{"id":"e56e4a7a-b8f0-407b-bc0f-a3a876c93a39","arxiv_id":"2411.12676","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"IE-PONet is a proposed C3D plus OpenPose plus Bayesian optimization pipeline claiming minor benchmark gains, with no reproducible evidence.","lead":"This paper combines three existing tools (C3D, OpenPose, and Bayesian optimization) into a pipeline called IE-PONet for athlete pose analysis, and claims small accuracy gains on two public action datasets. It provides no code or clear method details, so the reported improvements cannot be checked.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.6's mAP/AR values contradict Table 3, and Section 3.3 never defines a 2D-to-3D lifting step; the claimed superiority cannot be verified from the paper.","rationale":"I read the paper in good faith as claiming an IoT-enabled C3D plus OpenPose plus Bayesian optimization pipeline that beats listed baselines on NTURGB+D and FineGYM. For that claim to hold, the reported metrics must be internally consistent and the pipeline must actually produce 3D poses. The Section 4.6 text and Table 3 give incompatible mAP and AR values; adopting either version changes whether the model beats HRNet-32, so the central result is not a stable fact. The reader flagged the undefined 2D-to-3D fusion as the weakest assumption, and I agree that is a serious gap, but the more directly load-bearing problem is the contradictory reported evidence: even a perfectly described model cannot be evaluated when the paper contains two mutually exclusive versions of the headline numbers. The proposed audit is the minimal check that would settle both issues: rerun the evaluation from code and compare per-class AP/AR against both reported sets, while also confirming whether any 2D-to-3D lifting operation exists. Until such an audit is possible, the central claim remains unverifiable, so the REJECT verdict is appropriate and no adjustment is needed.","tokens_in":21683,"tokens_out":7903,"duration_ms":78516,"concrete_test":"Ask the authors for the trained model, code, and raw per-class AP/AR tables, then re-run the NTURGB+D and FineGYM evaluations with the stated protocol. Compute mAP as the mean of per-class AP and compare with Section 4.6 and Table 3; if the recomputed values are 73.0/72.8, Table 3 is wrong, if 74.3/74.0, the text is wrong, and if neither, the experiments are not reproducible. In the same code audit, locate the OpenPose-to-3D step: if the only output is 2D keypoint grouping from Eq. (17) with no depth regression or multi-view triangulation, the 3D pose output is not implemented.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The condition needed for the central claim is that Table 3 reports the actual performance of a well-defined 3D pose-estimation pipeline. That condition is not met. Section 4.6 text says IE-PONet's NTURGB+D mAP is 73.0 with AR 78.5, and FineGYM mAP is 72.8 with AR 78.4, but Table 3 (and the abstract) give mAP 74.3/74.0 and AR 79.3/79.1. The difference is not cosmetic: versus the strongest baseline HRNet-32 (mAP 73.4/73.0, AR 78.9/78.7), the table version beats it by about 1 mAP on both datasets, while the Section 4.6 text version is worse on both. So the headline superiority is an artifact of which internal source is believed. Independently, the described method does not generate 3D pose: Section 3.3 gives only 2D heatmaps (Eq. 16) and keypoint grouping (Eq. 17); no triangulation, depth regression, or explicit 2D-to-3D lifting is specified, and the 'feature fusion' that supposedly converts keypoints to 3D is never defined in Eq. (15) or elsewhere. Without code or weights, neither the contradictory numbers nor the missing lifting step can be checked.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes IE-PONet, an IoT-oriented pipeline that combines C3D spatiotemporal features, OpenPose keypoint detection, and Bayesian hyperparameter optimization for 3D pose estimation and action analysis of athletes. Experiments on NTURGB+D and FineGYM are reported, with AP^p50 values around 90.5-91.0 and mAP values around 74.0-74.3 claimed to exceed eight baselines, supported by an ablation table. The central claim is an empirical benchmark assertion, but the manuscript contains an internal inconsistency between the results text and Table 3, and the described architecture does not specify how 2D keypoints are lifted to 3D poses.","tokens_in":22010,"tokens_out":4576,"duration_ms":44188,"significance":"If the reported results were reproducible, the paper would offer a modest engineering contribution: a fusion of off-the-shelf C3D and OpenPose with Bayesian tuning that achieves competitive pose-estimation accuracy at low computational cost for sports analytics. The paper's strengths are the use of standard datasets, explicit metric definitions, and an ablation structure. However, no code or model weights are provided, the evaluation protocol is underspecified, and the central empirical claim is contradicted by internal numbers, so the significance cannot currently be