{"id":"20506596-b07d-4025-8180-ccac12c1dbff","arxiv_id":"2508.20920","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"COMETH uses multi-source convex inverse kinematics with biomechanical constraints and a Kalman filter to fuse 3D skeletons from multiple cameras, improving multi-person tracking accuracy over OpenPTrack and BeFine.","lead":"COMETH fuses 3D body poses from several cameras into one anatomically consistent skeleton using convex optimization plus a Kalman filter. It reports higher tracking and localization scores than two existing edge fusion methods on a public benchmark and in a factory lab.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Grid-searching all COMETH parameters on the evaluation split would inflate the reported gains; the paper does not document a validation split, so the SOTA comparison in Tables 2–3 is potentially unfair.","rationale":"After reviewing the manuscript, I find the central construction plausible and the code release is a real asset. The QPIK formulation and state observer are reasonable, and the empirical results are internally consistent. The Kalman update equation (10) appears to have a typo (it multiplies instead of adds the innovation), but since the code is public, this is not the main obstacle to the central claim. The weakest methodological link is the evaluation protocol. The paper does not state a validation split; it says all parameters were set through grid search and kept the same. In multi-object tracking evaluations, even a few hyperparameters (timeout, solver iterations, covariance scales) can change HOTA by several points if tuned per dataset. Without evidence that tuning was separated from the test sequences, the headline advantage over OpenPTrack and BeFine could be an artifact of fitting to the Panoptic test set. This is a standard correctness concern, not a disagreement with the community consensus. I would therefore maintain the CONDITIONAL verdict: the authors should disclose the split or provide a held-out evaluation. I only partially agree with the reader's weakest_assumption: the BioAMASS velocity limits may indeed not transfer to every population, but the reported Panoptic results already suggest the limits are not catastrophically wrong; the evaluation leakage is the more decisive issue.","tokens_in":16388,"tokens_out":7412,"duration_ms":79662,"concrete_test":"Ask the authors to specify the exact Panoptic sequences or frame ranges used for grid search. If no separate validation split exists, rerun the pipeline with a pre-registered split: hold out 20% of the sequences (or one full sequence per camera count) as validation, tune the six listed hyperparameters on that subset only, lock them, and recompute Table 2 on the remaining sequences. If COMETH's HOTA/LocA lead over OpenPTrack and BeFine shrinks below the reported margins (or reverses), the central claim is not supported. This test is feasible with the released code.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is the undisclosed relationship between hyperparameter selection and evaluation. Section 4.4 says: \"We set all COMETH parameters through grid search, keeping the same values throughout the experiments in both case studies.\" The only data described for the quantitative evaluation is the CMU Panoptic dataset (Section 4.1). No training/validation/test split is given, and the grid-searched parameters (Δ=0.07 s, 100 QPIK iterations, Λ/D identity, γ=1, KF covariance Q=0.5, R=I) are reported after the fact. If the grid search was performed on the same Panoptic sequences used to compute Table 2, then the reported 76.7% HOTA vs. 37.1% for BeFine is not a fair algorithmic comparison: the method's few free parameters have been fit to the test set, while the baselines use fixed default settings. This concern is more load-bearing than the BioAMASS velocity-limit transferability: even if limits are imperfect, the empirical claim could survive, but test-set tuning directly invalidates the headline comparisons. The released code is helpful, but it does not by itself reveal the split used during grid search.