{"id":"b68fb885-8a33-4792-bd39-ea147b80bb82","arxiv_id":"2608.08381","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DoRF++ recovers a latent 3D velocity sequence from CSI Doppler projections via rank-3 matrix factorization, re-projects it onto a sphere, and classifies gestures with spherical attention, reporting 80.4% cross-user accuracy on a four-gesture Wi-Fi dataset.","lead":"This paper builds a 3D motion representation from Wi-Fi Doppler signals by treating each antenna pair as a virtual camera, then classifies hand gestures with a spherical Transformer. It reports large cross-user accuracy gains over prior Wi-Fi sensing methods on its own collected gesture dataset, but without public data or significance tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader's verdict is CONDITIONAL with moderate confidence, and the weakest assumption is that each common-RX CSI-ratio stream's dominant MUSIC Doppler peak is a linear projection of the true hand velocity onto an unknown effective direction that stays constant over the trial. I agree that this is the most fragile physical modeling assumption in the derivation, and the paper defers the full derivation to a self-cited prior work (MORIC [14]) and does not validate the recovered latent motion V against ground-truth motion. However, the central empirical claim, strong cross-user generalization of DoRF++, is supported by the ablation structure: gains are consistently attributed to the common-RX sanitization, the DoRF factorization, and the spherical attention, with the full model exceeding all ablated baselines. The assumption's failure would undermine the physical interpretation of the learned representation but would not necessarily invalidate the classification result, since the DoRF construction is still a deterministic, structured transform of the Doppler projections that yields a discriminative and generalizable input for the spherical Transformer. The paper also honestly notes the orthogonal ambiguity and the non-identifiability of the regularized factors beyond Proposition 2's ideal setting. The reader's recommended conditions (code/data release, significance testing, validation of the linear-projection assumption, and cross-environment evaluation) are reasonable and do not overstate or understate the risk. No new concern was found that would change the verdict.","tokens_in":23511,"tokens_out":1117,"duration_ms":12182,"concrete_test":"A single verification step worth running: recompute DoRF++ AP1 accuracy after removing the common-RX phase-sanitization component and replacing it with a standard same-TX/different-RX CSI ratio, keeping all other pipeline elements fixed. If the sanitization component is indeed essential, the ablation should reproduce the reported 80.4% vs 66.9% gap under identical preprocessing and evaluation settings.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"No load-bearing concern identified. The paper's central claim is that DoRF++ significantly outperforms state-of-the-art Wi-Fi-based HAR methods in cross-user generalization accuracy, especially for a single receiver AP. The reported gains are consistently supported by the ablation study in Table IV, where the full model (80.4%) outperforms each ablated variant (excluding the common-RX phase sanitization drops to 31.2%; replacing it with conventional same-TX/different-RX CSI ratio drops to 66.9%; removing spherical position encoding drops to 74.1%; direct latent V representation reaches only 68.2%; PCA-based baselines reach 60.9-62.1%). The weakest assumption identified by the reader, that each common-RX CSI-ratio stream's dominant MUSIC Doppler peak is a linear projection of the true hand velocity onto a stream-specific effective direction that stays constant within a window (Eq. 12-14, 23), is a legitimate limitation for the physical interpretability and identifiability of the recovered latent motion. However, the empirical classification claim does not depend on recovering physically meaningful 3D motion: the DoRF representation is a fixed spherical re-projection of the fitted latent sequence, and the ablations show that this construction outperforms PCA-based low-rank fits, direct sequence models (LSTM/CNN), and the raw latent V. The paper honestly states that the recovered coordinate frame has an orthogonal ambiguity and explicitly cautions that Proposition 2 is an ideal reference, not a guarantee for the regularized factors from (27). Therefore, the conditional verdict, which asks for code/data release, significance testing, ground-truth validation of the linear-projection assumption, and cross-environment evaluation, is appropriate. No additional breaking objection is identified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes DoRF++ for Wi-Fi-based human activity recognition. The pipeline first forms common-receive-antenna CSI ratios between transmit streams to suppress