{"id":"fec74560-d449-4150-ae84-e1866f8a22ee","arxiv_id":"2507.09672","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A velocity-integrated spatiotemporal attention network is reported to improve WiFi-based 2D and 3D pose estimation, but the evaluation has temporal leakage and reproducibility gaps.","lead":"This paper presents a WiFi CSI based deep learning system for human pose estimation, adding a velocity branch to a dual-stream spatiotemporal transformer and reporting 92.2% PCK@50 on a new self-collected dataset. The headline accuracy is weakened by clip-level random splitting, test-set-based hyperparameter selection, and unavailable data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 92.2% PCK@50 rests on a random clip-level split; temporal leakage from shared 3-second action recordings likely inflates the result, and the paper provides no subject/action-disjoint evaluation to rule this out.","rationale":"The reader identified the same load-bearing weakness: random clip-level splitting from shared action recordings creates train/test leakage and invalidates the central quantitative claim. My independent reading of Section V.A confirms that 3,300 action segments are cut into 10 clips each and then randomly split, with no statement that clips from the same action segment are kept in the same fold. This is the single most load-bearing issue because the paper's main contribution is empirical: a new state-of-the-art accuracy and the claimed benefit of velocity modeling. If the split is corrected and the numbers drop, the paper's central claims no longer hold. I do not find an equally decisive flaw in the architectural derivation; the method is incremental but not internally inconsistent. The other concerns (no error bars, test-set hyperparameter selection, abstract mismatch, unavailable code) are secondary but real, and they reinforce the difficulty of verifying the result. Since the reader's verdict of REJECT already rests on this leakage concern, my assessment does not change the verdict; I would keep REJECT (or, equivalently, mark the core empirical claim unverified pending a corrected split and code/data release). No ad hominem is intended: the critique is about the evaluation protocol, not the authors' conduct.","tokens_in":12062,"tokens_out":3253,"duration_ms":38688,"concrete_test":"Request the exact train/test clip indices from the authors (the abstract's code link should contain the partition). Recompute Table II after enforcing that all clips derived from the same 3-second action segment are placed entirely in one fold, and also report leave-one-volunteer-out PCK@50. If the overall PCK@50 falls by more than about 2 points, or if DT-Pose no longer trails by the reported margin, the temporal-leakage concern is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Section V.C's 92.2% mean PCK@50 and Table II's 'outperforms existing approaches across all evaluation metrics.' The evaluation protocol in Section V.A splits 33,200 CSI clips randomly at a 4:1 ratio after each 3-second action segment is divided into 10 clips of 9 consecutive frames. Because all clips from one action segment share the same volunteer, room, action, and largely overlapping CSI content, random clip-level splitting places near-duplicate examples in both training and testing. This allows the model to memorize recording conditions rather than learn generalizable pose mapping. The concern directly affects the headline numbers: Table II's margins over DT-Pose (88.24 vs 92.25 PCK@50) and MetaFi++ (86.37 vs 92.25) could shrink or disappear under a disjoint action-segment or leave-one-subject-out split. Table VI's velocity ablation (+2.5 PCK@20) is equally exposed because it uses the same leaked partition. No error bars, multiple-seed runs, or subject-disjoint results are reported, and the abstract's '8.3%' improvement claim does not even match Table II's absolute margins. The paper's own limitation section (V.F) acknowledges cross-domain difficulties, which makes the absence of a strict split especially consequential.