{"id":"02d39c2b-3815-4c5e-a712-b82c6d38e1c0","arxiv_id":"2504.16655","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Wi-Fi CSI can be converted into 2D skeleton poses by a transformer encoder-decoder, and those poses support fall detection with about 90% accuracy in a 20-person in-house dataset.","lead":"The authors train a transformer-CNN network called TED-Net to estimate 2D human skeletons from Wi-Fi channel state information, then feed those skeletons into a graph network that classifies standing, walking, squatting, and falling. They report that CSI-based skeletons give fall detection accuracy close to RGB-camera skeletons, which could enable privacy-preserving home monitoring.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fall-detection claim rests on within-subject splits: every test subject also appears in training, so the 90.5% CSI fall accuracy likely reflects subject memorization, not cross-user generalization.","rationale":"Agreement with reader is partial: the reader's weakest assumption was protocol mismatch in Table III; I agree that is a serious flaw, but I read the custom-dataset within-subject split as even more load-bearing because it directly undermines the fall-detection application that motivates the paper. The architecture itself is a reasonable composition of convolutional encoders, a transformer, and a DGNN, and the hardware validation (coaxial-cable CSI test, ROS2 synchronization) is a genuine strength. However, the empirical support for the central claim is concentrated in two evaluations, and both have protocol weaknesses. The custom dataset evaluation uses subject-internal splits for every action, so the reported accuracies likely overstate cross-user performance. The MM-Fi comparison lacks evidence that the prior PCK numbers were computed under the same subject subset, action set, train/test split, and PCK normalization. Neither issue is an internal inconsistency in the math; both are testable empirically. A leave-one-subject-out experiment and a matched-protocol reproduction of [13] would resolve them. Given that no code or custom data are released, the appropriate verdict remains conditional pending those checks.","tokens_in":10575,"tokens_out":13255,"duration_ms":126765,"concrete_test":"Run a leave-one-subject-out evaluation on the custom dataset: train TED-Net and the DGNN on 19 subjects and test on the remaining subject, reporting per-subject fall accuracy, the four-class confusion matrix, and the same metrics for RGB-skeleton input. If the CSI fall accuracy under unseen subjects falls materially below the 90.5% in Table V, or below the RGB-skeleton LOSO baseline, then the central claim that CSI skeletons alone drive reliable fall detection for new users is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central application claim—that CSI-estimated skeletons can drive reliable fall detection—is not actually tested across people. In Section III-B, standing/walking/squatting sequences are split 80/20 by duration inside each subject, and fall trials are split by repetition inside each subject (first four repetitions train, fifth tests). Every test subject is therefore also in the training set for both TED-Net and the DGNN. With only four classes in a fixed room, the model can memorize subject-specific CSI-to-skeleton mappings and pose dynamics; the reported CSI fall accuracy of 90.5% and average 94.8% (Table V, Fig. 4) do not establish that an unseen user's fall would be detected. The same within-subject temporal split is used for the MM-Fi pose comparison (10-subject subset, 80/20 by duration), while the cited PCK10 values from [13] and [16] come from their own protocols; unless those protocols are matched, the 73.8% vs 63.5% margin in Table III may be a protocol artifact rather than an architectural advantage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TED-Net, a CNN-transformer encoder-decoder that estimates 17 2D skeleton keypoints from Wi-Fi CSI tensors captured by three antennas, and a customized Directed Graph Neural Network (DGNN) that classifies the resulting skeleton sequences into stand, walk, squat, and fall. TED-Net is trained with keypoints extracted from synchronized RGB frames by YOLOv11Pose as supervision. The method is evaluated on a 10-subject subset of the public MM-Fi dataset and on a newly collected 20-subject dataset that includes fall and non-fall actions. The paper reports an average PCK10 of 73.8% on MM-Fi, compares this with published values from MetaFi++ and MetaFi, and reports CSI-skeleton action-recognition accuracy of 90.5% for fall detection versus 94.3% for RGB-skeleton input.","tokens_in":10789,"tokens_out":4324,"duration_ms":45039,"significance":"If the headline results were obtained under matched protocols and subject-disjoint evaluation, the paper would offer a practically relevant contribution to privacy-preserving fall detection: the architecture is clearly described, the integrated pose-estimation-plus-action-recognition pipeline is sensible, and the custom hardware setup is reported in enough detail to be replicated. The explicit PCK tables and the inclusion of both fall and fall-like actions (squat) are useful. However, the current evidence does not establish the central claims, because the comparison with prior work is not controlled and the fall-detection evaluation is performed on a within-subject split in which every test subject