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes COMETH, a distributed multi-view 3D human pose fusion algorithm intended for real-time edge deployment. The pipeline performs temporal synchronization of per-camera 3D keypoints, Hungarian-based association, per-bone body scaling with outlier rejection, multi-source convex quadratic-programming inverse kinematics (QPIK) with kinematic and biomechanical joint limits, and a Kalman-filter state observer for temporal smoothing. The authors evaluate COMETH on the CMU Panoptic dataset using 1–5 RGB-D cameras against OpenPTrack and BeFine, reporting LocA, DetA, AssA, and HOTA, and they show qualitative results and latency measurements on an industrial laboratory setup. The central claim is that COMETH outperforms both baselines on tracking metrics, especially with more cameras, while remaining computationally feasible for real-time use.","tokens_in":16793,"tokens_out":4940,"duration_ms":55641,"significance":"If the empirical claims hold, the paper addresses a genuine engineering need: accurate multi-view human pose tracking with low communication overhead and edge-computing compatibility. The QPIK formulation with per-source slack variables, combined with a state observer and explicit biomechanical constraints, is a sensible and potentially reusable design. The release of the code and the public-dataset evaluation are commendable. However, the current validation has a load-bearing weakness: all COMETH parameters are selected by grid search, yet no train/validation/test split is documented, and the only quantitative evaluation is on the same CMU Panoptic dataset. In addition, a key equation in the state observer appears to be missing a plus sign, and Table 1 reports physically implausible velocity limits in the stated units. These issues must be resolved before the comparative claims can be accepted.","major_comments":[{"comment":"The paper states (Section 4.4): 'We set all COMETH parameters through grid search, keeping the same values throughout the experiments in both case studies.' The values found (Δ=0.07 s, 100 QPIK iterations, Λ=D=I, γ=1, Q=0.5·I, R=I) are then used to produce Tables 2 and 3 on the CMU Panoptic dataset. No validation split, cross-validation, or held-out sequence is described anywhere in Section 4.1 or 4.4. Since Panoptic is the only quantitative evaluation set, the grid search appears to have been performed on the same sequences used for the final reported numbers. This makes the comparison to OpenPTrack and BeFine, which are not tuned in the same way, potentially circular. The authors must either document a proper validation split, use nested cross-validation, or evaluate on sequences not touched by the grid search; otherwise the headline HOTA gains (e.g., 76.7% vs. 37.1% at five cameras) a","section":"§4.4 and §5.1"},{"comment":"Table 1 lists velocity limits in units of °/s with values such as Hip Flexion/Extension ˙qL=−1.6, ˙qU=1.9 and Knee Flexion/Extension −2.0 to 2.1. These values are three orders of magnitude too small for human joint angular velocities in degrees per second; typical hip and knee motions exceed 100°/s during walking. They are plausible only as rad/s (≈1.6 rad/s ≈ 92°/s). If the solver interprets the table as °/s, the velocity constraints in Eq. (5) would reject almost all ordinary human motion, contradicting the reported tracking performance. If the actual implementation uses rad/s, the table headers are wrong and must be corrected. Because these limits are a central component of the QPIK constraints and are taken from BioAMASS, the paper should also state the exact units used in the solver and justify their transferability to the Panoptic and ICE-Laboratory populations.","section":"Table 1 and §3.3.2"},{"comment":"The Kalman-filter correction update is written as Q_i^+[t] = Q_i^-[t] · K_i[t](z_i[t] - H Q_i^-[t]). In a standard Kalman filter the correct update is Q_i^+ = Q_i^- + K_i(z_i - H Q_i^-), i.e., the innovation is added, not multiplied elementwise. As printed, Eq. (10) is not a valid state estimate: the dimensions are inconsistent and the expression does not implement the correction phase described in the text. Since the state observer is one of the three core contributions, this equation must be fixed and the implementation checked against the corrected formula.","section":"Eq. (10), §3.3.3"},{"comment":"The reported metrics are single-point estimates with no error bars, no standard deviations across sequences, and no significance tests. Some differences are large and likely meaningful, but others are not: for example, at two cameras the LocA values are 82.0% (COMETH) vs. 82.9% (BeFine), a difference that could easily be within run-to-run or sequence-to-sequence variability. The claim that COMETH 'significantly outperforms' the baselines is not statistically supported by the present evidence. The authors should report results per Panoptic sequence, or at least over multiple runs / subsamples, and perform a paired test where appropriate.","section":"Tables 2–3 and §5.1"}],"minor_comments":[{"comment":"The baseline name is inconsistently spelled as 'Befine' in tables and text but 'BeFine' in the abstract/references. Please unify to 