synchronization phase errors, estimates a single dominant Doppler velocity per ratio stream via MUSIC, and then fits, per receive antenna, a regularized rank-3 matrix factorization Vr ≈ VR that is interpreted as latent 3D hand velocity observed through unknown effective view directions. The recovered velocity sequence is re-projected onto a fixed equiangular spherical grid (the Doppler radiance field), and classification is performed by a spherical Transformer with quadrature-aware attention and rotation-invariant relative position biases. Experiments on a self-collected 10-user, 4-gesture dataset (UTHAMO-5G) with three receiver APs, under a leave-one-subject-out protocol, report that DoRF++ reaches 80.4% accuracy on AP1 versus 66.4% for MORIC and 67.1% averaged over APs versus 54.9% for MORIC; ablations (Table IV) attribute the gains to the common-RX phase sanitization, the DoRF factorization, and spherical attention.","tokens_in":23788,"tokens_out":12826,"duration_ms":141403,"significance":"The claimed gains are practically relevant if they hold: fine-grained cross-user hand-gesture recognition from a single Wi-Fi receiver AP is a hard and useful setting. The evaluation has several strengths: the LOSO protocol keeps training, validation, and test subjects disjoint; the ablation that feeds the DoRF representation into the MORIC classifier (DoRF vs MORIC in Table I) isolates the representation from the classifier; and the paper is unusually honest about the ideal nature of its identifiability result (Proposition 2) and the residual orthogonal ambiguity. The alternating-minimization convergence argument is standard but correctly stated. The main weaknesses are statistical: no significance tests are reported for the headline comparisons, and the 'state-of-the-art' claim rests on a narrow baseline set. The physical interpretation of the recovered latent velocity is not validated against ground truth, so the radiance-field framing currently outruns the evidence, although this does not invalidate the empirical classification claim.","major_comments":[{"comment":"The central claim that DoRF++ 'significantly outperforms' prior methods is not backed by any statistical test. The ten LOSO folds are paired by test subject, so the authors should report paired tests (e.g., Wilcoxon signed-rank or paired t-test) for DoRF++ versus MORIC and DoRF++ versus DoRF, for each AP and for the AP-average, together with 95% confidence intervals. As reported, the AP1 gap (80.4±8.4 vs 66.4±7.8) and the AP3 gap (54.2±8.6 vs 42.4±6.8) overlap by roughly one standard deviation, so the reader cannot tell whether the gains are consistent across subjects or driven by a few folds. This is necessary to substantiate 'significantly outperforms' in the abstract and conclusion.","section":"§IV-C1, Table I"},{"comment":"The latent motion V is never validated against ground-truth motion, yet the abstract states that DoRF 'infers a latent 3D motion sequence.' Because the spherical DoRF field P is a deterministic re-projection of V, the ablation results show that the rank-3 factorization and spherical re-projection help classification but do not establish physical interpretability. The projection model in Eq. (23) is asserted with the derivation deferred to MORIC [14], and Proposition 2 is explicitly ideal. I recommend either (i) validating V with motion capture or with synthetic Doppler data generated from known 3D trajectories, or (ii) rephrasing the contribution as a learned low-rank spherical representation inspired by radiance fields, rather than as recovery of the physical 3D motion. The current wording overstates the evidence.","section":"§III-D, Eq. (23)-(27); Table IV"},{"comment":"The 'state-of-the-art' comparison is too narrow to support the claim. The only recent strong baseline is MORIC [14], which is from the same group; Widar3.0 [13], cited in the introduction as a cross-domain gesture recognition method, is not included, nor are other independently developed 2023-2025 Wi-Fi HAR systems. The authors should add at least one independent recent baseline or soften the claim to 'outperforms the evaluated methods on UTHAMO-5G.'","section":"§IV-A, Table I"}],"minor_comments":[{"comment":"The common-RX ratio cancels receiver-side STO/SFO but does not cancel TX-specific cyclic-shift-diversity delays or per-stream beamforming phases; the text should state that these residuals are static (hence Doppler-free) to avoid an imprecise claim about canceling all common phase terms.","section":"§III-A, Eq. (5)-(6)"},{"comment":"With four transmit antennas there are exactly six pairwise difference directions per receive antenna, and these six differences satisfy linear dependencies; the paper should comment on whether the spanning condition of Proposition 2 can actually be met under this construction, rather than only noting that N<6 is insufficient.","section":"§III-D, Proposition 