{"id":"775b02ce-5546-477c-8b5b-472df15efa11","arxiv_id":"2412.11632","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A multi-scale incremental model for human motion prediction is claimed to outperform prior state-of-the-art, but its reported gains are selective and its evaluation protocol is questionable.","lead":"This paper proposes PMS, a neural framework that predicts future human poses by modeling velocity and acceleration changes between frames at multiple time scales. The authors report large error reductions on four motion-capture datasets, but the results are inconsistently reported and the headline improvement range is cherry-picked.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on MPJPE numbers that appear to be computed in normalized coordinates, since Section 3.2 normalizes inputs to [-1,1] but Section 4.2 never denormalizes before applying Eq. 22.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the MPJPE values are likely computed in normalized coordinates without denormalization, making the SOTA claim unverifiable and likely false. I agree that this is the most load-bearing issue because every quantitative claim in the abstract, introduction, and conclusion depends on the comparability of Tables 1-5. The normalization step in Eqs. 1-2 is explicit, the MPJPE definition in Eq. 22 is in terms of raw 3D positions, and no denormalization is described anywhere in Section 4. Without denormalization, the reported numbers are not in the same units as the baselines. The 3DPW and AMASS tables provide strong internal evidence: values like 0.09 mm and non-monotonic error curves are not physically plausible for raw-coordinate MPJPE. The method description is also incomplete (placeholder symbols in Eq. 8-17, unspecified coefficients in Eq. 6-7), but even a complete method description would not rescue the claim if the evaluation metric is in the wrong space. A single re-evaluation with proper denormalization would settle the issue, and until that is done the paper should not be accepted as a reliable demonstration of state-of-the-art performance. I therefore recommend keeping the reader's REJECT verdict.","tokens_in":21547,"tokens_out":2832,"duration_ms":28670,"concrete_test":"Take one trained PMS checkpoint (or retrain per the paper's protocol) and run it on a held-out Human3.6M test sequence. Compute MPJPE twice: (a) directly on the network output in the normalized [-1,1] space of Eqs. 1-2, and (b) after applying the inverse of Eqs. 1-2 to map predictions back to the original metric coordinates before applying Eq. 22. If the published Table 1 values (e.g., 3.0 at 80ms for Walking) match version (a) but not version (b), then the reported errors are in normalized units, invalidating every comparison to baselines. If a checkpoint is unavailable, a simpler unit check suffices: compute the per-joint coordinate range on 3DPW and multiply the published PMS values (0.09-0.18) by that range; if the resulting millimeter-scale errors are 50-100x larger than any plausible human motion prediction error, the numbers cannot be in raw millimeters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PMS improves prediction accuracy by 16.3%-64.2% over previous methods is supported entirely by the MPJPE comparisons in Tables 1-5. The load-bearing assumption is that those MPJPE values are in the same physical units (millimeters) as the baselines. Section 3.2 normalizes each action's joint coordinates to [-1,1] via Eqs. 1-2, and the model is trained on these normalized coordinates. Section 4.2 defines MPJPE in Eq. 22 as the Euclidean distance between ground truth and predicted 3D positions, but the paper never states that predictions are denormalized before this computation. If the predictions remain in the normalized [-1,1] space, then all reported errors are dimensionless fractions of the per-action coordinate range and cannot be compared to baseline errors computed on raw coordinates. The 3DPW results highlight this: PMS reports 0.09-0.18, while the next-best method reports 1.0-1.66; in raw millimeters, human motion prediction errors of 0.09 mm are physically implausible, whereas these values are natural in a [-1,1] normalized space. The AMASS table also shows non-monotonic errors (8.01 at 320ms, then 2.26 at 400ms), which is impossible for a well-behaved autoregressive predictor and indicates that the reported numbers are not reliable as absolute errors. Thus, the empirical foundation of the SOTA claim is internally inconsistent: either the metric is computed in the wrong space, or the evaluation pipeline has another error that makes the tables unusable for comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Parallel Multi-scale Incremental Prediction (PMS), a framework for 3D human motion prediction that computes velocity and acceleration increments over multiple time intervals (10, 5, and 2 frames), fuses them through parallel LSTM branches, and trains with a multi-term full-timeline loss. The authors report experiments on Human3.6M, CMU Mocap, 3DPW, and AMASS-BMLrub, claiming state-of-the-art performance with 16.3%–64.2% accuracy improvements over prior methods, and they include an ablation study over network components and loss variants.","tokens_in":21915,"tokens_out":7183,"duration_ms":63937,"significance":"If the reported results were valid, the multi-scale incremental modeling idea would be a useful addition to human motion prediction, and the paper's attempt to model both velocity and acceleration increments across temporal scales is a reasonable inductive bias to investigate. The authors also provide a clear architectural description, evaluate on four datasets, and include a component-wise ablation in Table 6. However, the empirical foundation of the central claim is not sound as presented: the evaluation appears to compare normalized-coordinate errors against raw-coordinate baselines, several tables contain implausible or internally inconsistent numbers, and the headline improvement range is drawn from selected favorable action/horizon pairs. The paper does not supply code, data, or enough hyperparameter detail (e.g., the weighting coefficients in Eqs. 6–7 and attenuation coefficients in Eq. 17) to reproduce the experiments independently.","major_comments":[{"comment":"The paper normalizes each action's joint coordinates to [-1,1] using Eqs. (1)-(2) and trains on those normalized coordinates, but nowhere states that predictions are denormalized before applying the MPJPE formula in Eq. (22). As written, the reported errors are dimensionless fractions of the per-action coordinate range, not errors in millimeters, and they cannot be compared to baseline errors computed on raw coordinates. Table 4's 3DPW values (0.09 at 80 ms, 0.18 at 1000 ms) are physically implausible in millimeters but natural in a [-1,1] space. Because the entire state-of-the-art claim rests on these comparisons, this unit mismatch must be resolved; if the predictions are not denormalized, all tables must be recomputed.","section":"§3.2, Eqs. (1)–(2) and §4.2, Eq. (22)"},{"comment":"The claimed improvement range of 16.3%–64.2% is cherry-picked from selected favorable action/horizon pairs (notably Sitting Down, Posing, and short-term Sitting Down/Taking Photo) rather than from the full set of results. Table 2 shows that PMS is substantially worse than Rele-GCN on Walking at 560 ms (62.5 vs. 31.9) and at 1000 ms (158.4 vs. 37.6), and worse than Rele-GCN on Eating at 560 ms (33.0 vs. 27.4). The abstract's global statement is therefore not supported by the complete table; the authors should either report aggregate results across all actions and horizons or explicitly restrict the claim to the actions for which it holds.","section":"Abstract and Table 2"},{"comment":"The AMASS-BMLrub results are internally inconsistent. PMS is reported as 8.01 at 320 ms, then 2.26 at 400 ms and 2.31 at 560 ms, a non-monotonic error pattern that is implausible for cumulative forecast error on the same test set. The accompanying text states that PMS reduces errors by 79.2%, 63.5%, 55.5%, 89.7%, 92.0%, 6.7%, and 15.3% across seven indicators, but the table implies reductions of 76.7% at 1000 ms and 78.4% on average, not 6.7% and 15.3%. These discrepancies indicate that the reported numbers are not reliable as absolute errors and that the evaluation pipeline needs to be corrected and re-run.","section":"Table 5 and §4.3 (AMASS-BMLrub results)"},{"comment":"The narrative that PMS 'outperforms other methods for 6 actions' in the long-term Human3.6M comparison is contradicted by the table: at 560 ms, PMS is worse than the best baseline on Walking, Eating, and Smoking, and is better than the best baseline on 12 of the 15 listed actions. The statement that PMS achieves better performance on '20 out of 32 metrics' is not defined in a verifiable way (it is unclear whether other methods are compared per action and horizon separately, or per action averaged over horizons). Please report exact counts and specify the comparison protocol.","section":"§4.3, Table 2 discussion"}],"minor_comments":[{"comment":"The notation in Eq. (22) uses p and p' for ground-truth and predicted positions, but the problem definition in §3.1 uses p_t and q_t; the two notations should be unified for