'BeFine'.","section":"Throughout"},{"comment":"Several entries contain apparent spacing artifacts, e.g., '1 .9', '0 .5'. Please reformat the table so that all numeric values are unambiguous.","section":"Table 1"},{"comment":"In the extracted text, the region around Figure 1 contains uninterpretable glyph codes (e.g., '/gid00041/gid00084'). If this reflects the actual PDF rendering, the figure or its labels are corrupted and must be regenerated.","section":"Fig. 1"},{"comment":"Typo: 'mutiple' should be 'multiple'.","section":"§3.3.2"},{"comment":"The notation κ2(·) is used without definition. Clarify that it selects the second-smallest element of the set of Euclidean distances, as described in the text.","section":"Eq. (1)"},{"comment":"TRTPose is identified only by a footnote URL. A formal citation would be more appropriate for a comparison baseline.","section":"§4.3"},{"comment":"No details of the grid search are given: no ranges, no number of configurations, no criterion for selection. Even with a proper validation split, these details are needed for reproducibility.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The tuning-on-the-test-set concern is the main gatekeeper issue. It is fixable by re-running the evaluation with a documented split or with nested cross-validation, but without that the empirical contribution is not yet convincing. The units error in Table 1 and the incorrect-looking Kalman update in Eq. (10) reinforce the need for careful revision. The method itself is interesting and the code release is a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about arXiv:2508.20920. The method is a reasonable and clearly described system: multi-camera fusion using a multi-source QPIK formulation with per-source slack variables, biomechanical joint limits, body scaling, and a Kalman smoother, all built on Nimble and designed for edge devices. It's a legitimate extension of existing building blocks rather than a new theory, but it is a sensible engineering contribution, and the code is public.\n\nThe bigger issue is the evaluation. The paper reports large gains in HOTA and AssA over OpenPTrack and BeFine on the CMU Panoptic dataset (e.g., HOTA 76.7% vs. 37.1% with five cameras). But Section 4.4 says all parameters were set by grid search and kept fixed across experiments. There is no train/validation/test split described anywhere, and the only quantitative evaluation is on Panoptic. It looks like the grid search was done on the same sequences whose results appear in Tables 2 and 3. If that's what happened, the comparison is unfair: COMETH's parameters are fit to the test set, while the baselines run with their default settings. That directly undercuts the headline claim. This is a load-bearing problem, not a cosmetic one. Without a proper split or at least an honest statement of the tuning protocol, the reported margins are not trustworthy.\n\nOther soft spots are more minor. There are no error bars or significance tests; the real-world evaluation is qualitative; and equation (10) has a typo (the update should be Q+ = Q- + something, not dot). The velocity limits taken from BioAMASS are a reasonable but untested assumption for the Panoptic actors and factory workers. None of these break the method, but they add to the sense of an under-validated empirical claim.\n\nWhat the paper does well: the writing is clear, the algorithm is plausible, and the code release will let people test it. The idea of using biomechanical constraints and multi-source QPIK is a good practical answer to the noise and association problems in multi-view tracking.\n\nWho is this for? Practitioners working on multi-person 3D tracking from multiple cameras, especially on edge hardware. They'll find the architecture useful even if the numbers need re-verification. For a serious referee, yes—this deserves review, but it should come back with a request to clarify the grid-search protocol and ideally a train/test split. If the results hold under a fair comparison, it's a solid contribution.","headline":"A clearly described multi-view fusion system with a plausible method and released code, but the grid-search protocol as reported makes the headline comparison to baselines unfair without a proper train/validation/test split.","tokens_in":17202,"tokens_out":2701,"would_cite":false,"duration_ms":27939,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"COMETH claims that modeling multi-view human pose fusion as a constrained convex inverse-kinematics problem beats two established fusion algorithms in accuracy and tracking stability on public and industrial data.","keywords":["multi-view human pose estimation","3D