{"id":"b3840b6f-d638-4ee0-8c75-38f6acb7aab5","arxiv_id":"2505.04713","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Joint-based pose trackers with outlier smoothing and late fusion estimate sprint angles to within roughly 4 to 8 degrees of expert labels, still short of the roughly 2-degree accuracy needed for injury screening.","lead":"Six computer-vision body trackers were compared against expert-labeled sprint videos to see how well they measure running biomechanics. The best joint-based models, after smoothing and fusion, reach errors of about 4 to 8 degrees, useful for low-cost analysis but not yet clinical-grade.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground truth is single-annotator plus symmetry-repaired occluded joints; no inter-annotator or marker-based validation is reported, so the reported RMSEs may conflate tracker error with label error.","rationale":"The reader's weakest_assumption was that expert annotations (including symmetry-based occlusion estimates) are treated as error-free, with no inter-annotator agreement reported. I independently arrived at the same load-bearing concern. The paper's comparative ranking of trackers is likely directionally robust because the same ground truth is used across all trackers, so label noise would partially cancel in comparisons; however, the absolute RMSE values, which are the basis for the conclusion that trackers are approaching practical biomechanical accuracy, are directly affected by label reliability. The absence of inter-annotator agreement and marker-based validation is an explicit gap, and the occlusion-repair procedure makes the ground truth particularly vulnerable in exactly the frames where trackers also fail. I therefore recommend keeping the verdict CONDITIONAL: the central claim (useful but not yet high-precision) is plausible and supported by internal consistency, but the absolute accuracy numbers cannot be fully trusted until label reliability is quantified. A concrete, tractable check is a two-annotator re-labeling study on a stratified subset. No additional concerns, such as the RTMPose citation mismatch, change this verdict, since they affect provenance rather than the main quantitative argument.","tokens_in":9634,"tokens_out":1494,"duration_ms":13285,"concrete_test":"Have two or more biomechanical experts independently re-label a stratified sample of at least 200 frames spanning all five runners, both occluded and visible joint frames, and compute per-joint inter-annotator agreement (e.g., mean absolute difference and 95% limits of agreement) on joint positions and derived angles. If the inter-annotator spread is comparable to or larger than the reported RMSE improvements (e.g., 1–3 degrees), the reported RMSE values should be reinterpreted as bounded below by label noise rather than as tracker error.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative claim—that post-processed joint trackers reach RMSE values of 3.88–6.99 degrees against expert-labeled sprint kinematics—rests on the VideoRun2D manual annotations being accurate enough to serve as ground truth. Several features of that ground truth are unquantified. First, the paper states that three biomechanical experts labeled joints in Kinovea, but no inter-annotator agreement, repeated-labeling variance, or adjudication protocol is reported, so label noise is unknown. Second, for occluded joints the labels were estimated using anatomical symmetry (Section II-A.1); since the sagittal-plane camera view makes one body side systematically occluded, these symmetry-based labels are not independent measurements, and their error is likely correlated with the occlusion condition that most affects the trackers. Third, no external validation against marker-based motion capture or IMU is provided; the only comparisons to marker-based or IMU systems are qualitative references to other studies in the Discussion. If the manual labels contain bias or noise at the level of several degrees, the reported RMSE values do not measure true biomechanical accuracy. The SVR-based post-processing threshold (Section II-C.3) also implicitly optimizes toward the manual labels, so the improvement numbers are not independent of the ground-truth imperfections. This concern is load-bearing because the headline contribution is precisely the claim that markerless tracking plus post-processing is approaching clinically useful angular accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper benchmarks six visual trackers—CoTracker2, CoTracker3, PoseNet, MediaPipe, MoveNet, and RTMPose—for estimating trunk, hip, and knee joint angles during sprinting, using the VideoRun2D dataset (40 sprints, 513 strides, 5,870 frames). The authors introduce a post-processing module based on SVR smoothing, a 3-sigma outlier rule, left/right swapping, and a