{"id":"d1f742f1-4162-470c-aa0a-a1513a939fff","arxiv_id":"2505.22555","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"MultiFormer estimates human pose from WiFi CSI using dual time-frequency tokens, self-attention, and three-stage pose heatmap refinement, claiming gains over prior CSI pose methods.","lead":"This paper describes a WiFi-based system that estimates human joint positions from wireless channel reflections instead of camera images. It reports accuracy gains over prior methods on a new two-volunteer dataset and on the public MM-Fi benchmark, with the largest gains on wrists and elbows.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MM-Fi SOTA comparison lacks a documented shared protocol; the WPFormer parameter/FLOP discrepancy across tables suggests mixed-source numbers, so the headline accuracy gain is not yet established.","rationale":"I read MultiFormer as a system paper whose main assertion is empirical: on MM-Fi and a self-collected dataset, the proposed time-frequency dual-token transformer plus a multi-stage PCM/PAF decoder outperforms prior CSI pose estimators, especially at hands, wrists, and elbows. The architecture sections are internally coherent: TFDDT is described with equations, attention formulas, and stage-wise PCM/PAF decoders, and the ablations are directionally consistent. There is no hidden circularity in using OpenPose as a teacher: the paper discloses it, and the student is trained on its outputs, which is standard in this line of work, though it caps the validity of the self-collected ground truth. The load-bearing weakness is the SOTA comparison. The same named baseline, WPFormer, has incompatible parameter and FLOP counts in Tables VII and VIII; no shared protocol, split, or preprocessing is described; and the 17-keypoint MM-Fi annotations are never reconciled with the 18-keypoint output. If these numbers were drawn from different papers with different evaluation settings, the reported PCK differences, including the over-500% hand-joint improvements, could be artifacts of protocol rather than architecture. The one concrete check that settles this is a controlled re-run of all baselines on a single official split with common preprocessing and keypoint normalization. Because the reader's CONDITIONAL verdict already depends on exactly this testability, my read does not change the verdict; it reinforces the condition.","tokens_in":20341,"tokens_out":6972,"duration_ms":91344,"concrete_test":"Obtain the official checkpoints or code for WPFormer (MetaFi++), CSI2Pose, HPE-Li, and DT-Pose, and rerun them on MM-Fi using the official training/validation split, the same CSI preprocessing (upsampling, tokenization, normalization), the same 17-keypoint skeleton, and the same PCK head-size normalization as MultiFormer. If MultiFormer no longer achieves the highest PCK@20 (0.7225), or if the reported per-joint gains shrink materially (e.g., the hand PCK improvement drops below 10%), the headline SOTA claim is not supported. If no code is available, request the authors' exact baseline configurations and splits; without these, Table VIII cannot be audited.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MultiFormer outperforms prior CSI pose estimators on MM-Fi and on the authors' dataset. The load-bearing premise is that every baseline in Tables VII and VIII was evaluated under one comparable protocol. That premise is not documented and is contradicted by the tables. WPFormer is listed as 30.81M parameters / 8.26G FLOPs on the collected dataset (Table VII) and as 26.42M / 507.89G FLOPs on MM-Fi (Table VIII). A roughly 60x FLOPs difference for the same named model implies different input resolutions, model variants, or numbers taken from different publications. The paper never states which split of MM-Fi was used, how the 17-keypoint MM-Fi annotations were reconciled with the 18-keypoint output, whether baselines were retrained or imported from their papers, or how PCK normalization was applied across datasets. If baselines were imported from heterogeneous sources, the head-to-head ranking—especially the 0.7225 vs 0.6841 PCK@20 lead over CSI2Pose and the over-500% hand-joint gains over HPE-Li—could be an artifact of different evaluation code, label sets, or data splits rather than of the architecture. This is the most load-bearing concern because it directly controls the abstract's central claim. The teacher-label issue is secondary: OpenPose-derived labels are disclosed and standard for the literature, but they cap the self-collected accuracy. No code, data, or baseline configuration is released, so the comparison cannot be audited.