headline motion-prediction numbers are quoted from prior work rather than reproduced, the feedback signal is demonstrated only on ground-truth Vicon skeletons, and the JPE signal is acknowledged to lack therapist-annotation validation. These gaps are load-bearing because the paper's claimed contribution is the integration of quality labels and joint-level deviation signals from marker-free video. Because the missing experiments and claim qualification are within reach of a revision, the work is potentially publishable, but the current manuscript overstates what has been demonstrated.","major_comments":[{"comment":"Table II is explicitly described as 'as presented in [12]' and Section VII.D states that the STARS benchmark is taken from [12]; no STARS training, inference, or reproduced evaluation is reported in this paper. The abstract and conclusion nevertheless present the 75.8 mm MPJPE at 560 ms as a result of the proposed system. Furthermore, the qualitative feedback demonstration in Section VII.E uses STS-GCN, not the STARS model adopted in Section V.C, so the behavior of the actual predictor inside the feedback loop is never demonstrated. Please either report a reproduced STARS evaluation on Human3.6M or clearly label these values as literature-reported in every summary, and either run the feedback visualization with STARS or explicitly state that the demonstration uses a surrogate model.","section":"VII.D, Table II, and VII.E"},{"comment":"The pipeline in Section III specifies MediaPipe/OpenPose skeletons from marker-free RGB video, and Eq. (8) computes JPE between predicted and observed poses, so pose-estimation noise directly enters the joint-level feedback signal. No experiment exercises the predictor under this marker-free condition: the only predictor-on-video evidence is Section VII.E, which uses ground-truth Vicon skeletons from Human3.6M, and PROZIS cannot support graph-based predictors because it lacks kinematic-tree metadata (Section VI.B). Since typical monocular 3D pose estimation error at distal joints is on the order of the long-horizon MPJPE values in Table II, the claimed marker-free joint-level feedback is not established. Please add an experiment with noisy or pose-estimator-derived skeletons (for example, running MediaPipe or OpenPose on Human3.6M RGB frames, or adding calibrated noise to the input skeletons) and report how JPE and the color-coded overlays are affected.","section":"III, V.D, VI.B, and VII.E"},{"comment":"The 96.45% mean-class accuracy on squats is the headline classification result, but it comes from a single 70/15/15 split of a proprietary dataset with no cross-validation, confidence intervals, or per-class sample breakdown. The companion classes with 70-300 samples fail to converge, so the reliability of the squat result is unclear. Please provide repeated-split or cross-validated statistics with error bars, a per-class performance table, and details of the class distribution; if such statistics are unavailable, the claim should be explicitly downgraded to a single-split feasibility observation.","section":"VII.C and Fig. 2"},{"comment":"The paper itself states in Section VIII that the joint-level error signal is derived from motion-prediction error rather than therapist-annotated deviation ground truth and has not been clinically validated. Because the title and abstract present 'joint-level performance assessment' as a contribution, this is a load-bearing limitation rather than merely future work. The manuscript should either reframe the contribution as a candidate visualization mechanism that is not yet shown to correspond to clinically meaningful deviations, or add a small expert-annotation comparison (even on a handful of sequences) to connect JPE magnitude to therapist judgment.","section":"V.D and VIII"}],"minor_comments":[{"comment":"Equation (5) is hard to parse: X and X^+ are both described as samples from the same action category, and the anchor argument in f(.) is not clear. Please define whether f(X) denotes the embedding of a single sample or of a sample set, and align the notation with the text explaining the loss.","section":"IV.C, Eq. (5)"},{"comment":"The claim that interpolation 'was also evaluated but yielded marginally inferior results (≤2 percentage points in mean-class accuracy)' is not supported by any table or quantitative comparison in Section VII. Please report this comparison or remove the claim.","section":"III"},{"comment":"The layer-normalization equations use the symbol g_{c,j}^t without defining it, so the mean and variance computation in Eq. (1) is not reproducible. Please define all quantities in the equations.","section":"IV.A, Eqs. (1)-(2)"},{"comment":"Table I reproduces FPR values from Coskun et al. [5], but the caption does not say so, unlike Table II. Please add a caption note so that readers can distinguish reproduced numbers from new experimental results.","section":"Table I caption"},{"comment":"The color thresholds that map JPE to green/yellow/orange/red are never stated. Since the color-coded overlay is