late-fusion strategy that averages the three best joint trackers. The main results are RMSE values against manual expert annotations, with the best fused configurations reaching 3.88–6.99 degrees RMSE for the six angles considered.","tokens_in":9928,"tokens_out":8969,"duration_ms":81929,"significance":"The paper addresses a practical problem—markerless biomechanical analysis of sprinting—and provides a side-by-side comparison that is useful because the VideoRun2D dataset is public. The proposed post-processing and fusion ideas are simple and could be of practical value. If the ground-truth labels are reliable and the hyperparameter choices generalize, the reported RMSE values would be a useful reference point for the markerless-sprinting community. However, the manuscript does not yet establish label reliability or statistical significance, so its quantitative conclusions should be read as preliminary.","major_comments":[{"comment":"The manual ground truth is not sufficiently characterized to support the absolute RMSE claims in Tables I–III. The paper states that three biomechanical experts labeled the joints, but it reports no inter-annotator agreement, repeated-labeling variance, or adjudication protocol, and occluded joints were filled using anatomical symmetry. Because the sagittal camera view makes one body side systematically occluded, the symmetry-based labels are not independent measurements, and their errors are likely correlated with the tracker errors on the same occluded joints. Please report inter-annotator or repeated-labeling variability and analyze symmetry-inferred joints separately from directly visible joints, or reframe the results as a relative comparison among trackers rather than an absolute accuracy benchmark.","section":"Section II-A.1"},{"comment":"No uncertainty quantification accompanies the reported RMSE values. The evaluation aggregates 40 sprints and 513 strides, but there are no per-sprint standard deviations, confidence intervals, or paired statistical tests; several key differences are sub-degree (e.g., Table III right trunk 4.04 vs. 4.38, left hip 6.99 vs. 7.48). The phrase 'substantial reduction' in Section III.C is therefore not supported. Please report per-sprint error distributions and appropriate paired comparisons, or explicitly downgrade the strength of the comparative claims.","section":"Section III, Tables I-III"},{"comment":"The post-processing module has free hyperparameters—SVR kernel, C, epsilon, the 3-sigma outlier threshold, and the left/right swapping rule—and the paper does not state how they were selected or whether the reported improvements in Table II were obtained on training data. Without a validation split, cross-validation, or sensitivity analysis, the improvements may not generalize to new sprints or camera setups. Please document the parameter-selection procedure and evaluate the module on held-out data.","section":"Section II-C.3"},{"comment":"The Discussion claims the best systems are 'competitive compared to more expensive marker-based solutions,' but no marker-based or IMU ground truth was collected in this study; the comparison rests on references to other works. This claim goes beyond the evidence. Please either add a direct validation against marker-based or IMU measurements, or soften the conclusion to state that the systems agree with manual expert labeling within the reported RMSE.","section":"Section IV"}],"minor_comments":[{"comment":"The numerical summary is internally inconsistent with the tables; Table I shows a minimum joint-tracker RMSE of 3.88 degrees (RTMPose right trunk), so the stated range '11.41 degrees to 4.37 degrees' is incorrect, and the 6.99-degree value in the abstract comes from late fusion rather than from post-processing alone. Please reconcile the abstract with Tables I–III.","section":"Abstract"},{"comment":"The Abstract says 'five professional runners' while Section II.A says 'five healthy amateur soccer players' and Section IV says 'amateur football players.' Please correct the participant description for consistency.","section":"Section II.A and Section IV"},{"comment":"Section III.A reports CoTracker3's mean RMSE as 24.97 degrees, but Table I lists 24.83 degrees; the mean of the six angle values is 24.97 degrees, so the table entry appears to be a typo.","section":"Section III.A, Table I"},{"comment":"Tables I–III cite RTMPose as [37], but Section II-C.2 attributes RTMPose to Jiang et al. [22]; [37] is the HRNet paper and is not the correct reference for RTMPose. Please update the citations.","section":"Tables I-III"},{"comment":"Table III's first row reads 'MediaPipe+MoveNet+RTMPose4.14', missing a space before the numerical value, and Section II-C.3's definition of e_{m,j} should state explicitly whether the norm is computed on one coordinate at a time before thresholding, to make the outlier rule reproducible.","section":"Table III and Section II-C.