human tracking","inverse kinematics","convex optimization","quadratic programming","Kalman filter","biomechanical constraints","edge computing"],"falsifier":"Run the same fusion pipeline on the public benchmark but with joint-velocity limits recomputed from the evaluation subjects' own recorded motions; if the accuracy gap over the clustering baseline shrinks or disappears, the claimed gains depend on the constraint population rather than the method. A related check: count how many ground-truth benchmark poses violate the published Table 1 velocity limits when passed through the same inverse-kinematics solver.","tokens_in":16337,"feed_emoji":"🧍","tokens_out":14976,"duration_ms":127787,"temperature":0.7,"pith_summary":"The paper tries to establish that multi-camera 3D human pose tracking can be both accurate and fast enough for real-time industrial monitoring if the fusion step is built on a kinematic body model instead of generic clustering or filtering. Its central claim is that posing the fusion as a convex inverse-kinematics problem — one optimization that must satisfy every camera's keypoints under anatomical joint limits and velocity limits — produces more accurate joint positions and more stable person identities than two established fusion algorithms. On a public multi-camera benchmark the method reports the best localization, detection, association, and overall tracking scores in nearly every camera configuration, reaching 76.7 percent HOTA (the overall tracking score) with five cameras, and it keeps tracking stable with up to seven people. In a production-line laboratory it also recovers poses that a clustering-based baseline gets wrong, such as a misdetected elbow and partially occluded legs. If these results hold, edge-deployed pose tracking would no longer require trading accuracy for scalability.","feed_headline":"Beats prior fusion for edge multi-view pose tracking","feed_subtitle":"A lightweight convex optimizer with anatomical limits holds joints and identities steady across up to seven people.","key_machinery":"The central object is multi-source QPIK — quadratic-programming inverse kinematics for several simultaneous targets. At each fusion step it solves for the joint velocities of a 49-degree-of-freedom biomechanical skeleton such that the forward-kinematic joint positions match detections from every camera, with per-source slack variables that let the optimizer discount conflicting measurements. It is wrapped by two supporting mechanisms: per-bone body scaling with outlier removal, which estimates a subject's proportions before solving and discards keypoints that disagree with those proportions, and a second-order Kalman state observer that filters the 49 joint angles over time for temporal cons","core_discovery":"COMETH's central discovery is that inverse kinematics, a robotics tool, can serve as the fusion rule for multi-view human pose: rather than averaging or clustering keypoints across cameras, it treats each camera's keypoints as targets for a 49-degree-of-freedom biomechanical skeleton and solves a convex quadratic program for the joint configuration closest to all of them at once. That optimization is constrained by anatomical range-of-motion limits and by joint-velocity limits from a large motion-capture corpus, and its outputs pass through a Kalman state observer before forward kinematics reconstructs the pose. On a public benchmark it reports the highest HOTA in every camera configuration","pith_inferences":["The same multi-source QPIK update could in principle be applied to any articulated body — a robot arm, an animal, or a hand — by substituting the kinematic tree and the constraint tables; nothing in the optimization itself is human-specific.","The paper reports only aggregate gains over baselines; ablating the three modules (per-bone scaling, convex constraints, and Kalman smoothing) separately would show which one drives the improvement.","Because only labeled 3D keypoints travel to the aggregator, the architecture could be extended to privacy-sensitive deployments where raw video must stay at the camera node.","The velocity-limit calibration should be re-estimated per deployment population: fixed 5th–95th percentile bounds from one motion-capture corpus are a prior that may need updating for worker populations with restricted mobility."],"forward_implications":["Adding cameras steadily improves localization: reported LocA rises from 78.4% with one camera to 87.4% with five.","Association accuracy rises steeply with camera count, from 56.1% to 86.7%, so multi-view fusion helps identity consistency rather than hurting it.","Tracking remains accurate in