2"},{"comment":"Please state explicitly whether the DoRF matrix factorization is fit independently per trial (i.e., without using labels or other trials), since this is important for ruling out information leakage in the LOSO protocol.","section":"§IV-B"},{"comment":"The 'mean±std' in Tables I and III should specify that the standard deviation is across the ten LOSO test subjects, and it would help to report per-fold results in a supplementary table.","section":"§IV-A"},{"comment":"Minor presentation issues: the SFO term notation f'_c/f_c-1 in Eq. (5) is hard to parse, and the lower-triangular entries of the binary-accuracy matrix in Table III appear to be missing or misaligned in the typeset version.","section":"Eq. (5) and Table III"}],"recommendation":"major_revision","confidential_remarks":"The paper is plausible and the core empirical claim appears to hold on this dataset, but I could not verify it from the reported aggregates alone; the authors should provide per-fold numbers and significance tests. The narrow baseline set and the unvalidated physical interpretation of V are the main issues. I see no reason to reject on scientific grounds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: DoRF++ is a solid, honest empirical paper that makes a plausible case for treating Wi-Fi Doppler projections as low-rank views of a latent 3D velocity, then classifying a spherical re-projection with a spherical Transformer. The ablation is the real contribution: it isolates the representation from the classifier in a clean leave-one-subject-out protocol, and the full model beats the ablated variants by margins that look consistent. I came away convinced the pipeline does something useful on their dataset.\n\nWhat's new: the common-RX CSI-ratio phase sanitization is a simple, sensible fix to a known problem; factoring Doppler projections into V and R is a reasonable extension of MORIC; and applying quadrature-aware spherical attention to the re-projected field is a natural fit. The relative positional bias sensibly addresses the orthogonal ambiguity. The paper is also candid: it explicitly says Proposition 2 is an ideal reference, that the recovered frame has an orthogonal ambiguity, and that spherical attention is only approximately equivariant. That's the right tone.\n\nThe soft spots are proportionate. The load-bearing assumption—dominant MUSIC peak per CSI-ratio stream is a linear projection of hand velocity onto a fixed effective direction—is deferred to MORIC and never validated against ground truth. If that fails, the recovered V is a fitted artifact. But the classification results don't depend on physical interpretability; the ablation shows the re-projection helps even if the latent motion isn't the true hand motion. More consequential is the evidence base: one room, one body orientation, ten participants, no significance tests, no code or data. AP3's gain is real but small and noisy. Also, the main comparison is to the authors' own MORIC; that's the right baseline, but independent replication would help. Self-citation to MORIC for the projection model is fine, yet the deferred derivation makes the core model hard to evaluate internally.\n\nWho is this for? Wi-Fi sensing folks working on cross-user generalization, and anyone interested in spherical attention for structured representations. It deserves a serious referee—I'd send it out, not desk reject—but I'd condition acceptance on releasing data/code, adding significance testing, and ideally a second environment or a ground-truth check of the projection model.","headline":"DoRF++ is a well-ablated empirical paper with a plausible central claim, but the unvalidated linear-projection assumption and single-environment evidence keep it conditional.","tokens_in":24363,"tokens_out":4024,"would_cite":true,"duration_ms":44118,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Wi-Fi hand gestures can be recognized across unseen users from a single access point by reconstructing Doppler projections as views of a latent 3-D motion.","keywords":["Wi-Fi sensing","channel state information","Doppler radiance fields","human activity recognition","spherical Transformer","cross-user generalization","MUSIC Doppler estimation","hand gesture recognition"],"falsifier":"Take synthetic CSI generated from a known 3-D hand velocity with known projection directions, run MUSIC and DoRF, and check whether the recovered $V$ matches the ground-truth velocity up to a global orthogonal transform; in real data, record the same gestures with an optical motion-capture system alongside the Wi-Fi receiver and test whether dominant Doppler peaks and the recovered latent trajectory track the measured hand velocity. If classification accuracy stays high while the recovered motion fails to align with ground truth, the geometric mechanism is not the actual source of the reported