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits a computer-vision application venue, and the public VideoRun2D dataset is a strength. My main reservation is that the headline accuracy numbers rest on an unvalidated annotation process; I would encourage the editor to require either inter-annotator reliability statistics or a reframing of the contribution as a relative benchmark. The RTMPose citation error should also be corrected. I regard this as a major-revision rather than a reject because the comparative ranking could survive the required additions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a fair, useful comparison paper. It does not oversell itself. It evaluates six pose trackers on 40 sprints (5,870 frames) from the group's VideoRun2D dataset, proposes a simple SVR-based outlier correction plus late fusion, and reports RMSE against expert manual labels. The new bits that matter: CoTracker3 and RTMPose hadn't been benchmarked in sprint biomechanics, and the left-side occlusion problem is exactly the one practitioners hit in sagittal-plane video. The post-processing genuinely helps the knee and hip angles on the occluded side, and the paper is clear that the best numbers (3.88-6.99 degrees) still fall short of the ~1.9 degrees needed for injury-related decisions.\n\nWhat it does well: the pipeline is described in enough detail to reproduce (SVR, 3-sigma outlier rule, left/right swap then smoothed value), the tables are internally consistent, and the authors do not dress up the point trackers' poor results. They also compare with the relevant prior work on markerless gait and sprint analysis.\n\nSoft spots, in proportion. The big one is the ground truth. Three experts labeled joints in Kinovea, but there is no inter-annotator agreement, no repeated-label variance, and no adjudication protocol. On the sagittal view the far side is systematically occluded, and those occluded joints were 'estimated based on anatomical symmetry' (Sec II-A.1). So the manual labels for the occluded side are not independent measurements. The RMSEs for the left side could be measuring tracker error partly as label error. No marker-based or IMU validation is provided, only qualitative references to other studies. Given the paper's main claim is the size of the RMSE, that is a real gap, not a nit. It is not fatal: the comparative ranking of trackers is against the same labels for all systems, so the ordering is likely directionally right. But the absolute RMSEs should be read with caution, and the paper should have said so more clearly.\n\nSecond, there are no error bars, confidence intervals, or statistical tests across the 40 sprints and 513 strides. Some of the differences (e.g., late fusion gains of 0.3-1.0 degrees) may be within label noise. A per-sprint bootstrap would fix this.\n\nThird, a citation mismatch: the paper says RTMPose by Jiang et al. [22] but Table I cites [37], which is Sun et al.'s HRNet. That should be corrected; the identity of the tracker matters.\n\nOne thing worth pushing back on: the stress-test note says the SVR post-processing 'implicitly optimizes toward the manual labels.' That is not what the text describes. The SVR is fit to the tracker trajectory itself, and the 3-sigma threshold is computed from the tracker-vs-smoothed residual, not from ground truth. So the post-processing improvement is not an artifact of fitting to labels.\n\nBottom line: this is a solid, honest benchmark for a specific applied community. It deserves peer review, not desk rejection. The fixes are adding inter-annotator stats, bootstrap errors, and fixing the RTMPose citation. I'd probably send it to someone in sports biomechanics or applied pose estimation. I wouldn't cite the absolute RMSE numbers in my own work until the label reliability is quantified, but I'd cite it as a useful comparison and use the post-processing recipe.","headline":"A useful, honest benchmark of six trackers for sprint kinematics, with a real post-processing recipe; the numbers are only as good as the unquantified manual labels, which the paper should have addressed.","tokens_in":10468,"tokens_out":3216,"would_cite":false,"duration_ms":30363,"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":"Joint-based pose trackers, after SVR outlier correction and late fusion, reproduce expert-labeled sprint knee, hip, and trunk angles with RMSE between 3.88 and 6.99 degrees, making markerless video useful but not high-precision for…","keywords":["human pose estimation","sprint biomechanics","markerless motion capture","joint angle estimation","outlier correction","support vector regression","late fusion","VideoRun2D"],"falsifier":"Re-annotate a subset of the same frames with a second independent expert team