{"id":"3da277d6-3b63-40ab-896e-ea4f552683b1","arxiv_id":"2507.12132","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Doppler radiance fields, built by factorizing Wi-Fi Doppler projections into 3D velocities and resampling them on a fixed sphere grid, improve cross-user activity recognition accuracy from 51.5% to 54.8% on the UTHAMO dataset.","lead":"This paper proposes Doppler radiance fields (DoRF), a method that reconstructs a 3D velocity description of human motion from Wi-Fi Doppler measurements and then projects it onto a uniformly sampled sphere before activity classification. The authors report improved cross-user generalization accuracy over a prior Doppler-based baseline on a small hand-gesture dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The factorization in Eq. (9) identifies V only up to an arbitrary 3D rotation; without a canonical Procrustes reference, the DoRF is not view-invariant and the reported accuracy may depend on the random initialization.","rationale":"The central claim is that uniform sampling of a reconstructed 3D velocity field yields a view-invariant representation that improves cross-user accuracy. For this to hold, the reconstructed V must be canonical, or the downstream classifier must be invariant to its ambiguity. The paper's factorization has a continuous O(3) gauge symmetry: V and R can be jointly rotated without changing the data fit or any regularizer. The Procrustes step cannot resolve this without an external reference, and none is specified. I considered alternatives: statistical significance of the 3.3 percentage point gap on six subjects is a real weakness, but it affects the strength of evidence rather than whether the method is well-defined; time-invariance of R affects physical fidelity, but the method could still yield a useful latent representation. The gauge problem is more load-bearing because it undermines the uniqueness and view-invariance at the core of the proposed representation, and it is directly testable by seed and rotation perturbation. The reader's conditional verdict already captures the need for additional evidence; this concern makes the condition concrete: the authors must demonstrate rotation invariance or define a canonical alignment before the claimed generalization can be attributed to DoRF.","tokens_in":7251,"tokens_out":12903,"duration_ms":163159,"concrete_test":"Run Algorithm 1 with at least 10 random initializations of R on the same UTHAMO trial and compare the resulting DoRF tensors (e.g., mean pairwise cosine distance after optimal alignment). Then, with a single trained classifier, take one held-out subject's estimated V, apply 20 random orthogonal rotations Q to V, regenerate P via Eq. (16), and measure the classification accuracy. If per-seed LOSO accuracy varies by more than the reported 7.2% standard deviation, or rotated DoRFs change predictions materially, the gauge freedom invalidates the view-invariance claim and a canonical alignment must be specified; if accuracy is invariant, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.2, the objective (Eq. 9) and its constraints are invariant under (V,R) -> (VQ,RQ) for any orthogonal Q: Vr = V R^T is unchanged, and both regularizers are preserved. The recovered velocity sequence V is therefore determined only up to an arbitrary global rotation/reflection. The Procrustes step (Algorithm 1, line 5) is described only as applying a rotation to stabilize orientation; without an external reference frame it cannot remove this gauge. The DoRF of Section 3.3 is P(s,m,n) = v(s)^T d_mn, so under V -> VQ it becomes v(s)^T Q d_mn, which for a general Q is not a permutation of the fixed M-by-2M grid values. The MORIC max-pooling in Section 3.4 is invariant to channel permutation, but a general rotation linearly mixes the channels before feature extraction and pooling; element-wise max of transformed features is not generally invariant. Hence the claimed view-invariant, complete representation is not guaranteed, different random initializations of R in Algorithm 1 can produce different DoRFs, and the reported 54.8% versus 51.5% gap may be an artifact of a particular seed rather than evidence for the uniform-sampling mechanism. This is the weakest point of the central claim: even if the accuracy difference is real, the method as specified is not well-defined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Doppler Radiance Fields (DoRF), a Wi-Fi CSI-based motion representation for human activity recognition. The method first estimates a 3D velocity sequence V from 1D Doppler projections via alternating constrained least squares (Eq. (9)), then projects each velocity vector onto a uniform spherical grid to form P(s,m,n), and finally feeds this DoRF into the MORIC classifier. On the UTHAMO dataset with leave-one-subject-out evaluation, the method reports mean accuracy 54.8% (SD 7.2%) versus 51.5% (SD 8.7%) for MORIC and lower accuracy for other baselines. The paper claims that uniform sampling provides a view-invariant, complete motion representation that improves cross-user generalization.","tokens_in":7546,"tokens_out":4447,"duration_ms":50669,"significance":"If the claimed improvement were statistically reliable, the paper would contribute a novel way to turn sparse Doppler projections into a structured, dense representation for Wi-Fi HAR, and the LOSO evaluation on a public dataset is a strength. The algorithmic description is mostly complete, and comparison against several baselines is useful. However, the evidence for the central claim is currently weak: the accuracy gap is within the reported standard deviations, no significance testing or ablation is provided, and the factorization underlying the DoRF has an unresolved rotational gauge freedom that undermines the well-definedness of the proposed representation. The idea is promising but requires substantial additional validation.","major_comments":[{"comment":"The optimization objective and both regularizers are invariant to the simultaneous transformation (V,R) -> (VQ,RQ) for any orthogonal Q, since V_r = V R^T is unchanged, ||v(s)|| is unchanged, and ||r_i||=1 is preserved. The Procrustes step (Algorithm 1, line 5) is described only as 'stabilize orientation' and, without an external reference frame, cannot remove this gauge freedom. Under the DoRF construction of Section 3.3, P(s,m,n)=v(s)^T d_{mn} becomes v(s)^T Q d_{mn}, which for a general Q is not a permutation of the fixed grid values. Therefore the claimed view-invariant, complete representation is not guaranteed and different random initializations of R can produce different DoRFs and different classification inputs. The authors should either fix a canonical orientation (e.g., Procrustes against