gains.","tokens_in":23305,"feed_emoji":"📡","tokens_out":6734,"duration_ms":71031,"temperature":0.7,"pith_summary":"The paper claims that the many Doppler velocity signals extracted from Wi-Fi CSI are not independent noisy descriptors but sparse, direction-specific observations of one underlying three-dimensional hand motion. It introduces DoRF, which recovers that latent motion by a regularized rank-three matrix factorization, then re-projects it onto a fixed sphere; DoRF++ adds spherical attention for classification. On a collected four-gesture dataset, this raises leave-one-subject-out accuracy from 66.4% to 80.4% for the best single receiver, and from 54.9% to 67.1% averaged over three receiver placements. If correct, the practical upshot is that privacy-preserving gesture recognition needs no wearable and no dense sensor array: a single multi-antenna router can do it.","feed_headline":"One Wi-Fi access point reads hand gestures across strangers","feed_subtitle":"Treating Doppler velocity trails as camera views, then learning on the sphere, lifts single-AP accuracy to 80%.","key_machinery":"The load-bearing object is the Doppler Radiance Field tensor $\\mathcal{P}\\in\\mathbb{R}^{T\\times M\\times 2M}$, formed by re-projecting the recovered latent velocity sequence $V\\in\\mathbb{R}^{T\\times 3}$ onto an equiangular latitude-longitude grid of unit directions $d_{mn}$ via $P(s,m,n)=v(s)^\\top d_{mn}$. The recovery step treats the observed Doppler projections $V_r$ as the product $VR+E$ of latent motion and unknown effective projection directions, solved by alternating ridge-regression updates; a global orthogonal ambiguity remains, which motivates the spherical learning stage. DoRF++ then treats each spherical direction as a token, extracts temporal features with fixed random convolution kernels, and applies spherical multi-head attention whose scores carry a relative positional bias built from Legendre polynomials of the cosine between directions and are weighted by spherical quadrature weights.","core_discovery":"On its own terms, the paper claims that the seemingly chaotic multipath Doppler information inside Wi-Fi CSI can be organized into a coherent picture of the underlying body motion. For each receive antenna it forms ratios between transmit streams that share that antenna, removing receiver-side synchronization errors; from each ratio stream it extracts one dominant Doppler velocity per time window with MUSIC; then it solves a regularized rank-three factorization $V_r \\approx VR$ for a latent 3-D velocity sequence $V$ and unknown effective projection directions $R$. The recovered motion is re-projected onto an equiangular sphere, and a spherical Transformer with quadrature-aware attention classifies the resulting Doppler radiance field. In the paper's leave-one-subject-out experiment, this raises single-receiver accuracy to 80.4% from 66.4% for the strongest prior method, with the biggest gains on gesture pairs that look alike from some viewpoints.","pith_inferences":["Nothing in the method constrains the hand to be the only moving object; if the dominant MUSIC peak occasionally tracks a torso or a stationary reflector, the recovered latent motion would be a fitted artifact. Comparing recovered $V$ against optical motion capture in a controlled setup would make the geometric claim directly measurable; the paper stops at classification accuracy.","The same common-RX phase-sanitization and DoRF construction should transfer to other Doppler-capable RF sensors and to swapped transmitter-receiver roles, since the model only assumes linear projections of a latent 3-D velocity; this transfer is not tested here.","The dominant-peak truncation discards secondary Doppler peaks, which likely carry cross-user information for the confusing gesture pairs; a full-spectrum or multi-peak generalization of the factorization is a natural extension.","The per-trial orientation ambiguity means the recovered latent motion cannot yet be compared across trials without alignment; canonical orientation alignment would turn DoRF from a classifier input into a motion-capture-like representation for pose tracking."],"forward_implications":["A single multi-antenna Wi-Fi receiver is enough for practical cross-user gesture recognition, with accuracy tied to how directly the user lies on the dominant propagation path.","Unordered aggregation of Doppler projections leaves accuracy on the table; recovering a shared 3-D motion before classification is what unlocks the reported gain.","Because the global orientation of the recovered latent motion is arbitrary per trial, classifiers must be rotation-aware, and spherical attention plus max-pooling over antennas is the paper's proposed cure.","The largest remaining confusions are between gestures that become equivalent under a rotation of the coordinate axes, such as left-right versus push-pull, which is an intrinsic limit of