crowded scenes: with seven subjects and five cameras, reported HOTA reaches 85.8%.","The fusion step is fast enough for real-time use: aggregator latency stays below the 30 Hz frame budget even as the number of incoming measurements grows.","The model-based constraints can correct false and missing keypoints, such as a misdetected elbow or occluded lower limbs, that clustering-based fusion propagates."],"supporting_citations":[{"why":"The first comparison baseline: a fusion method using assignment and unscented Kalman filters.","marker":"[7]"},{"why":"The second comparison baseline: a distributed edge-fusion method using clustering and temporal association.","marker":"[10]"},{"why":"Supplies the public multi-camera benchmark and ground-truth poses used in all quantitative comparisons.","marker":"[11]"},{"why":"Supplies the biomechanical skeleton model with 49 degrees of freedom and the motion-capture data from which joint-velocity limits are derived.","marker":"[24]"},{"why":"Provides the underlying motion-capture archive from which the biomechanical dataset used for velocity bounds is built.","marker":"[25]"},{"why":"Provides normative elbow range-of-motion values used to set one of the positional joint limits.","marker":"[31]"},{"why":"Supplies the Kalman-filter formulation used as the state observer for temporal smoothing.","marker":"[34]"},{"why":"Defines the HOTA, DetA, AssA, and LocA metrics used to compare tracking performance.","marker":"[36]"}],"fun_headline_variants":["Convex fusion sharpens multi-view pose","Edge pose fusion via convex optimization","Robotics trick tightens multi-camera pose","Convex IK fuses views for accurate pose","Fast multi-view pose with convex IK"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The joint-velocity limits are fixed from the 5th–95th percentiles of one motion-capture population, and the method assumes those same limits apply to the people in the evaluation scenes; if the deployed population moves differently, valid poses will be clipped and the reported gains will not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Convex fusion sharpens multi-view pose","Edge pose fusion via convex optimization","Robotics trick tightens multi-camera pose","Convex IK fuses views for accurate pose","Fast multi-view pose with convex IK"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000573,"raw_usage":{"total_tokens":2536,"prompt_tokens":728,"completion_tokens":1808,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":1743}},"tokens_in":472,"tokens_out":1808,"duration_ms":13032,"temperature":1.0,"reasoning_tokens":1743,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:41:53.742094+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same fusion pipeline on the public benchmark but with joint-velocity limits recomputed from the evaluation subjects' own recorded motions; if the accuracy gap over the clustering baseline shrinks or disappears, the claimed gains depend on the constraint population rather than the method. A related check: count how many ground-truth benchmark poses violate the published Table 1 velocity limits when passed through the same inverse-kinematics solver.","supporting_citations":[{"cited_title":"Real-time marker-less multi-person 3D pose estimation in RGB-Depth camera networks","cited_arxiv_id":"1710.06235","evidence_quote":"The first comparison baseline: a fusion method using assignment and unscented Kalman filters."},{"cited_title":"Boldo, M","cited_arxiv_id":null,"evidence_quote":"The second comparison baseline: a distributed edge-fusion method using clustering and temporal association."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the public multi-camera benchmark and ground-truth poses used in all quantitative comparisons."},{"cited_title":"Keller, K","cited_arxiv_id":null,"evidence_quote":"Supplies the biomechanical skeleton model with 49 degrees of freedom and the motion-capture data from which joint-velocity limits are derived."},{"cited_title":"Mahmood, N","cited_arxiv_id":null,"evidence_quote":"Provides the underlying motion-capture archive from which the biomechanical dataset used for velocity bounds is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides normative elbow range-of-motion values used to set one of the positional joint limits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Kalman-filter formulation used as the state observer for temporal smoothing."},{"cited_title":"HOTA: A Higher Order Metric for Evaluating Multi-Object Tracking","cited_arxiv_id":"2009.07736","evidence_quote":"Defines the HOTA, DetA, AssA, and LocA metrics used to compare tracking performance."}],"review_version":1}