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MultiFormer, a CSI-based multi-person 2D pose estimation system. It introduces Time-Frequency Dual-Dimensional Tokenization (TFDDT), which converts upsampled CSI amplitude into separate temporal and frequency token streams processed by two parallel multi-head self-attention encoders; the extracted features are fed to a Multi-Stage Feature Fusion Network (MSFN) that iteratively refines PCM and PAF heatmaps using channel and spatial attention, and the final pose is decoded via NMS and Hungarian matching. Training labels are obtained from an OpenPose teacher on synchronized webcam video, and the student network is evaluated on a self-collected two-person dataset and on the public MM-Fi dataset. The paper reports consistent improvements in PCK over several baselines, with the largest gains on wrist and elbow keypoints, and claims state-of-the-art accuracy across PCK@5 to PCK@40.","tokens_in":20580,"tokens_out":5272,"duration_ms":63658,"significance":"If the reported results are reliable, the architecture is a plausible contribution to privacy-preserving, camera-free multi-person pose estimation: the dual-token transformer design is a clean way to capture both subcarrier and temporal structure, the multi-stage fusion with PAPM is well motivated, and the teacher-student label pipeline is disclosed rather than hidden. The ablation study is internally consistent and isolates single factors, which is a genuine strength. However, the current evidence does not yet support the abstract's headline claim of superiority over state-of-the-art methods: the baseline comparison protocol is undocumented and internally inconsistent, the self-collected evaluation uses only two volunteers and no error bars, and the MM-Fi skeleton is not reconciled with the model's 18-keypoint output. The engineering ideas are interesting and potentially publishable, but the experimental validation needs substantial rework.","major_comments":[{"comment":"The baseline comparison is not based on a documented shared protocol, and the tables contradict the assumption of comparability. WPFormer is listed as 30.81M parameters and 8.26G FLOPs in Table VII but as 26.42M parameters and 507.89G FLOPs in Table VIII; WISPPN is 20.19M/7.70G in Table VII but 26.78M/159.81G in Table VIII. A fixed architecture cannot have different parameter counts across tables, so these numbers must come from different model variants or different publications with different input resolutions and evaluation settings. The paper never states the MM-Fi train/test split, whether baselines were retrained or imported, the input size used for each baseline, or how PCK normalization was applied. This makes the headline MM-Fi ranking, including the 0.7225 versus 0.6841 PCK@20 lead over CSI2Pose, unsupported as a head-to-head comparison.","section":"IV-C, Tables VII and VIII"},{"comment":"Teacher labels are generated by OpenPose from an approximately 8 FPS webcam and used as ground truth for both training and evaluation of the student. The paper does not describe how the 50 Hz CSI frames are synchronized with 8 FPS video, nor how frames in which OpenPose fails (or produces uncertain keypoints) are handled. Because the evaluation uses the same teacher outputs used for training, the self-collected PCK measures agreement with the teacher rather than physical pose accuracy. This is especially relevant to the claimed robustness in darkness in Section IV-C, since the student is trained on OpenPose labels rather than on independently verified pose data.","section":"IV-A and IV-C, Eqs. (13)-(15), Fig. 15"},{"comment":"MM-Fi is described as containing 17-keypoint pose annotations, while MultiFormer outputs an 18-keypoint OpenPose-style skeleton with a 19-channel PCM and 38-channel PAF representation. The paper never states how MM-Fi's 17 keypoints are mapped to the 18-keypoint output, how the teacher PCM/PAF labels are generated on MM-Fi, or how the 17-keypoint results in Figs. 13 and 14 are derived when the model produces 18 keypoints. The keypoint names in Fig. 13 (e.g., Bot Torso, Center Torso) are not the standard OpenPose/COCO 18-keypoint names, which suggests a separate label convention. Without this mapping, the MM-Fi results in Table VIII and Figs. 13-14 cannot be audited.","section":"Table I, Section IV-A, Figs. 13 and 14"},{"comment":"The PCK definition in Eq. (16) is ambiguous or incorrect as written. The denominator is sqrt(rs^2 + lh^2), where rs and lh are described as the 'positions' of the right shoulder and left hip. A standard PCK normalizes the keypoint error by a scale derived from the torso or head segment length, not by the Euclidean norm of two coordinate vectors, so Eq. (16) is not a standard PCK and its normalization is unclear. Since all reported results are PCK values, this issue affects every table and the cross-dataset comparability of Tables VII and VIII.","section":"IV-A, Eq. (16)"},{"comment":"Every ablation and comparison table reports a single run with no error bars or cross-validation, and the self-collected dataset has only two volunteers (Table I). The central quantitative claims, such as the 11% improvement from dual-token over single-token modeling (Table III) and the 12%-64% stage-wise improvements in Table IV, would be strengthened substantially by