also appears in training. These issues are fixable within the paper's scope, but they require additional experiments and a more cautious interpretation of the results.","major_comments":[{"comment":"The comparison with MetaFi++ [13] and MetaFi [16] is not controlled. TED-Net is evaluated on a 10-subject subset of MM-Fi in one environment with an 80/20 temporal split described in Section III-B, while the cited PCK values are taken from the original papers, which used their own subject subsets, environments, and evaluation protocols. Because PCK depends on the subject pool, action set, and normalization, the reported margin at PCK10 (73.8% vs. 63.5%) may be a protocol artifact rather than an architectural advantage. The claim that TED-Net outperforms existing approaches therefore needs either a same-protocol re-implementation of the baselines or an explicit statement that the comparison is only indicative.","section":"Section IV-A, Table III"},{"comment":"The fall-detection evaluation uses a within-subject split: for each subject, the first four fall repetitions are used for training and the fifth repetition for testing, so every test subject is also present in the training set of both TED-Net and the DGNN. The reported CSI fall accuracy of 90.5% and average accuracy of 94.8% do not demonstrate that a fall by an unseen user would be detected. To support the claimed applicability to home fall detection, the authors should report leave-one-subject-out or otherwise subject-disjoint results, per-subject accuracy, and confidence intervals or variance over repeated runs.","section":"Section III-B, Table V, Figure 4"},{"comment":"The text states that TED-Net is 'superior across all PCK metrics' compared with [16] and outperforms [13], but the table itself shows that TED-Net is better than [13] only at PCK10 (73.8% vs. 63.5%), nearly identical at PCK20 (87.0% vs. 86.7%), and worse at PCK30, PCK40, and PCK50 (91.9/94.5/96.0 vs. 93.1/95.9/97.3). The claim as written is contradicted by the paper's own table and should be revised, or the higher-threshold behavior should be explained.","section":"Section IV-A, Table III"},{"comment":"The skeleton keypoints labeled as ground truth are produced by YOLOv11Pose from RGB images, so the PCK scores measure agreement with a learned pose estimator rather than with an independent physical measurement of joint positions. Missing or displaced keypoints are also corrected by interpolation, which can bias the evaluation. The paper should explicitly acknowledge this limitation and, ideally, validate a subset of the keypoints against manually annotated or marker-based data.","section":"Section III-A and IV-A"}],"minor_comments":[{"comment":"The sentence 'an average age of ±29 years and an average height of ±172 cm' appears to be missing the means and standard deviations; please report the actual values, e.g., mean ± std.","section":"Section III-B"},{"comment":"The definition of PCK is slightly inaccurate: the text says α is 'the minimum required proportion of correctly detected joints within a skeleton,' but Equation (1) defines PCK as the fraction of keypoints whose normalized distance is below α. Please align the wording with the equation.","section":"Section III-C1"},{"comment":"The input shape is described as (10, 3, 114) in the text but as (1, 114, 10) per receiver in Table I, and the receiver axis is not clearly indicated. Please clarify the tensor layout and any permutation or reshaping between acquisition and network input.","section":"Section III-A and Table I"},{"comment":"The statement that two example subjects demonstrate that TED-Net 'is not subject-dependent' is an overgeneralization from two individuals; please qualify this claim or provide quantitative per-subject results.","section":"Section IV-A, Figure 3"},{"comment":"The meaning of the X and O columns is explained only in the text; please define these symbols directly in the table caption for clarity.","section":"Table IV"},{"comment":"The phrase 'preliminary artificial dataset' is confusing; the dataset is a real collected dataset, so 'custom' or 'in-house' would be more accurate.","section":"Contributions, Section I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would benefit from releasing the custom dataset and code, since the current comparison with prior work cannot be independently verified without them. The central claims are plausible but the evaluation protocol is the main weakness; a subject-disjoint evaluation and a controlled baseline comparison are needed before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper builds a plausible CSI-to-skeleton estimator (TED-Net) and a DGNN action classifier, and contributes a new 20-subject fall dataset. The pose-estimation results on MM-Fi are decent, and the paper is honest that fall-time skeletons are much worse than non-fall ones (PCK10 28.5% vs 83.4%). What it does not do is demonstrate cross-user fall detection: all splits are within-subject, so every test subject is in training for both TED-Net and the DGNN. That is the load-bearing issue.\n\nThe architecture is a reasonable combination of known components: convolutional encoders, a two-layer transformer, a decoder, and a directed GNN with temporal windows. No claim of fundamental novelty, and that is fine. The dataset is the real contribution: synchronized CSI from three antennas plus RGB at 30 Hz for 20 subjects doing stand, walk, squat, and two kinds of falls, with a careful calibration check of the CSI chain. That is worth having.