assessed from the manuscript.","major_comments":[{"comment":"The reported IE-PONet results are internally inconsistent. The text in Section 4.6 states NTURGB+D mAP=73.0 and AR=78.5, and FineGYM mAP=72.8 and AR=78.4, while Table 3 reports mAP=74.3/74.0 and AR=79.3/79.1 for the same model. The FineGYM AP^p50 and AP^p75 values also differ (90.7/81.0 in the text versus 91.0/81.5 in the table). This is not a cosmetic discrepancy: relative to the strongest baseline, HRNet-32 (mAP 73.4/73.0), the table version of IE-PONet is about 1 mAP better, whereas the text version is worse. The headline superiority claim is therefore not reproducible from the manuscript's own evidence.","section":"Section 4.6 / Table 3"},{"comment":"The claimed 3D pose output is never defined. Equations (10)-(17) describe preprocessing, 2D convolution and pooling, feature fusion of multi-layer feature maps, keypoint heatmaps, and keypoint grouping, but there is no triangulation, depth regression, or any other 2D-to-3D lifting operation. Section 3.5 states that 'feature fusion techniques' convert 2D keypoint data into 3D pose information, but no equation or architectural detail specifies this conversion. Because the paper's central claim is 3D pose estimation, this undefined fusion step is load-bearing and cannot be validated or reproduced.","section":"Section 3.3"},{"comment":"The evaluation protocol is underspecified. The paper does not state the train/validation/test split, whether the NTURGB+D cross-subject or cross-view protocol is used, how FineGYM clips are selected for pose estimation, or whether the baselines in Table 3 are retrained under identical conditions. No error bars, standard deviations, or repeated-run statistics are provided, and no code or trained models are released. Consequently, the numerical comparisons in Table 3 cannot be independently checked.","section":"Section 4.3 / 4.6"},{"comment":"The Bayesian optimization module is described only through generic equations, without the concrete settings needed to support the claimed contribution. No search space, number of evaluations, acquisition function details, or optimized hyperparameter values are reported, and Eq. (24) contains an unspecified trade-off parameter lambda. The ablation in Table 4 attributes a performance gain to Bayesian optimization, but the mechanism cannot be verified without these details.","section":"Section 3.4"}],"minor_comments":[{"comment":"The text reports training 'accuracy' reaching approximately 0.95 on NTURGB+D and 0.90 on FineGYM, but accuracy is not defined for pose estimation, and its relationship to the AP/mAP metrics used elsewhere is unclear.","section":"Section 4.5"},{"comment":"The prose introducing Table 3 mentions input size as a reported metric, but the table has no input size column; the table should either include this information or the text should be corrected.","section":"Section 4.6 / Table 3"},{"comment":"The text refers to 'Figure 11' when describing the overall structure, but only Figure 1 is defined; this cross-reference should be fixed.","section":"Section 3.1"},{"comment":"The manuscript contains numerous grammatical errors and typos (e.g., 'there face challenges' in Section 2.1, '3 d' in the introduction), and the reference list includes many citations that appear unrelated to the topics under discussion; a careful language edit and reference relevance check are needed.","section":"Throughout"},{"comment":"The IoT component is described only conceptually, with no experiments involving actual IoT sensors, data transmission, or edge-computing devices; the title and framing overstate what is validated in the paper.","section":"Section 3.5"}],"recommendation":"reject","confidential_remarks":"My recommendation rests on the load-bearing internal inconsistency between Section 4.6 and Table 3 and on the undefined 2D-to-3D lifting step; these are not merely presentational issues. The manuscript would need corrected numbers, a precise architectural description, and a full evaluation protocol before it could be considered for publication. If the authors can provide those, a resubmission could be worth evaluating."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a desk reject. The paper stacks C3D, OpenPose, and Bayesian optimization into a pipeline called IE-PONet and reports small gains on NTURGB+D and FineGYM, but the numbers are internally inconsistent and the core 3D lifting step is never defined. The stress-test note is correct: Section 4.6 text gives mAP 73.0/72.8 while Table 3 and the abstract give 74.3/74.0; AR differs too. Against the strongest baseline HRNet-32, the table version wins by about 1 mAP, the text version loses. The headline superiority depends on which internal source you believe.\n\nWhat is actually new: not much. The combination is a routine integration of known components, and the equations in Sections 3.2-3.4 are textbook definitions for convolution, pooling, Gaussian process updates, and expected improvement. To the paper's credit, the intended architecture is described clearly and the ablation table at least shows each module contributes something. But the 'optimal' hyperparameters from Eq. (23) are never reported, no code or weights are provided, and there is no evaluation protocol, split, or error bar.