and measure the inter-annotator difference on occluded-side joints; if that difference is comparable to the reported RMSE values or to the fusion improvement of about $1^\\circ$, the claimed accuracy is not separable from label noise. Alternatively, record the same sprints with a marker-based 3D motion-capture system and recompute all RMSE values against that reference.","tokens_in":9471,"feed_emoji":"🏃","tokens_out":11093,"duration_ms":93404,"temperature":0.7,"pith_summary":"This paper asks whether off-the-shelf visual pose trackers, without markers or wearable sensors, can provide biomechanically usable joint-angle measurements for sprinting. On 5,870 frames from 40 sprints by five runners, the authors compare six trackers against manual expert annotations for trunk, hip, and knee angles. They report that joint-based trackers (MediaPipe, MoveNet, RTMPose) achieve raw RMSE values of $4.37^\\circ$ to $11.41^\\circ$, and that their proposed SVR-based outlier correction plus late fusion brings the best values down to $3.88^\\circ$ (right trunk) and $6.99^\\circ$ (left hip). The conclusion is that markerless video analysis is a useful, non-invasive screening tool for sprint biomechanics, but it is not yet accurate enough for applications that require near $1.9^\\circ$ precision, such as injury-risk assessment.","feed_headline":"Markerless trackers estimate sprint angles within about 4–7 degrees","feed_subtitle":"Outlier correction and tracker fusion bring video-only sprint analysis close to expert-labelled kinematics.","key_machinery":"The load-bearing mechanism is a post-processing module attached to the tracker output. For each joint, the module smooths the raw x and y trajectories with support vector regression, computes the residual error curve between raw and smoothed trajectories, and flags times where the residual exceeds three times its standard deviation. Flagged outliers are first corrected by swapping left and right joint labels, the most common confusion in side-view sprints, and if that fails, replaced by the smoothed value. A late-fusion step then averages joint positions from MediaPipe, MoveNet, and RTMPose before angles are computed. This mechanism is what converts noisy tracker outputs into angle curves that visually match expert ground truth.","core_discovery":"The paper's central claim is that joint-based pose estimators, once their output is smoothed and corrected, reproduce expert-labeled sprint kinematics closely enough for practical biomechanical analysis. Using the VideoRun2D dataset of 40 sprints, the authors show that point-tracking models (CoTracker2, CoTracker3) are unsuited to this task, with mean RMSE of $73.78^\\circ$ and $24.83^\\circ$, while the best skeleton-based trackers produce angle curves similar to ground truth. Their post-processing module, which smooths joint trajectories with support vector regression, detects outliers at three standard deviations of the residual error, and swaps left-right labels or replaces the outlier with the smoothed value, reduces hip and knee RMSE, for example from $10.43^\\circ$ to $8.27^\\circ$ for the hip angles of MediaPipe and MoveNet. Late fusion of the three best trackers further improves the occluded left-side joints, lowering left hip RMSE from $8.00^\\circ$ to $6.99^\\circ$. The best single tracker, RTMPose, already achieves $3.88^\\circ$ on the right trunk and $4.27^\\circ$ on the right knee, leading the authors to conclude that markerless tracking is a valuable but not high-precision resource for sprint biomechanics.","pith_inferences":["Because the post-processing correction is rule-based, a natural test is to compare it against a learned denoiser or a Kalman smoother on the same trajectories; the paper does not report such a comparison.","The reported left-side improvements may partly reflect the ground truth's own anatomical-symmetry assumption for occluded joints; a multi-camera 3D capture could separate tracker error from label bias.","Averaging joint positions from three trackers treats all trackers as equally reliable, so weighting each tracker by per-joint confidence, such as heatmap scores, could yield further gains on occluded joints.","The $1.9^\\circ$ clinical threshold comes from one hamstring-injury study, and whether the achieved $4^\\circ$--$7^\\circ$ error matters depends on the specific clinical decision, which the paper does not test."],"forward_implications":["Joint-based pose trackers with the proposed post-processing can serve as a non-invasive, cost-effective alternative to marker-based and IMU systems for routine sprint monitoring, since their angle curves resemble those reported for marker-based systems.","Near-side joints visible to the side-view camera can already be measured with errors around $4^\\circ$--$5^\\circ$, indicating that single-view video suffices