reporting mean and standard deviation over multiple runs or subject-wise cross-validation. As reported, the differences could be within run-to-run or subject-dependent noise. This does not invalidate the architecture, but it limits confidence in the magnitude of the claimed gains.","section":"IV-A, IV-B, Tables III-VIII"}],"minor_comments":[{"comment":"The abstract contains a duplicated word: 'Extensive experiments conducted on on the public MM-Fi dataset'.","section":"Abstract and Section III-C"},{"comment":"The text has another duplication: 'at at different stages' in Section III-C and the caption of Fig. 7.","section":"Section III-C and Fig. 7"},{"comment":"Several values in Table VI are missing decimal points or formatting, e.g., '02809', '02603', '02550', which should be '0.2809', '0.2603', '0.2550'.","section":"Table VI"},{"comment":"The text and captions for Figures 3, 4, and 8 contain unresolved glyph codes such as '/uni00000013' instead of readable labels; the figures need to be regenerated cleanly before publication.","section":"Figures 3, 4, and 8"},{"comment":"Reference [27], cited for CSI2Pose, is titled 'Towards cross-modal forgery detection and localization on live surveillance videos', which does not match the CSI-pose-estimation method described in the text; the citation should be checked and corrected.","section":"References"},{"comment":"The spelling of the WiSPPN baseline is inconsistent: the text uses both 'WISPPN' and 'WiSPPN', and Table VI uses 'WISPPN' while Table VII uses 'WiSPPN'. Please unify the notation.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready for acceptance in its current form because the SOTA comparison, which is the core claim, rests on an undocumented and internally inconsistent benchmark protocol. The architectural contributions are plausible and the ablations are sensible, so I would not reject: the authors should be asked to rerun or re-report all baselines under a single documented protocol, clarify the skeleton mapping for MM-Fi, correct the PCK definition, and provide variance estimates or a larger evaluation cohort. If these cannot be provided, the claims should be correspondingly weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Start with the thing worth knowing: the dual-token architecture is the real contribution here, and the ablations support it. MultiFormer tokenizes CSI into separate frequency and temporal streams, runs two independent multi-head attention stacks, and feeds the fused features through a three-stage PCM/PAF decoder with channel-spatial attention. That combination is not present in the cited prior work. The ablations are internally consistent: dropping either token stream costs about 10% PCK@5, and cutting from three decoder stages to one hits wrists and elbows hardest. The OpenPose teacher dependency is disclosed up front and is standard for this subfield. So the internal story is coherent.\n\nThe soft spot is the benchmarking, and it is load-bearing. Table VIII lists WPFormer at 26.42M parameters and 507.89G FLOPs on MM-Fi, while Table VII lists the same model at 30.81M and 8.26G on the collected dataset. A fixed architecture cannot change parameter count between datasets; the numbers were clearly pulled from different sources with different splits or variants. The paper never states which MM-Fi split was used, whether baselines were retrained, or how the 17-keypoint MM-Fi annotations were reconciled with the 18-keypoint skeleton. That means the leading PCK numbers over CSI2Pose and the huge hand-joint gains over HPE-Li are not yet trustworthy. On top of that, the self-collected data comes from two volunteers, every table reports a single run with no error bars, and the multi-person claim is supported only by qualitative figures, not a quantitative metric.\n\nNone of this sinks the architecture. The dual-token idea is worth testing, and the ablations suggest it does real work. But the paper overclaims in its current form. The fix is concrete: release code and data, standardize one evaluation protocol across all baselines, add error bars, provide a quantitative multi-person evaluation, and reconcile the skeleton definitions.\n\nFor a reader: this is a subfield paper for people working on CSI pose estimation. I would not cite the SOTA numbers until the comparison is redone, but I would send it to review because the core idea and ablations deserve referee scrutiny. Expected outcome: major revision.","headline":"The dual-token CSI-Transformer idea is genuinely new and the ablations are coherent, but the SOTA comparison is undermined by an undocumented, internally inconsistent baseline protocol, so the headline accuracy claims are not yet established.","tokens_in":21173,"tokens_out":2040,"would_cite":false,"duration_ms":25103,"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":"A WiFi-only network can estimate multi-person body poses, including wrists and elbows, from radio signals.","keywords":["human pose estimation","Channel