central to the feedback claim, please report the threshold values used in the figures and state whether they are fixed or data-dependent.","section":"VII.E and Figs. 3-4"},{"comment":"The speed-robustness experiment with 50% and 25% subsampling is purely qualitative; no error values are reported for the subsampled variants. Please include quantitative JPE or MPJPE comparisons, even at the sequence level, to support the tempo-invariance claim.","section":"VII.E"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest and clearly written, but the central quantitative claims rest on quoted prior results and a proprietary dataset, which makes independent verification difficult. The most important fix is a marker-free/noisy-skeleton experiment for the feedback pipeline; without it, the abstract's 'operating on marker-free RGB video' claim is not supported. If the authors can add that experiment and qualify the literature-reported numbers, the paper could be acceptable as a feasibility study for this venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main thing to know: this paper is honest about what it hasn't done, and what it hasn't done is the actual integration. The abstract describes a two-module pipeline for telerehabilitation feedback from marker-free RGB video, but the motion-prediction module is only ever evaluated on Vicon ground-truth skeletons (Human3.6M), and Table II is explicitly quoted from prior work rather than reproduced. The qualitative demo uses STS-GCN on a ground-truth walking sequence. So the load-bearing claim that the system can generate joint-level deviation signals from noisy RGB pose estimates is simply untested.\n\nWhat the paper does well: it is clearly written, well structured, and unusually transparent about its limitations. Section VIII lists the missing end-to-end evaluation, the absence of therapist-annotated deviation ground truth, and the PROZIS kinematic-tree constraint that prevents running the graph predictor on rehabilitation data. The classifier protocol for the one new result—96.45% mean-class accuracy on PROZIS squats with MMD-NCA—is reproducible in structure, with training curves and a 70/15/15 split. That is a legitimate new benchmark result, though on a proprietary dataset, one exercise class, and no cross-validation or error bars.\n\nThe soft spots are real but mostly acknowledged. The JPE formulation is a trivial per-joint decomposition of MPJPE, which the authors admit is prediction error, not clinically meaningful deviation. The color thresholds are arbitrary. The non-squat exercises failed to converge, which the authors attribute to data scarcity—plausible but not demonstrated. Most importantly, the marker-free gap is not just a future-work item; it undermines the central technical feasibility claim. If pose-estimation noise is comparable to the prediction error (the long-horizon number is ~75 mm), the red/yellow overlays could flag estimation artifacts rather than movement faults. The stress-test note is correct on this point.\n\nWho is this for? A workshop or short-conference audience interested in telerehabilitation pipelines would get value from the classifier result and the honest framing. It is not yet a journal-level validated system paper.\n\nRecommendation: this deserves a serious referee, but with the expectation of major revision. The authors should either run the predictor on noisy skeletons (e.g., a public RGB+mocap dataset) or reframe the paper to claim only the classifier evaluation plus a proposed feedback module. As it stands, the title and abstract overstate what is shown.","headline":"Honest, well-scoped paper whose central integrated-pipeline claim is untested: the motion predictor has never run on marker-free pose estimates, and the only new result is one classifier benchmark on PROZIS squats.","tokens_in":11907,"tokens_out":1667,"would_cite":false,"duration_ms":17170,"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 two-module pipeline on marker-free RGB video can label a rehabilitation exercise as correct or incorrect and, in parallel, produce per-joint deviation maps by comparing predicted and observed skeletons.","keywords":["telerehabilitation","exercise quality assessment","human motion prediction","skeleton-based action recognition","MMD-NCA metric learning","joint-level feedback","marker-free RGB video","pose forecasting"],"falsifier":"Collect a set of exercise repetitions with both therapist fault labels and the system's predicted-vs-observed per-joint errors: if the red-threshold joints do not match the joints therapists flag, or if clearly faulty movements produce no elevation in JPE, the central feedback claim is refuted. A simpler check is to measure whether JPE stays low for correct repetitions across the full prediction horizon; if it rises sharply even for correct motion, the overlay would mislead users.","tokens_in":10884,"feed_emoji":"🩺","tokens_out":5815,"duration_ms":49180,"temperature":0.7,"pith_summary":"This paper is trying to establish that a rehabilitation coach can be built from two complementary