Doppler-only sensing.","The ablation results indicate that the common-RX phase-sanitization step is responsible for the single largest accuracy swing, so its quality directly determines whether the rest of the pipeline has clean phase dynamics to work with."],"supporting_citations":[{"why":"Supplies the virtual-camera Doppler projection model and the classifier baseline that DoRF builds on and must beat.","marker":"[14]"},{"why":"Introduces the neural radiance field concept that DoRF adapts from 2-D view synthesis to Doppler observations.","marker":"[17]"},{"why":"Shows how unposed images can be factorized into a shared scene and unknown poses, the template for recovering latent motion and unknown effective directions together.","marker":"[4]"},{"why":"MUSIC is the spectral algorithm used to extract the dominant Doppler frequency from each CSI-ratio stream.","marker":"[21]"},{"why":"Introduces the CSI-ratio phase-cancellation construction that the paper reworks into its common-RX form.","marker":"[19]"},{"why":"Prior Doppler-velocity extraction and antenna-pair-selection pipeline that serves as a generalization baseline.","marker":"[20]"},{"why":"Provides the quadrature-aware spherical attention module that DoRF++ adopts as its classifier backbone.","marker":"[25]"},{"why":"Supplies the fixed random-convolution kernels that generate the per-direction temporal feature tokens feeding the spherical Transformer.","marker":"[23]"},{"why":"Raw-CSI magnitude baselines that set the low end of the cross-user comparison for DoRF++.","marker":"[24]"}],"fun_headline_variants":["One Wi-Fi AP reads gestures via spherical Doppler radiance fields","Spherical Doppler radiance fields let one Wi-Fi AP see gestures","DoRF++: spherical learning reads Wi-Fi gestures on one AP","Treat Wi-Fi Doppler as camera views, learn on sphere for gestures","Single Wi-Fi AP recognizes gestures via spherical Doppler views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole reconstruction rests on assuming that the dominant Doppler peak extracted from each common-RX, different-TX CSI-ratio stream is a linear projection of the same hand velocity onto a direction that stays fixed during the trial; if the peak follows some other scatterer or the effective direction rotates mid-trial, the recovered 3-D motion is a fitted artifact rather than true hand motion.","fun_headline_variants_meta":{"raw":{"variants":["One Wi-Fi AP reads gestures via spherical Doppler radiance fields","Spherical Doppler radiance fields let one Wi-Fi AP see gestures","DoRF++: spherical learning reads Wi-Fi gestures on one AP","Treat Wi-Fi Doppler as camera views, learn on sphere for gestures","Single Wi-Fi AP recognizes gestures via spherical Doppler views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000764,"raw_usage":{"total_tokens":3421,"prompt_tokens":1007,"completion_tokens":2414,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":2327}},"tokens_in":623,"tokens_out":2414,"duration_ms":20242,"temperature":1.0,"reasoning_tokens":2327,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:37:51.817203+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take synthetic CSI generated from a known 3-D hand velocity with known projection directions, run MUSIC and DoRF, and check whether the recovered $V$ matches the ground-truth velocity up to a global orthogonal transform; in real data, record the same gestures with an optical motion-capture system alongside the Wi-Fi receiver and test whether dominant Doppler peaks and the recovered latent trajectory track the measured hand velocity. If classification accuracy stays high while the recovered motion fails to align with ground truth, the geometric mechanism is not the actual source of the reported gains.","supporting_citations":[{"cited_title":"Wi-Fi CSI-based device-free sensing: From fresnel zone model to CSI-ratio model,","cited_arxiv_id":null,"evidence_quote":"Introduces the CSI-ratio phase-cancellation construction that the paper reworks into its common-RX form."},{"cited_title":"Enhancing generalization in human activity recognition through improved Wi-Fi channel state information phase processing and antenna pair selection,","cited_arxiv_id":null,"evidence_quote":"Prior Doppler-velocity extraction and antenna-pair-selection pipeline that serves as a generalization baseline."},{"cited_title":"Rocket: exceptionally fast and accurate time series classification using random convolutional kernels,","cited_arxiv_id":null,"evidence_quote":"Supplies the fixed random-convolution kernels that generate the per-direction temporal feature tokens feeding the spherical Transformer."},{"cited_title":"Joint human orientation-activity recognition using Wi-Fi signals for human-machine interaction,","cited_arxiv_id":null,"evidence_quote":"Raw-CSI magnitude baselines that set the low end of the cross-user comparison for DoRF++."}],"review_version":1}