clarity.","section":"§4.2, Eq. (22)"},{"comment":"References [45] and [49] are the same paper (Mao et al., 'Learning trajectory dependencies for human motion prediction', ICCV 2019), yet they are listed as two distinct references. This should be corrected.","section":"References"},{"comment":"The Rele-GCN row is missing for the WalkingTogether action in both Table 1 and Table 2, although the text repeatedly compares PMS against Rele-GCN. Please either provide the missing values or explicitly state that Rele-GCN was not evaluated on that action.","section":"Tables 1 and 2"},{"comment":"The description of the 3DPW dataset says each sample contains 72 positions, whereas Human3.6M uses 28 joints with 3 coordinates each; the correspondence between the 72 values and the joint definitions is not explained, and the per-action normalization described in Eqs. (1)-(2) is not specified for 3DPW.","section":"§4.1, 3DPW setup"},{"comment":"The multi-scale weighting coefficients alpha_k, beta_k and the attenuation coefficients gamma_i are introduced with summation constraints but no values, initialization scheme, or learning procedure. The ablation entry labeled '0.4 0.6' is the only concrete numerical hint, and it is not tied directly to the equations, so the reported method is not fully reproducible from the text.","section":"§3.2, Eqs. (6)–(7) and (17)"}],"recommendation":"reject","confidential_remarks":"The central quantitative claim is not supported by the manuscript as written because of the likely normalized-coordinate error computation and the internal inconsistencies in Tables 4 and 5. I would not recommend inviting a revision unless the authors re-run the entire evaluation with a clearly stated denormalization step and report full results rather than selected action/horizon pairs. The duplicate reference [45]/[49] and unreported hyperparameters should also be checked in any resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Juncheng Zou's paper has a sensible core: model velocity and acceleration increments at multiple time intervals, fuse them in parallel branches, and train with a full-timeline loss. That is a legitimate inductive bias for human motion prediction, and the ablation study in Table 6 suggests the pieces do something—removing the T=10 interval or the acceleration correction degrades results. The related work survey is also adequate.\n\nThe problem is that the empirical foundation collapses on inspection. Section 3.2 normalizes all joint coordinates to [-1,1] per action. Section 4.2 defines MPJPE in Eq. 22 as the Euclidean distance between predicted and ground-truth 3D positions, but never states that predictions are denormalized before computing the error. The 3DPW numbers make this plain: 0.09–0.18 mm is physically implausible for raw human motion, but natural for a normalized space. The AMASS table is even more telling: 8.01 at 320 ms then 2.26 at 400 ms is non-monotonic in a way an autoregressive predictor should not be. So the tables are not comparable to the baselines, and the abstract's 16.3–64.2% improvement is not supported.\n\nThe cherry-picking makes it worse. The headline percentage gains come from selected action/horizon cells (Sitting Down, Posing). Other cells in the same tables show PMS far worse than baselines—e.g., Walking at 560 ms is 62.5 vs. GA-MIN's 35.5, and at 1000 ms it is 158.4 vs. 42.8. The text acknowledges this only indirectly.\n\nThere are also reproducibility gaps. The weighting coefficients alpha, beta, and the attenuation coefficients gamma are never given values; the paper says they are tuned on the target datasets. That is a free-parameter concern, and with the placeholder symbols and missing equation terms, a reader cannot reimplement the method from the text.\n\nWho is this for? Someone building an incremental multi-scale motion predictor might borrow the architecture idea, but nobody should rely on the reported accuracy numbers. I would not cite this in its current form. A serious referee could help the authors fix the evaluation—add an explicit denormalization, report all action/horizon cells, and provide code or full coefficient values—so I would not desk-reject it outright, but it needs major revision before the empirical claims are credible.","headline":"The multi-scale incremental idea is sensible, but the reported MPJPE numbers appear to be computed in normalized coordinates, so the SOTA claims do not hold as written.","tokens_in":22431,"tokens_out":3405,"would_cite":false,"duration_ms":31491,"reading_group":"maybe","serious_thinker":"no","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 parallel multi-scale framework that predicts velocity