for trunk and knee angles when the side is unoccluded.","The far side of the runner remains the main error source, and late fusion of multiple trackers recovers only about $1^\\circ$ there, so occlusion handling, not tracker choice, is the limiting factor.","For applications that require about $1.9^\\circ$ accuracy, such as injury-risk assessment, the current pipeline is not sufficient and would need additional views or complementary sensing."],"supporting_citations":[{"why":"Provides the MMPose framework in which RTMPose is implemented for the experiments.","marker":"[10]"},{"why":"Supplies the VideoRun2D dataset with 40 sprints and the manual expert joint annotations used as ground truth.","marker":"[18]"},{"why":"Presents RTMPose, the best-performing single tracker in the comparison.","marker":"[22]"},{"why":"Provides CoTracker3, the improved point tracker whose poor performance shows the limits of point-tracking for this task.","marker":"[24]"},{"why":"Provides CoTracker2, the point tracker whose failure establishes the point-tracking baseline.","marker":"[25]"},{"why":"Supplies the 1.9-degree accuracy target for injury-prevention biomechanics used to judge sufficiency.","marker":"[26]"},{"why":"Provides the MediaPipe pose tracker, one of the three best joint-based trackers.","marker":"[31]"},{"why":"Presents the joint-estimation model the paper evaluates under the name PoseNet.","marker":"[33]"},{"why":"Provides the MoveNet-Thunder model, one of the three best joint-based trackers.","marker":"[39]"},{"why":"Establishes the markerless recording protocol and prior angle-comparison results that this work builds on.","marker":"[42]"}],"fun_headline_variants":["Sprint angles from video: 4-7° accuracy with tracker fusion","Markerless tracking hits 4° on sprint angles after outlier fix","Joint trackers beat point trackers for sprint biomechanics","Video-only sprint analysis: within 4-7° of expert labels","Outlier correction narrows pose-tracking error to ~4° in sprints"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument treats the manual expert annotations as error-free ground truth, even though occluded joints were filled in by anatomical symmetry and no inter-annotator agreement is measured.","fun_headline_variants_meta":{"raw":{"variants":["Sprint angles from video: 4-7° accuracy with tracker fusion","Markerless tracking hits 4° on sprint angles after outlier fix","Joint trackers beat point trackers for sprint biomechanics","Video-only sprint analysis: within 4-7° of expert labels","Outlier correction narrows pose-tracking error to ~4° in sprints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000867,"raw_usage":{"total_tokens":3809,"prompt_tokens":1049,"completion_tokens":2760,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":665,"completion_tokens_details":{"reasoning_tokens":2675}},"tokens_in":665,"tokens_out":2760,"duration_ms":19039,"temperature":1.0,"reasoning_tokens":2675,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:22:39.489168+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-annotate a subset of the same frames with a second independent expert team and measure the inter-annotator difference on occluded-side joints; if that difference is comparable to the reported RMSE values or to the fusion improvement of about $1^\\circ$, the claimed accuracy is not separable from label noise. Alternatively, record the same sprints with a marker-based 3D motion-capture system and recompute all RMSE values against that reference.","supporting_citations":[{"cited_title":"Contributors","cited_arxiv_id":null,"evidence_quote":"Provides the MMPose framework in which RTMPose is implemented for the experiments."},{"cited_title":"Garrido-Lopez, L","cited_arxiv_id":null,"evidence_quote":"Supplies the VideoRun2D dataset with 40 sprints and the manual expert joint annotations used as ground truth."},{"cited_title":"Karaev, I","cited_arxiv_id":null,"evidence_quote":"Provides CoTracker2, the point tracker whose failure establishes the point-tracking baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 1.9-degree accuracy target for injury-prevention biomechanics used to judge sufficiency."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MediaPipe pose tracker, one of the three best joint-based trackers."},{"cited_title":"Papandreou, T","cited_arxiv_id":null,"evidence_quote":"Presents the joint-estimation model the paper evaluates under the name PoseNet."},{"cited_title":"MoveNet: Ultra fast and accurate pose detection model","cited_arxiv_id":null,"evidence_quote":"Provides the MoveNet-Thunder model, one of the three best joint-based trackers."},{"cited_title":"Yang and K","cited_arxiv_id":null,"evidence_quote":"Establishes the markerless recording protocol and prior angle-comparison results that this work builds on."}],"review_version":1}