\n\nThe soft spots are load-bearing. Section 3.3 says OpenPose generates 3D pose 'through feature fusion,' yet the only equations are 2D heatmaps (Eq. 16) and keypoint grouping (Eq. 17). No triangulation, depth regression, or any defined 2D-to-3D lifting is specified. The method as described cannot produce the 3D pose that the evaluation presumably measures. The reference list also contains several unrelated citations, which looks like padding.\n\nWho is this for? A reader tracking pose-estimation benchmarks would find nothing reliable here. The paper does not deliver a reproducible result or a new capability. It is not coherent enough on its own terms to warrant a serious referee; it should be returned for a complete rewrite with actual method details, consistent numbers, and code. If the authors can provide those, the idea might support a short workshop paper. As submitted, it does not deserve reviewer time.","headline":"The paper's reported benchmark superiority is internally inconsistent and the 3D pose pipeline is never actually defined; this is a desk reject, not a revision.","tokens_in":22530,"tokens_out":1686,"would_cite":false,"duration_ms":16450,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"IE-PONet claims that combining C3D spatiotemporal features, OpenPose real-time keypoints, and Bayesian hyperparameter tuning yields high-accuracy 3D pose estimation, with $AP^{p50}$ of 90.5 on NTURGB+D and 91.0 on FineGYM at about 8 GFLOPS.","keywords":["Human Pose Estimation","IoT Sensors","Motion Capture","Multi-View Pose Estimation","Deep Learning","Athlete Action","C3D","OpenPose"],"falsifier":"Reproduce the reported $AP^{p50}$ of 90.5 on NTURGB+D using only the described C3D, OpenPose, and Bayesian optimization components with the stated settings; if the 3D pose accuracy cannot be obtained because OpenPose outputs only 2D keypoints and no lifting mechanism is provided, the central accuracy claim would not be supported by the described architecture.","tokens_in":21456,"feed_emoji":"🏃","tokens_out":9621,"duration_ms":72990,"temperature":0.7,"pith_summary":"This paper proposes IE-PONet, a three-module pipeline for 3D pose estimation and motion analysis of athletes. It joins C3D for spatiotemporal video features, OpenPose for real-time body-keypoint detection, and Bayesian optimization to tune the model's hyperparameters. The authors report that this combination reaches $AP^{p50}$ scores of 90.5 on NTURGB+D and 91.0 on FineGYM, with mAP of 74.3 and 74.0, beating eight listed baselines while keeping computational cost near 8 GFLOPS. A coach or athlete could in principle use such a system for immediate technical feedback and injury-risk assessment during training.","feed_headline":"Athlete pose network posts 91.0 AP via C3D-OpenPose fusion","feed_subtitle":"C3D motion features plus OpenPose keypoints, tuned by Bayesian search, beat eight baselines on NTURGB+D and FineGYM.","key_machinery":"The load-bearing mechanism is the IE-PONet pipeline itself, a fusion of three established components: C3D, whose 3D convolutions extract spatiotemporal volume features; OpenPose, which detects body keypoints through heatmaps and Part Affinity Fields; and Bayesian optimization, which treats accuracy as a black-box function and guides hyperparameter search with a Gaussian-process surrogate. Bilinear pooling inside C3D fuses a feature system and an attention system to capture high-order motion detail. Each module is intended to add a distinct capability: motion dynamics, keypoint geometry, and automated tuning.","core_discovery":"The paper's central claim is that merging complementary feature types—3D convolutional motion features, 2D keypoint detections, and Bayesian-tuned hyperparameters—outperforms each component alone and the existing baselines on two public action datasets. Ablation results support the claim by showing that removing C3D, OpenPose, or Bayesian optimization each lowers $AP^{p50}$ by roughly one to two points. The full model reports $AP^{p50}$ of 90.5/91.0 and mAP of 74.3/74.0 on NTURGB+D/FineGYM, ahead of the eight comparison models including HRNet-32 and skeleton-based graph networks.","pith_inferences":["The paper never specifies how OpenPose's 2D keypoint heatmaps are lifted to 3D poses or how they are fused with C3D features; a direct test would be to reimplement just that stage and measure its independent contribution.","A natural testable extension is to replace the undefined 3D lifting with an explicit depth-estimation or triangulation module and compare the reported scores.","The experiments use pre-recorded datasets, so the IoT real-time transmission and edge-computing claims remain untested in live training conditions; a field trial with streaming video would be needed to validate that part.","The gains