and acceleration increments between frames improves human motion forecasting accuracy by 16.3%-64.2% over previous methods, with gains on four benchmark…","keywords":["human motion prediction","incremental modeling","multi-scale architecture","multi-stage training","human-robot collaboration","velocity increments","acceleration increments","MPJPE"],"falsifier":"Re-run the evaluation by inverting the per-action normalization (Eqs. 1-2) on the predicted poses before computing MPJPE, then compare the denormalized errors with the same baselines on Human3.6M and 3DPW; if the errors jump from the reported 0.1-1 range to baseline-scale values, the state-of-the-art claim would not survive.","tokens_in":21329,"feed_emoji":"🤖","tokens_out":8502,"duration_ms":70346,"temperature":0.7,"pith_summary":"The paper sets out to improve long-term human motion prediction for human-robot collaboration. Its central claim is that explicitly modeling the velocity and acceleration differences between consecutive frames, at several time scales at once, gives a model better short- and long-term forecasts than predicting absolute joint positions. The proposed PMS framework computes these increments in parallel branches, refines them iteratively, and trains with a loss that spans past, present, and future frames. The authors report state-of-the-art results on Human3.6M, CMU Mocap, 3DPW, and AMASS-BMLrub, with prediction accuracy improved by 16.3%-64.2% over previous methods. A sympathetic reader would take the core claim to be that incremental multi-scale dynamics is a more learnable target for motion forecasting than raw poses.","feed_headline":"Human motion forecast error drops up to 64%","feed_subtitle":"Predicting velocity and acceleration changes across time scales sharpens long-term robot-human coordination forecasts","key_machinery":"The load-bearing object is the multi-scale incremental computation module: it converts normalized joint-position history into velocity and acceleration differences at intervals of 10, 5, and 2 frames, then synthesizes them with weighted sums (Eqs. 6-7). Parallel LSTM branches learn future increments from these features, and the acceleration branch corrects the velocity branch (Eq. 16) before attenuation regulation produces the pose (Eq. 17). An iterative adjustment mechanism and a full-timeline loss $L_t = L_p + L_c + L_f$ tie the multi-scale increments to both short- and long-range temporal context. This machinery is what carries the argument: it replaces absolute-pose regression with a hierarchy of motion increments.","core_discovery":"The paper's central claim is that PMS, a Parallel Multi-scale Incremental Prediction framework, can forecast human poses more accurately by predicting how joint positions change rather than predicting absolute positions. Given 50 observed frames, it computes velocity differences $\\Delta v_k$ and acceleration differences $\\Delta\\Delta v_k$ over three time intervals ($\\delta = 10$, $5$, and $2$ frames), combines them with learned weights, and runs parallel LSTM branches that predict future increments. The predicted acceleration is used to correct the predicted velocity before attenuation-regulated integration reconstructs future poses. A full-timeline loss sums L1 errors over past, present, and future horizons. On Human3.6M, CMU Mocap, 3DPW, and AMASS-BMLrub, the authors report that PMS outperforms the compared baselines and raises prediction accuracy by 16.3%-64.2%.","pith_inferences":["Beyond the paper, the same incremental velocity/acceleration representation could be applied to non-autoregressive transformers or diffusion-based motion generators, which currently predict absolute poses or residual positions rather than multi-scale increments.","Beyond the paper, the multi-scale increments could be extended to predict uncertainty: instead of point increments, a model could output a distribution over velocity/acceleration changes and sample consistent motion sequences.","Beyond the paper, the full-timeline loss recipe (past/current/future with multiple strides) is a transferable training strategy for any sequential predictor, not only human-motion models.","Beyond the paper, replacing the hand-set interval weights with an attention mechanism over scales might let the model adaptively emphasize short-range increments for fast actions and long-range increments for sustained motions."],"forward_implications":["If the central claim is right, modeling inter-frame velocity and acceleration increments is a better inductive bias for human motion forecasting than predicting absolute joint positions, and it reduces error accumulation over