perception modules working on ordinary RGB video: one that labels each repetition as correctly or incorrectly executed, and one that predicts how the skeleton should move a fraction of a second ahead and turns the difference between predicted and actual joint positions into spatially localized feedback. If the claim holds, home-based patients could receive per-repetition quality labels plus a color-coded joint-level overlay indicating where their movement deviates, without marker suits or a therapist in the room. The system's numbers come from two separate benchmarks: 96.45% mean-class accuracy on squat sequences and a mean per-joint prediction error of 75.8 mm at a 560 ms horizon. The paper is explicit that this is a technical feasibility result, not a clinically validated feedback system.","feed_headline":"One pipeline grades rehab exercises and pinpoints joint deviations","feed_subtitle":"Marker-free video feedback could let patients exercise at home with joint-level correction, no therapist in the room.","key_machinery":"The load-bearing object is the per-joint position error (JPE), the per-joint, per-frame Euclidean distance between predicted and observed 3D skeleton positions defined in Eq. (8). It converts a motion-prediction model into a feedback generator by localizing deviation in space and time. On the classification side, the mechanism is a self-attentive BiLSTM trained with the MMD-NCA loss of Eq. (5), which measures distributional divergence between class embedding distributions via maximum mean discrepancy with a Gaussian kernel mixture; this yields tempo-robust holistic quality labels. On the prediction side, the mechanism is the STARS generative model built on an STS-GCN backbone, operating in the DCT frequency domain with learnable spatial and temporal anchors and kinematic-tree-guided adjacency pruning; its predicted trajectories fill the JPE map.","core_discovery":"The central claim is that rehabilitation feedback can be decomposed into two complementary signals—a holistic correct/incorrect exercise label and a spatially localized per-joint deviation map—and that both can be generated from marker-free RGB video. The exercise-quality module classifies full squat repetitions with 96.45% mean-class accuracy on the PROZIS dataset, using a self-attentive bidirectional LSTM trained with MMD-NCA metric learning. The prediction module adopts STARS, a generative graph-based motion predictor, which reaches a mean MPJPE of 75.8 mm at 560 ms on Human3.6M and outperforms graph and recurrent baselines at every tested horizon. The per-joint position error $L_{\\mathrm{JPE}}(v,k)$, Eq. (8), between predicted and observed skeletons is thresholded into green/yellow/orange/red overlays. The paper is explicit that this establishes technical feasibility of the components, not a clinically validated feedback system.","pith_inferences":["If JPE is later validated against physiotherapist-annotated faults, the same two-module skeleton could generate feedback for exercises beyond squats and walking, provided a motion prior can be learned for each exercise.","The feedback semantics could be sharpened by conditioning the motion predictor on the quality label, so the system compares the user's motion against a predicted correct trajectory rather than against the user's own observed trajectory.","A testable extension is to replace the observed ground-truth trajectory in Eq. (8) with a therapist-approved ideal trajectory; if prediction error against that ideal correlates with clinician ratings better than prediction error against observed motion, the feedback signal becomes clinically meaningful.","The gap that forces separate evaluation—PROZIS lacking kinematic-tree metadata and Human3.6M lacking rehabilitation annotations—could be closed by adding kinematic skeletons and therapist labels to rehabilitation exercise videos, enabling an end-to-end test of the whole feedback loop."],"forward_implications":["A patient at home could receive per-repetition feedback stating both whether a repetition was correctly executed and which joints deviated, using only an ordinary camera.","The classifier's 96.45% accuracy on squats indicates that the metric-learning approach works on rehabilitation data when enough balanced sequences are available, while the paper's own convergence failures set a data requirement of several hundred balanced sequences per exercise class.","The STARS predictor's advantage over graph and recurrent baselines at all horizons up to 1000 ms means the anchor-based generative formulation can supply joint-level error signals over the time range relevant for exercise feedback.","Tempo robustness is supported by subsampling a walking sequence to 50% and 25% of its original frame rate without retraining or dynamic time warping, suggesting the feedback can tolerate the slower execution common in rehabilitation.","Because end-to-end inference takes 2 to 5 seconds, the system is suited to per-repetition feedback, and the paper recommends buffering at least 10 frames, or 400 ms, of active movement