\n\nThe main soft spot is the evaluation protocol. For the custom dataset, standing/walking/squatting are split by duration 80/20 per subject, and fall trials are split by repetition (first four train, fifth test) per subject. So the 90.5% CSI fall accuracy and the 94.8% average are within-subject numbers. The model can memorize subject-specific CSI-to-skeleton mappings and pose dynamics; an unseen-user fall might behave far worse. The text even says 'TED-Net is not subject-dependent' based on two example subjects that are both in training. That claim does not follow.\n\nSecond soft spot: the comparison with MetaFi++ [13] and MetaFi [16] in Table III is not controlled. The authors use a 10-subject subset of MM-Fi with their own 80/20 split, while the cited numbers come from those papers' protocols. Unless the normalization and subset are matched, the 73.8 vs 63.5 PCK10 margin is not interpretable. The reader flagged this and I agree.\n\nThird, the ground truth for PCK is YOLOv11Pose output, not an independent measurement. That is a minor point for comparing methods under the same GT, but it means the absolute PCK values should be taken as 'agreement with a model,' not accuracy against a marker-based system.\n\nThe paper is clearly written, and the discussion of fall-time pose degradation is refreshingly honest. The central engineering question—whether WiFi-only skeletons can drive fall detection—remains plausible but untested across people. A serious referee would want a cross-subject split, matched baselines, and ideally code/data release. I would not desk-reject it; the dataset and pipeline justify a round of review, but the current claims are overstated.","headline":"A useful WiFi-skeleton pipeline and a new fall dataset, but the headline fall-detection accuracy is not tested across users because the split is within-subject.","tokens_in":11332,"tokens_out":2180,"would_cite":false,"duration_ms":19737,"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":"WiFi signals alone can reconstruct a human skeleton and detect falls with accuracy close to camera-based systems.","keywords":["human skeleton pose estimation","human action recognition","channel state information (CSI)","transformer network","graph neural networks","fall detection","Wi-Fi sensing","privacy-preserving monitoring"],"falsifier":"Retrain TED-Net and the prior baseline on the full public dataset with the same split and PCK normalization; if TED-Net's PCK10 no longer exceeds the baseline's, the superiority claim fails.","tokens_in":10381,"feed_emoji":"📡","tokens_out":8759,"duration_ms":75472,"temperature":0.7,"pith_summary":"The paper sets out to show that Wi-Fi Channel State Information (CSI) — the signal data already available from commodity Wi-Fi hardware — carries enough information to reconstruct a person's 2D skeleton, and that those skeletons can drive activity recognition and fall detection. It introduces TED-Net, a Transformer-based encoder-decoder that maps CSI tensors to 17 body-keypoint coordinates, and couples it with a Directed Graph Neural Network that classifies stand, walk, squat, and fall at the frame level. On the public multi-modal benchmark, TED-Net reports a PCK10 of 73.8%, above the 63.5% reported for the prior CSI pose estimator. On the authors' own fall dataset, CSI-derived skeletons give 90.5% fall accuracy, compared with 94.3% when the same graph network is fed RGB-derived skeletons. If these results hold, home Wi-Fi alone could provide privacy-preserving fall monitoring without cameras.","feed_headline":"WiFi signals can spot falls nearly as well as cameras","feed_subtitle":"A transformer network maps Wi-Fi signals to body keypoints, then a graph network detects falls at 90.5% accuracy.","key_machinery":"The load-bearing mechanism is TED-Net's CNN-Transformer encoder-decoder: three convolutional encoders per receiving antenna compress the CSI tensor, the encoded features from three antennas are concatenated and reshaped into a sequence of 34 tokens (17 keypoints × 2 coordinates), and a two-layer, eight-head Transformer applies self-attention across that sequence so globally distant joints can influence each other; a convolutional decoder then up-samples the attended features and outputs the 17 (x, y) coordinates. The second mechanism is the customized DGNN, which uses directed edges between keypoints, a dynamically adjusted adjacency matrix, and temporal convolution to classify each frame as stand, walk, squat, or fall. The Transformer's global cross-joint attention is what lets the network infer coordinated body poses from spatially coarse CSI, and the DGNN's directed graph structure is what lets it propagate joint information along anatomical connections to separate falls from similar postures such as squats.","core_discovery":"The central claim is that a Transformer-based encoder-decoder can estimate 17 two-dimensional skeleton keypoints directly from raw CSI and that those keypoints are accurate enough for downstream action recognition. The paper reports that TED-Net outperforms prior CSI-based pose estimators on the public dataset, with an average PCK10 of 73.8% versus 63.5%, and that the resulting skeletons preserve body-center joints best (pelvis 99.3%, shoulders about 85%) while hands are hardest. On the custom fall dataset, the