State Information","WiFi sensing","Transformer","multi-head self-attention","multi-person pose estimation","part affinity fields","multi-stage feature fusion"],"falsifier":"Run the original released code for every compared method on a single fixed train/test split of each dataset and recompute PCK; if any model's parameter count or FLOPs changes between the two reported tables, the numbers were drawn from different publications and the comparison is not controlled.","tokens_in":20039,"feed_emoji":"📡","tokens_out":5692,"duration_ms":58793,"temperature":0.7,"pith_summary":"MultiFormer is a WiFi-based system that estimates the 2D skeleton of one or more people from Channel State Information (CSI), the fine-grained radio measurements of how a WiFi signal reflects off a room and its occupants. The paper's claim is that a dual-token Transformer, which models subcarrier correlations and temporal dynamics in separate streams, combined with a three-stage heatmap refinement network, yields more accurate pose estimates than previous WiFi pose estimators, especially for the wrists and elbows. The authors report the highest Percentage of Correct Keypoints scores among compared methods on both a self-collected dataset (PCK@5 of 0.5209) and a public benchmark dataset (PCK@20 of 0.7225), and demonstrate that the system works in darkness and in two-person scenes. If correct, this means privacy-preserving, camera-free, device-free human pose tracking is feasible with commodity WiFi hardware.","feed_headline":"CSI-only transformer beats prior WiFi pose estimators on wrists","feed_subtitle":"Multi-person skeletons from radio reflections, with the biggest gains on high-mobility joints like wrists and elbows.","key_machinery":"The load-bearing object is the Time-Frequency Dual-Dimensional Tokenization (TFDDT): raw CSI amplitude is upsampled in both time and frequency, then grouped into 64 frequency tokens, each spanning time and antennas for one subcarrier, and 64 temporal tokens, each spanning subcarriers and antennas at one time instant. Two parallel multi-head self-attention streams with independent parameters process the two token types, and a reconstruction layer reshapes the one-dimensional outputs into 36x36 feature maps. The second mechanism is the Multi-Stage Feature Fusion Network (MSFN), whose Pose-Attentive Perception Module computes channel and spatial attention from the previous stage's PCM/PAF heatmaps and multiplies them into the CSI features before the next heatmap decoder. This iterative refinement is what allows the system to associate keypoints across multiple people and to focus progressively on body regions rather than background clutter.","core_discovery":"The central discovery is that CSI amplitude, when organized into separate time tokens and frequency tokens and processed by parallel multi-head self-attention modules, carries enough information to reconstruct multi-person 2D skeletons. Unlike prior approaches that treat CSI as an image-like tensor, regress keypoint coordinates directly, or encode only pairwise keypoint distances, MultiFormer produces Part Confidence Maps and Part Affinity Fields as intermediate heatmaps and refines them across three decoding stages. A Pose-Attentive Perception Module re-weights the CSI features channel-wise and spatially using the previous stage's heatmap output, which the authors argue enforces anatomical constraints. On their own dataset the full model reaches PCK@5 of 0.5209 and PCK@20 of 0.8885, and on the public benchmark PCK@20 of 0.7225, with the largest gains on wrists and elbows.","pith_inferences":["The dual-token separation suggests that treating frequency and time as independent relational dimensions before fusing them may transfer to other CSI tasks, such as activity recognition or device-free localization, where similar tokenization could help.","The reported wrist and elbow gains could be probed more directly by evaluating on a larger public dataset with more subjects and multiple room layouts; if the margins persist across environments, the improvement is tied to the architecture rather than to the specific recording setup.","The teacher-student setup means the system inherits whatever biases the vision-based teacher has, so a natural next test is training from motion-capture labels to see whether the CSI estimator can exceed the teacher's precision.","A three-stage decoder with attention reweighting is a generic refinement recipe that could be grafted onto other sensor modalities, not just WiFi, wherever intermediate heatmaps are available."],"forward_implications":["A CSI-only system can estimate multi-person skeletons without cameras, wearables, or specialized radar hardware, so pose tracking remains available in darkness and privacy-sensitive settings.","The largest accuracy jumps occur on wrists and elbows, so gesture recognition and hand-motion analysis are the most promising near-term applications.","Multi-stage heatmap refinement with attention