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VST-Pose, a WiFi CSI-based human pose estimation method that combines a dual-stream spatiotemporal transformer backbone (ViSTA-Former) with an explicit velocity modeling branch. The authors collect a 2D home-care dataset with five volunteers performing 15 actions, report 92.25% PCK@50 on it, compare against MetaFi++, Hpeli, and DT-Pose, and also evaluate on the public MMFi dataset for 3D pose estimation. The central claims are state-of-the-art accuracy on the self-collected dataset and improved fine-grained motion modeling via velocity supervision.","tokens_in":12277,"tokens_out":6101,"duration_ms":61107,"significance":"If the reported accuracy were obtained under a valid evaluation protocol, the contribution would be useful: explicit velocity modeling for WiFi pose is a sensible idea, the dual-stream attention design builds on established prior work, the new dataset could be a community resource, and the code is publicly released. However, the evaluation protocol as described—random clip-level splitting, test-set-based hyperparameter selection, and single-run results—means the headline numbers cannot currently be interpreted as generalization performance. The MMFi results also do not show uniform superiority over all baselines on all metrics. The architectural contribution is plausible, but the evidence for its claimed advantage is not yet convincing.","major_comments":[{"comment":"The evaluation protocol in Section V.A splits 33,200 CSI clips randomly at a 4:1 ratio after each 3-second action segment is divided into 10 clips. Because clips from the same action segment share the same volunteer, room, action, and contiguous time interval, random clip-level splitting places temporally and environmentally correlated examples in both training and testing. This can inflate the headline 92.25 PCK@50 in Table II and the velocity ablation in Table VI by allowing the model to memorize recording conditions rather than learn generalizable pose mappings. The paper must report results under a subject-disjoint and/or action-segment-disjoint split (e.g., leave-one-subject-out), and should present the same baseline comparisons under that split. The limitation discussion in Section V.F makes the absence of such a strict split especially consequential.","section":"Section V.A, Table II"},{"comment":"Tables IV, V, VII, and VIII select the sliding-window size, DST-block depth, velocity branch source, and fusion strategy by comparing performance on the same test set that later produces the final reported metrics. This is test-set-based model selection, which can systematically overstate the accuracy of the chosen configuration. The authors should use a validation split for these design choices and then report final results on a genuinely held-out test set, or use nested cross-validation.","section":"Section V.E, Tables IV-VIII"},{"comment":"All reported numbers appear to come from a single training run with no error bars. Several key comparisons are small; for example, Table VII shows 92.25 PCK@50 for the TS branch versus 92.06 for TS+ST, and Table VIII shows 92.25 with fusion versus 91.68 without. Without multiple seeds and standard deviations, these differences are not statistically interpretable. Please report mean and standard deviation over at least three random seeds, and indicate statistical significance where appropriate.","section":"Section V.C, Tables I-VIII"},{"comment":"The abstract claims that VST-Pose outperforms existing methods by 8.3% in PCK@50, but Table II shows absolute margins of 4.01, 5.88, and 6.65 percentage points over DT-Pose, MetaFi++, and Hpeli, respectively, corresponding to relative improvements of roughly 4.5%, 6.8%, and 7.8%. The paper should specify the baseline and formula used for the 8.3% figure, or correct the claim.","section":"Abstract, Table II"},{"comment":"On the MMFi dataset, VST-Pose's PA-MPJPE is 105.9 mm, which is worse than Hpeli (105.3) and DT-Pose (105.6). The text in Section V.D says the method gives 'significant improvements in PCK@20 and MPJPE,' and the conclusion states the model 'consistently outperforms existing approaches,' but this is not supported on MMFi across all metrics. Please qualify the MMFi claims or provide additional evidence for robustness.","section":"Section V.D, Table III"}],"minor_comments":[{"comment":"Typo: 'estimateon' should be 'estimation'.","section":"Section II.A"},{"comment":"Typo: 'diraction' should be 'direction'.","section":"Section V.A"},{"comment":"The dataset description says 3,300 action segments, each divided into 10 clips, which gives 33,000 clips, not 33,200. Please reconcile the count.","section":"Section V.A"},{"comment":"Reference numbering is inconsistent: MetaFi++ is cited as [37] in Table II but appears as reference [35] in the reference list, where [37] is assigned to a different work (HPE-Li).","section":"Section V.D, Table II"},{"comment":"Equation (5) is garbled in the text; please rewrite the loss function and the definition of ground-truth velocity clearly.","section":"Section IV.D, Eq. (5)"},{"comment":"In the Network Depth Selection paragraph, the sentence 'Table IV is composed of stacked DST-Blocks' appears to be a caption duplication; the reference should be to the architecture description or to Table V.","section":"Section V.E"}],"recommendation":"major_revision","confidential_remarks":"The core architectural ideas are reasonable and the dataset collection is a useful effort, but the evaluation protocol needs a fundamental overhaul before the claims can be accepted. At minimum, the authors should re-run experiments with subject-disjoint or action-segment-disjoint splits, select hyperparameters on a validation set, and report multiple seeds with error bars. If those experiments reveal that the reported margins largely disappear, the paper's contribution would be substantially weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's central claim on the self-collected dataset should not be trusted as reported. The evaluation splits 33,200 short clips randomly at 4:1, but each clip comes from a 3-second action segment that was cut into 10 overlapping clips. Clips from the same segment share the same volunteer, room, action, and mostly overlapping CSI content, so the test set contains near-duplicates of training clips. That allows the model to memorize recording conditions rather than learn a generalizable pose mapping. The reported margins over DT-Pose and MetaFi++ on Table II could shrink or disappear under a subject-disjoint or action-segment-disjoint split. The paper's own limitation section acknowledges cross-domain difficulties, which makes the absence of a strict split especially consequential.\n\nWhat is actually new: a velocity modeling branch added to a dual-stream spatiotemporal attention backbone, applied to WiFi CSI. The architecture is explicitly inherited from DST-Former/MotionBERT, so the novelty is incremental, but the velocity branch is a sensible twist for CSI and the ablation suggests it helps. The self-collected home-care 2D pose dataset is a useful contribution in principle, though the code and data are not actually released despite the abstract's claim. The MMFi experiments use a public benchmark and show a modest but plausible gain (87.37 vs 86.38 PCK@50), which is less exposed to the leakage problem because it follows the official split.\n\nOther soft spots: hyperparameters (window size, depth, velocity branch source, fusion strategy) are selected by comparing on the same test set used for the final reported numbers, so the reported metrics are cherry-picked. There are no error bars or multiple-seed runs. The abstract's \"8.3%\" improvement does not match Table II, where the absolute PCK@50 margins over the next best method are 4-6 points.\n\nWho this is for: researchers working on WiFi-based pose estimation who want to see a velocity-integrated transformer variant and a home-care dataset description. The MMFi result is worth verifying, and the architecture is easy enough to reproduce once code is out. It deserves a serious referee because the core idea is legitimate and the public-benchmark result is potentially useful, but the self-collected claim needs a strict split and error bars before it can be accepted.\n\nMy recommendation: send to peer review with major revision. Ask for a leave-one-subject-out or disjoint action-segment split, multiple seeds, and a corrected abstract.","headline":"The velocity-integrated attention idea is reasonable and the MMFi results are worth a look, but the self-collected 92.2% PCK@50 claim is inflated by a clip-level random split that leaks near-duplicate temporal windows into both train and test.","tokens_in":12864,"tokens_out":1946,"would_cite":false,"duration_ms":22907,"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":"The paper claims that explicitly modeling keypoint velocity lets a spatiotemporal attention network estimate human pose from WiFi CSI at 92.2% PCK@50, outperforming prior WiFi pose estimators on a self-collected dataset.","keywords":["WiFi sensing","channel state information","human pose estimation","spatiotemporal attention","velocity modeling","transformer","PCK","smart home monitoring"],"falsifier":"Evaluate VST-Pose with a subject-exclusive split, training on four of the five volunteers and