over HRNet-32 (mAP 74.3 vs 73.4) may be largely attributable to Bayesian hyperparameter selection rather than the architecture; ablating with and without tuning at matched hyperparameters would isolate the source."],"forward_implications":["If the reported numbers hold, real-time athlete pose analysis is achievable at the stated roughly 8 GFLOPS, which is within reach of edge or mobile processors used at training venues.","The ablation gaps imply that both motion features and keypoint geometry contribute materially, so removing either loses about two $AP^{p50}$ points.","The model is claimed to generalize across two different sport-action datasets, suggesting the same pipeline could be applied to other sports without architectural change.","The comparison table positions IE-PONet ahead of both video-based models and skeleton-based graph networks, suggesting the fusion route is competitive with both families."],"supporting_citations":[{"why":"Supplies the NTURGB+D benchmark on which the reported 90.5 $AP^{p50}$ is measured.","marker":"[90]"},{"why":"Supplies the FineGYM benchmark for the reported 91.0 $AP^{p50}$.","marker":"[91]"},{"why":"I3D is the video-action baseline IE-PONet must beat on both datasets.","marker":"[94]"},{"why":"X3D is an efficient video baseline whose GFLOPS are compared against.","marker":"[95]"},{"why":"CTR-GCN is a skeleton-based baseline outperformed in the comparison table.","marker":"[97]"},{"why":"InfoGCN is another skeleton baseline IE-PONet claims to surpass.","marker":"[98]"},{"why":"HRNet-32 provides the closest baseline on mAP, setting the bar for the reported gains.","marker":"[101]"},{"why":"Source of the OpenPose keypoint-detection module used in the pipeline.","marker":"[85]"},{"why":"Provides the C3D-style spatiotemporal feature extraction idea used by the C3D module.","marker":"[83]"},{"why":"Source of the Bayesian optimization method used for hyperparameter tuning.","marker":"[76]"}],"fun_headline_variants":["C3D-OpenPose fusion posts 91.0 AP for athlete pose","Bayesian-tuned C3D+OpenPose hits 91 AP in athlete motion","91 AP athlete pose via C3D and OpenPose","Athlete 3D pose: C3D+OpenPose beat 8 baselines","IE-PONet: C3D+OpenPose fusion reaches 91 AP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline's reported accuracy rests on the assumption that OpenPose's detected 2D keypoints are actually lifted to 3D and combined with C3D's spatiotemporal features by a feature-fusion step that the paper describes only in general terms, without a concrete method or standalone validation.","fun_headline_variants_meta":{"raw":{"variants":["C3D-OpenPose fusion posts 91.0 AP for athlete pose","Bayesian-tuned C3D+OpenPose hits 91 AP in athlete motion","91 AP athlete pose via C3D and OpenPose","Athlete 3D pose: C3D+OpenPose beat 8 baselines","IE-PONet: C3D+OpenPose fusion reaches 91 AP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000916,"raw_usage":{"total_tokens":3890,"prompt_tokens":863,"completion_tokens":3027,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":2918}},"tokens_in":479,"tokens_out":3027,"duration_ms":20123,"temperature":1.0,"reasoning_tokens":2918,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:15:34.849946+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the reported $AP^{p50}$ of 90.5 on NTURGB+D using only the described C3D, OpenPose, and Bayesian optimization components with the stated settings; if the 3D pose accuracy cannot be obtained because OpenPose outputs only 2D keypoints and no lifting mechanism is provided, the central accuracy claim would not be supported by the described architecture.","supporting_citations":[{"cited_title":"Shahroudy, J","cited_arxiv_id":null,"evidence_quote":"Supplies the NTURGB+D benchmark on which the reported 90.5 $AP^{p50}$ is measured."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the FineGYM benchmark for the reported 91.0 $AP^{p50}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"I3D is the video-action baseline IE-PONet must beat on both datasets."},{"cited_title":"Feichtenhofer, X3d: Expanding architectures for efficient video recognition, in: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, 2020, pp","cited_arxiv_id":null,"evidence_quote":"X3D is an efficient video baseline whose GFLOPS are compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CTR-GCN is a skeleton-based baseline outperformed in the comparison table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"InfoGCN is another skeleton baseline IE-PONet claims to surpass."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"HRNet-32 provides the closest baseline on mAP, setting the bar for the reported gains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the OpenPose keypoint-detection module used in the pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the C3D-style spatiotemporal feature extraction idea used by the C3D module."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the Bayesian optimization method used for hyperparameter tuning."}],"review_version":1}