to the previous stage's output offers a way to enforce anatomical consistency that direct regression and pairwise-distance methods lack.","The full model uses 11.93M parameters and 15.12G FLOPs, while smaller variants still beat comparable-parameter baselines, so the approach can run at modest compute budgets.","Because the pipeline outputs the same part-confidence maps and part-affinity fields used in vision-based pose estimation, existing pose-decoding and multi-person association algorithms can be reused directly."],"supporting_citations":[{"why":"Supplies the vision-based teacher network that generates the PCM/PAF ground-truth labels and the NMS/Hungarian association machinery used at inference.","marker":"[5]"},{"why":"A transformer-based WiFi pose estimation baseline whose PCK results and parameter counts appear in both comparison tables.","marker":"[20]"},{"why":"The CSI-Former baseline that models inter-subcarrier correlations with attention, which MultiFormer extends by also modeling temporal dependencies.","marker":"[22]"},{"why":"The WiSPPN baseline using pose adjacency matrices, used to illustrate single-person and multi-person failure cases that MultiFormer overcomes.","marker":"[23]"},{"why":"The heatmap-based multi-person CSI pose estimation baseline that MultiFormer outperforms in per-keypoint comparisons.","marker":"[27]"},{"why":"Defines the multi-head self-attention mechanism that the dual-token feature extractor is built on.","marker":"[34]"},{"why":"Provides the public MM-Fi benchmark dataset used for the second set of evaluation experiments.","marker":"[35]"},{"why":"Defines the Percentage of Correct Keypoints (PCK) metric used for all accuracy reporting.","marker":"[36]"}],"fun_headline_variants":["WiFi pulse skeletons: attention model tops joint accuracy","Transformer turns WiFi CSI into multi-person skeletons","Radio reflections reveal poses: attention beats old methods","CSI attention model excels at high-mobility wrist poses","Multi-person WiFi pose via attention on CSI frames"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline accuracy comparisons assume that every prior model was trained and evaluated under one shared protocol on each dataset, a premise the paper never documents and its own tables appear to contradict; the system also assumes that the vision-based teacher's 8 FPS webcam labels are accurate ground truth for both training and evaluation.","fun_headline_variants_meta":{"raw":{"variants":["WiFi pulse skeletons: attention model tops joint accuracy","Transformer turns WiFi CSI into multi-person skeletons","Radio reflections reveal poses: attention beats old methods","CSI attention model excels at high-mobility wrist poses","Multi-person WiFi pose via attention on CSI frames"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000557,"raw_usage":{"total_tokens":2618,"prompt_tokens":881,"completion_tokens":1737,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":1665}},"tokens_in":497,"tokens_out":1737,"duration_ms":13294,"temperature":1.0,"reasoning_tokens":1665,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:05:57.208058+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the original released code for every compared method on a single fixed train/test split of each dataset and recompute PCK; if any model's parameter count or FLOPs changes between the two reported tables, the numbers were drawn from different publications and the comparison is not controlled.","supporting_citations":[{"cited_title":"Openpose: Re- altime multi-person 2d pose estimation using part affinity fields,","cited_arxiv_id":null,"evidence_quote":"Supplies the vision-based teacher network that generates the PCM/PAF ground-truth labels and the NMS/Hungarian association machinery used at inference."},{"cited_title":"MetaFi++: WiFi-Enabled Transformer-Based Human Pose Estimation for Metaverse Avatar Simulation,","cited_arxiv_id":null,"evidence_quote":"A transformer-based WiFi pose estimation baseline whose PCK results and parameter counts appear in both comparison tables."},{"cited_title":"CSI-Former: Pay More Attention to Pose Estimation with WiFi,","cited_arxiv_id":null,"evidence_quote":"The CSI-Former baseline that models inter-subcarrier correlations with attention, which MultiFormer extends by also modeling temporal dependencies."},{"cited_title":"Towards cross- modal forgery detection and localization on live surveillance videos,","cited_arxiv_id":null,"evidence_quote":"The heatmap-based multi-person CSI pose estimation baseline that MultiFormer outperforms in per-keypoint comparisons."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"Defines the multi-head self-attention mechanism that the dual-token feature extractor is built on."},{"cited_title":"2d human pose estimation: New benchmark and state of the art analysis,","cited_arxiv_id":null,"evidence_quote":"Defines the Percentage of Correct Keypoints (PCK) metric used for all accuracy reporting."}],"review_version":1}