testing on the fifth, and compare the PCK@20 gap between the velocity-on and velocity-off versions. If the 2.5-point gain shrinks to near zero, or if the overall PCK@50 drops well below 92.2%, the paper's core claim that velocity modeling drives fine-grained accuracy would not survive.","tokens_in":11811,"feed_emoji":"📡","tokens_out":6216,"duration_ms":63973,"temperature":0.7,"pith_summary":"The paper tries to establish that fine-grained, video-free human pose estimation is achievable from commodity WiFi channel state information, not just coarse activity recognition. It proposes VST-Pose, which reads short sequences of CSI frames and predicts both joint positions and joint velocities through a dual-stream spatiotemporal Transformer. On a self-collected dataset of 15 home-care actions, the method reaches an overall mean PCK@50 of 92.2% and reports gains over three previous WiFi pose estimators across all evaluated metrics. If the result holds, indoor health monitoring and fall detection could run without cameras, preserving privacy while keeping coverage through walls and in low light.","feed_headline":"WiFi pose estimation hits 92.2% with velocity modeling","feed_subtitle":"Body keypoints come from channel state information, so monitoring works without cameras in homes.","key_machinery":"The central object is the DST-Block inside the ViSTA-Former backbone: it runs two parallel attention streams, one that applies spatial attention then temporal attention and another that applies temporal then spatial, and attaches a local velocity branch to the time-first stream. The velocity branch learns keypoint displacement patterns within a short sliding window, and its features are fused late with the main keypoint features via F_K = 0.5 F_V + F_N, so the decoder sees both static joint structure and short-term motion. The velocity signal is computed as the difference between the last and first frames of the keypoint sequence, making explicit motion dynamics the mechanism that distinguishes subtle limb movements that are hard to see in instantaneous CSI amplitude.","core_discovery":"VST-Pose's central claim is that a dual-stream spatiotemporal attention backbone, with an auxiliary branch that explicitly regresses keypoint velocity, can map short sequences of WiFi CSI amplitude frames into temporally coherent human skeletons. The paper reports an overall mean PCK@50 of 92.2% and PCK@20 of 71.9% on its self-collected dataset of 33,200 clips, stating that this outperforms MetaFi++, Hpeli, and DT-Pose on PCK@50, PCK@20, MPJPE, and PA-MPJPE. The velocity branch is credited with a 2.5-point PCK@20 improvement over the version without velocity modeling, which the authors interpret as evidence that short-term motion dynamics, not just instantaneous spatial structure, carry the fine-grained pose signal. A separate evaluation on the public MMFi dataset shows improvements in PCK@20 and MPJPE for 3D pose estimation, supporting the claim that the mechanism transfers beyond the authors' own hardware and annotation pipeline.","pith_inferences":["The paper's performance gap over DT-Pose, especially the 8.75-point PCK@20 difference, would need corroboration on a subject-exclusive split because clips from the same 3-second recordings and same five volunteers may overlap between training and testing, inflating all methods' numbers.","A natural testable extension is to train VST-Pose on the MMFi training set and evaluate on unseen subjects or unseen rooms; the paper reports only random splits, not cross-domain splits.","The velocity branch uses only the difference between the last and first frame of the window; replacing this with per-frame instantaneous velocity might change the reported 2.5-point fine-grained gain.","The ground-truth labels come from a video-based keypoint detector, so the system's ceiling is bounded by that detector's accuracy; evaluating against other annotation sources would clarify the true limit of WiFi-only pose estimation."],"forward_implications":["If the 92.2% PCK@50 result survives subject-independent evaluation, WiFi-based pose estimation becomes usable for continuous home-care monitoring without any camera in the room.","The reported 2.5-point PCK@20 gain from velocity modeling suggests that explicit motion features are a general lever for fine-grained keypoint accuracy in radio-frequency pose estimation.","The improvement on MMFi indicates the architecture transfers to 3D pose estimation and to different