before activating the overlay.","The two modules are validated separately, so the paper's claim is about component feasibility; end-to-end evaluation on a single rehabilitation dataset remains future work."],"supporting_citations":[{"why":"Supplies the MMD-NCA metric-learning loss and the self-attentive BiLSTM architecture that the classification module adopts, along with the baseline FPR comparison.","marker":"[5]"},{"why":"Provides the STS-GCN backbone used by the motion predictor and the STS-GCN MPJPE values compared in the benchmark table.","marker":"[11]"},{"why":"Provides the STARS anchor-based generative predictor adopted for joint-level prediction and the MPJPE benchmark numbers that support the central prediction claim.","marker":"[12]"},{"why":"The LTD-50-25 DCT-domain graph model is a baseline in the motion-prediction comparison, representing a strong prior trajectory-dependency method.","marker":"[10]"},{"why":"Human3.6M is the dataset on which the motion-prediction module is benchmarked, supplying the standardized evaluation protocol.","marker":"[14]"},{"why":"The PROZIS Challenge dataset provides the annotated squat sequences used for the exercise-quality classification evaluation.","marker":"[16]"},{"why":"Related PROZIS workout video processing work supplies context for the dataset's origin and annotation scheme.","marker":"[17]"}],"fun_headline_variants":["Squat classification at 96% accuracy from marker-free video","Joint deviation maps without markers: a telerehab pipeline","Grading and predicting motion for autonomous rehab feedback","LSTM grades exercises, graph predicts joints, video only","Two-module rehab AI: scores moves, localizes joint errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire feedback layer rests on the assumption that the per-joint difference between a predicted pose and the pose the user actually makes corresponds to a movement fault a therapist would flag; the paper states this signal has not been validated against therapist-annotated deviations.","fun_headline_variants_meta":{"raw":{"variants":["Squat classification at 96% accuracy from marker-free video","Joint deviation maps without markers: a telerehab pipeline","Grading and predicting motion for autonomous rehab feedback","LSTM grades exercises, graph predicts joints, video only","Two-module rehab AI: scores moves, localizes joint errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000915,"raw_usage":{"total_tokens":3930,"prompt_tokens":945,"completion_tokens":2985,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":2902}},"tokens_in":561,"tokens_out":2985,"duration_ms":20048,"temperature":1.0,"reasoning_tokens":2902,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:14:18.990914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a set of exercise repetitions with both therapist fault labels and the system's predicted-vs-observed per-joint errors: if the red-threshold joints do not match the joints therapists flag, or if clearly faulty movements produce no elevation in JPE, the central feedback claim is refuted. A simpler check is to measure whether JPE stays low for correct repetitions across the full prediction horizon; if it rises sharply even for correct motion, the overlay would mislead users.","supporting_citations":[{"cited_title":"Human Motion Analysis with Deep Metric Learning","cited_arxiv_id":"1807.11176","evidence_quote":"Supplies the MMD-NCA metric-learning loss and the self-attentive BiLSTM architecture that the classification module adopts, along with the baseline FPR comparison."},{"cited_title":"Space-Time-Separable Graph Convolutional Network for Pose Forecasting","cited_arxiv_id":"2110.04573","evidence_quote":"Provides the STS-GCN backbone used by the motion predictor and the STS-GCN MPJPE values compared in the benchmark table."},{"cited_title":"Diverse human motion prediction guided by multi-level spatial-temporal anchors,","cited_arxiv_id":null,"evidence_quote":"Provides the STARS anchor-based generative predictor adopted for joint-level prediction and the MPJPE benchmark numbers that support the central prediction claim."},{"cited_title":"Learning trajectory dependencies for human motion prediction,","cited_arxiv_id":null,"evidence_quote":"The LTD-50-25 DCT-domain graph model is a baseline in the motion-prediction comparison, representing a strong prior trajectory-dependency method."},{"cited_title":"Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments,","cited_arxiv_id":null,"evidence_quote":"Human3.6M is the dataset on which the motion-prediction module is benchmarked, supplying the standardized evaluation protocol."},{"cited_title":"Prozis challenge,","cited_arxiv_id":null,"evidence_quote":"The PROZIS Challenge dataset provides the annotated squat sequences used for the exercise-quality classification evaluation."},{"cited_title":"Transformers for workout video segmentation,","cited_arxiv_id":null,"evidence_quote":"Related PROZIS workout video processing work supplies context for the dataset's origin and annotation scheme."}],"review_version":1}