DGNN achieves 90.5% accuracy on falls from CSI skeletons and 94.3% from RGB skeletons, and the paper argues this gap is small enough that CSI skeletons can replace camera-based skeletons for fall detection. The paper also claims that pose estimation is not subject-dependent and that the system can track pelvis/torso position with under 10% normalized error.","pith_inferences":["Beyond the paper: the training loop still requires RGB-skeleton supervision from a camera, so the system is privacy-preserving at inference time only; deployment would need a one-time camera-based calibration phase.","Beyond the paper: the reported 90.5% fall accuracy is measured on the authors' own data collected in a fixed room with a fixed transmitter-receiver geometry; moving furniture, another person, or a different room layout would likely degrade performance, so the claim is for semi-constrained settings only.","Beyond the paper: because pelvis tracking error is the lowest, a simpler detector could use only torso/pelvis trajectory to flag falls, potentially removing the need for full-keypoint graph classification.","Beyond the paper: a natural stress test is to evaluate TED-Net on the full public dataset with all subjects and across multiple environments rather than the 10-subject single-environment subset; if the gap over prior work narrows, the claimed advantage may be protocol-dependent."],"forward_implications":["A home Wi-Fi setup with one transmitter and three receivers could run continuous fall detection without cameras, preserving privacy in bathrooms and bedrooms.","CSI-derived skeletons can replace RGB-derived skeletons for frame-level action recognition in semi-constrained indoor environments at a small accuracy cost.","The DGNN's confusion matrix shows fall misclassifications mostly as squat, so adding a dedicated squat class keeps false alarms low.","Torso and pelvis tracking with under 10% normalized error enables coarse position monitoring alongside action recognition.","The architecture can be extended to other healthcare monitoring tasks that rely on large-scale body dynamics rather than fine joint detail."],"supporting_citations":[{"why":"supplies the prior transformer-based CSI pose estimation method and its published PCK values that TED-Net is compared against.","marker":"[13]"},{"why":"provides the public multi-modal benchmark dataset used for the comparative pose-estimation evaluation.","marker":"[14]"},{"why":"provides the second baseline whose PCK metrics TED-Net is claimed to outperform.","marker":"[16]"},{"why":"defines the original directed graph neural network architecture that the action-recognition module adapts.","marker":"[15]"},{"why":"supplies the RGB-based pose estimator used to generate the ground-truth keypoints that supervise TED-Net's training.","marker":"[1]"}],"fun_headline_variants":["WiFi maps bodies to spot falls as well as cameras","TED Net turns WiFi into 17 keypoints for fall detection","CSI skeletons reach 90.5% fall accuracy, rivaling RGB","Privacy-safe falls detection: WiFi skeletons rival cameras"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison to prior work assumes that the published PCK numbers were produced under the same evaluation protocol — same subject subset, same train/test split, same torso-length normalization — as the paper's 10-subject subset.","fun_headline_variants_meta":{"raw":{"variants":["WiFi maps bodies to spot falls as well as cameras","TED Net turns WiFi into 17 keypoints for fall detection","CSI skeletons reach 90.5% fall accuracy, rivaling RGB","Privacy-safe falls detection: WiFi skeletons rival cameras"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1438,"prompt_tokens":949,"completion_tokens":489,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":419}},"tokens_in":565,"tokens_out":489,"duration_ms":5032,"temperature":1.0,"reasoning_tokens":419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:57:59.884116+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain TED-Net and the prior baseline on the full public dataset with the same split and PCK normalization; if TED-Net's PCK10 no longer exceeds the baseline's, the superiority claim fails.","supporting_citations":[{"cited_title":"Metafi++: Wifi-enabled transformer-based human pose estimation for metaverse avatar simulation,","cited_arxiv_id":null,"evidence_quote":"supplies the prior transformer-based CSI pose estimation method and its published PCK values that TED-Net is compared against."},{"cited_title":"Mm-fi: Multi-modal non-intrusive 4d human dataset for versatile wireless sensing,","cited_arxiv_id":null,"evidence_quote":"provides the public multi-modal benchmark dataset used for the comparative pose-estimation evaluation."},{"cited_title":"Metafi: Device-free pose estimation via commodity wifi for metaverse avatar simulation,","cited_arxiv_id":null,"evidence_quote":"provides the second baseline whose PCK metrics TED-Net is claimed to outperform."},{"cited_title":"Skeleton-based action recog- nition with directed graph neural networks,","cited_arxiv_id":null,"evidence_quote":"defines the original directed graph neural network architecture that the action-recognition module adapts."},{"cited_title":"Ultralytics yolo11,","cited_arxiv_id":null,"evidence_quote":"supplies the RGB-based pose estimator used to generate the ground-truth keypoints that supervise TED-Net's training."}],"review_version":1}