WiFi hardware than the Intel 5300 setup used for data collection.","Because a three-frame window works best, the method implies that very short-term dynamics suffice for pose regression, which could keep latency low in real-time monitoring systems."],"supporting_citations":[{"why":"Defines the multi-head self-attention mechanism used in both spatial and temporal blocks.","marker":"[27]"},{"why":"Provides the dual-stream DST-Former design that ViSTA-Former adapts for its two attention branches.","marker":"[28]"},{"why":"Supplies the 2D keypoint ground truth extracted from synchronized video frames.","marker":"[4]"},{"why":"MetaFi++ is a primary baseline and the source of the PCK evaluation metric.","marker":"[35]"},{"why":"DT-Pose is the strongest competing WiFi pose estimator on the self-collected dataset.","marker":"[24]"},{"why":"MM-Fi provides the public 3D pose dataset used to test generalization beyond the authors' data.","marker":"[29]"},{"why":"HPE-Li, listed as Hpeli, is a competing WiFi pose estimator used in the comparison tables.","marker":"[37]"}],"fun_headline_variants":["WiFi pose estimation reaches 92.2% PCK with velocity modeling","Velocity-aware spatiotemporal WiFi attention improves pose estimation","No cameras needed: WiFi CSI drives pose estimation to 92.2% accuracy","VST-Pose: WiFi signals map human skeletons with 92.2% precision","Dual-stream WiFi attention beats prior pose methods by 8.3% PCK"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that randomly splitting the 33,200 CSI clips into training and testing sets at a 4:1 ratio produces independent samples, but because clips come from the same 3-second action recordings and the same five volunteers, adjacent clips can appear in both sets and the reported accuracy may reflect memorizing recording conditions rather than generalizing to new people or actions.","fun_headline_variants_meta":{"raw":{"variants":["WiFi pose estimation reaches 92.2% PCK with velocity modeling","Velocity-aware spatiotemporal WiFi attention improves pose estimation","No cameras needed: WiFi CSI drives pose estimation to 92.2% accuracy","VST-Pose: WiFi signals map human skeletons with 92.2% precision","Dual-stream WiFi attention beats prior pose methods by 8.3% PCK"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001,"raw_usage":{"total_tokens":4247,"prompt_tokens":975,"completion_tokens":3272,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":3172}},"tokens_in":591,"tokens_out":3272,"duration_ms":23934,"temperature":1.0,"reasoning_tokens":3172,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:50:21.024960+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate VST-Pose with a subject-exclusive split, training on four of the five volunteers and testing on the fifth, and compare the PCK@20 gap between the velocity-on and velocity-off versions. If the 2.5-point gain shrinks to near zero, or if the overall PCK@50 drops well below 92.2%, the paper's core claim that velocity modeling drives fine-grained accuracy would not survive.","supporting_citations":[{"cited_title":"MotionBERT: A unified perspective on learning human motion representations,","cited_arxiv_id":null,"evidence_quote":"Provides the dual-stream DST-Former design that ViSTA-Former adapts for its two attention branches."},{"cited_title":"OpenPose: Realtime multi-person 2D pose estimation using part affinity fields,","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D keypoint ground truth extracted from synchronized video frames."},{"cited_title":"MetaFi++: WiFi-enabled transformer-based human pose estimation for metaverse avatar simulation,","cited_arxiv_id":null,"evidence_quote":"MetaFi++ is a primary baseline and the source of the PCK evaluation metric."},{"cited_title":"AdaPose: Toward cross-site device-free human pose estimation with commodity Wi-Fi,","cited_arxiv_id":null,"evidence_quote":"DT-Pose is the strongest competing WiFi pose estimator on the self-collected dataset."},{"cited_title":"MM-Fi: Multi-Modal Non-Intrusive 4D Human Dataset for Versatile Wireless Sensing","cited_arxiv_id":"2305.10345","evidence_quote":"MM-Fi provides the public 3D pose dataset used to test generalization beyond the authors' data."},{"cited_title":"HPE-Li: Wi-Fi-enabled lightweight dual selective kernel convolution for human pose estimation,","cited_arxiv_id":null,"evidence_quote":"HPE-Li, listed as Hpeli, is a competing WiFi pose estimator used in the comparison tables."}],"review_version":1}