longer horizons.","The multi-scale branches let the model capture both quick joint-level changes and whole-body trajectory shifts, which is exactly what a robot sharing workspace with a person needs for collision avoidance.","The full-timeline loss, by supervising past, present, and future frames jointly, should make predictions more continuous and less prone to the blurry mean-pose problem of recurrent predictors.","Because the gains are reported on four datasets with different capture protocols, the incremental representation should transfer across mocap systems and action vocabularies."],"supporting_citations":[{"why":"Defines the motion-increment paradigm that the paper extends to multiple scales.","marker":"[1]"},{"why":"Non-autoregressive parallel decoding baseline whose parallel prediction idea PMS contrasts with.","marker":"[4]"},{"why":"The Motionmixer MLP baseline that PMS is compared against on Human3.6M and AMASS-BMLrub.","marker":"[37]"},{"why":"Rele-GCN baseline used in long-term Human3.6M comparisons and for action-variance analysis.","marker":"[42]"},{"why":"Multi-level motion attention baseline on 3DPW that PMS reports large MPJPE reductions over.","marker":"[47]"},{"why":"Human3.6M dataset and its train/test protocol.","marker":"[48]"},{"why":"3DPW in-the-wild dataset used to test generalization.","marker":"[52]"},{"why":"AMASS library, source of the ACCAD training data and BMLrub test set.","marker":"[53]"}],"fun_headline_variants":["Multi-scale motion model boosts robot-human predictions by 64%","Predicting motion changes, not positions, cuts forecast error","Incremental motion forecasting sharpens human-robot coordination","Velocity and acceleration drive smarter human motion forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on comparing errors computed after scaling each action's coordinates to [-1,1] against baseline errors, without an explicit step that scales the predictions back to original units before computing MPJPE.","fun_headline_variants_meta":{"raw":{"variants":["Multi-scale motion model boosts robot-human predictions by 64%","Predicting motion changes, not positions, cuts forecast error","Incremental motion forecasting sharpens human-robot coordination","Velocity and acceleration drive smarter human motion forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000474,"raw_usage":{"total_tokens":2314,"prompt_tokens":866,"completion_tokens":1448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":1383}},"tokens_in":482,"tokens_out":1448,"duration_ms":8908,"temperature":1.0,"reasoning_tokens":1383,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:44:19.858260+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the evaluation by inverting the per-action normalization (Eqs. 1-2) on the predicted poses before computing MPJPE, then compare the denormalized errors with the same baselines on Human3.6M and 3DPW; if the errors jump from the reported 0.1-1 range to baseline-scale values, the state-of-the-art claim would not survive.","supporting_citations":[{"cited_title":"Self-Supervised Pretraining Based on Noise-Free Motion Reconstruction and Semantic-Aware Contrastive Learning for Human Motion Prediction[J]","cited_arxiv_id":null,"evidence_quote":"Rele-GCN baseline used in long-term Human3.6M comparisons and for action-variance analysis."},{"cited_title":"Spatio-Temporal Branching for Motion Prediction using Motion Increments","cited_arxiv_id":"2308.01097","evidence_quote":"Defines the motion-increment paradigm that the paper extends to multiple scales."},{"cited_title":"Pose transformers (potr): Human motion prediction with non-autoregressive transformers[C]//Proceedings of the IEEE/CVF InternationalConferenceonComputerVision.2021:2276-2284","cited_arxiv_id":null,"evidence_quote":"Non-autoregressive parallel decoding baseline whose parallel prediction idea PMS contrasts with."},{"cited_title":"Multi-level motion attention for human motion prediction[J].Internationaljournalofcomputervision,2021,129(9):2513-2535","cited_arxiv_id":null,"evidence_quote":"Multi-level motion attention baseline on 3DPW that PMS reports large MPJPE reductions over."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Human3.6M dataset and its train/test protocol."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"3DPW in-the-wild dataset used to test generalization."},{"cited_title":"AMASS: Archive of motion capture as surface shapes[C]//ProceedingsoftheIEEE/CVF internationalconferenceoncomputervision.2019: 25 / 25 5442-5451","cited_arxiv_id":null,"evidence_quote":"AMASS library, source of the ACCAD training data and BMLrub test set."}],"review_version":1}