a fixed template or SVD-based alignment of V) or use a rotation-invariant feature representation, and report sensitivity of accuracy to initialization.","section":"Section 3.2, Eq. (9) and Algorithm 1"},{"comment":"The reported improvement of 54.8% over 51.5% is less than the subject-level standard deviations (7.2% and 8.7%), and no significance test, confidence interval, or per-user results are reported. With only six users in LOSO, this difference could be within chance fluctuation. The claim that uniform sampling 'noticeably enhances' generalization is therefore not supported by the evidence. The authors should report paired per-user accuracy, a statistical test (e.g., Wilcoxon signed-rank or paired bootstrap), and effect sizes. They should also include an ablation that separates the contribution of the 3D velocity reconstruction (Section 3.2) from the uniform sphere sampling (Section 3.3) by comparing, e.g., raw MORIC projections, reconstructed velocity sequence used directly, and the full DoRF.","section":"Section 4.3, Table 1"},{"comment":"The model assumes a single time-invariant set of unit directions R such that V_r(s,i)=v(s)^T r_i + n(s,i). In a multipath environment, the effective reflection directions move with the body and change over time; the factorization is then not the physical model implied by Eq. (5). This matters because the DoRF is built from the recovered v(s); if the model is misspecified, the 'velocity' is a latent factor rather than an actual motion estimate. The authors should either justify the approximation with measurements (e.g., reconstruction error or comparison with an independent motion capture), or soften the physical claims and treat the factorization as a learned low-rank representation. A simple diagnostic would be to report the fit of Eq. (9) on held-out data and the sensitivity of accuracy to the number of retained directions N.","section":"Section 3.2, Eqs. (5)-(9)"}],"minor_comments":[{"comment":"The loss is described as 'DTW-based', but Eq. (9) uses squared error and no DTW definition is given; please specify the exact loss and how DTW is incorporated into the alternating updates.","section":"Section 3.2, Algorithm 1, line 6"},{"comment":"The regularization parameters λ and γ in Eq. (9) are not reported; without them the experiments are not reproducible.","section":"Section 4.2"},{"comment":"Please clarify how the baselines in Table 1 were obtained—whether they are re-implemented with the same preprocessing and training setup or copied from the original papers—and report per-user accuracies.","section":"Section 4.2/4.3"},{"comment":"References [14] and [18] are arXiv preprints; if the methods/dataset have been published elsewhere, please cite the final versions.","section":"References [14] and [18]"},{"comment":"Figure 2 is not referenced in the text at its first occurrence; please add a cross-reference and define the color coding.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central comparison is against the authors' own MORIC method and dataset, and the reported gain is not statistically supported. In addition, the rotational gauge issue means the method as specified may not be deterministic. Given these issues, I cannot recommend acceptance, but the core idea is suitable for a major revision if the authors add a canonical alignment, statistical tests, and ablations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: DoRF is a sensible but incremental extension of MORIC. The genuinely new piece is the low-rank factorization of Doppler projections into a 3D velocity sequence followed by deterministic resampling on a sphere grid. The writing is clear, the evaluation is on a public dataset, and the comparison set is reasonable. But the headline gain—54.8% vs. 51.5%—is within the reported user-level standard deviations, there is no significance test, and there is no ablation isolating the contribution of the new steps. More seriously, the factorization in Eq. (9) is invariant under a global rotation (V,R)->(VQ,RQ), so the recovered velocity is only defined up to an arbitrary rotation. The Procrustes step cannot fix this without an external reference frame. Consequently, the DoRF P(s,m,n)=v(s)^T d_mn changes under that rotation, and the claimed view-invariance is not guaranteed; different random initializations of R can produce different DoRFs and potentially different accuracies. This is not just a technicality: it undermines the central argument that uniform sampling gives a complete, view-invariant representation. The paper would need to specify a canonical reference—say, aligning V to its SVD axes or to a fixed template—and show that results are stable across random seeds.\n\nThe empirical support is thin. A single dataset with six subjects, four gestures, and no confidence intervals is not enough to establish a robust claim. The authors should report per-subject results, run a paired significance test, and add an ablation where the DoRF is replaced by the raw projections (MORIC baseline) and where the 3D velocity estimation is replaced by a simpler averaging scheme. The missing code and unspecified lambda/gamma also make replication harder. The MORIC baseline is the authors' own prior work, which is fair, but an independent implementation would be more convincing.\n\nIf the gauge issue is fixed and the statistics are strengthened, the method could be a useful incremental step for Wi-Fi-based HAR. As it stands, it is a reasonable idea with weak evidence. I'd send it to review, but flag the rotation invariance as a load-bearing issue that needs to be addressed before acceptance.","headline":"A modest, plausible extension of MORIC whose central view-invariance claim is undermined by a rotation gauge in the factorization, and whose reported gain is within noise.","tokens_in":8064,"tokens_out":3246,"would_cite":false,"duration_ms":35893,"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":"This paper claims that reconstructing a 3D velocity field from Wi-Fi Doppler projections and uniformly sampling it over a sphere — a Doppler radiance field (DoRF) — yields a view-invariant motion representation that improves cross-user…","keywords":["Wi-Fi sensing","human activity recognition","channel state information","Doppler velocity","neural radiance fields","Doppler radiance field","cross-user generalization","multipath"],"falsifier":"Place a motion-capture system beside the Wi-Fi setup and compare the recovered center-of-mass velocity v(s) with the ground-truth hand or body velocity during each gesture. If the factorization is correct, the two trajectories should align up to a fixed rotation; if the DoRF accuracy gain over MORIC disappears when the room's reflecting furniture is moved during a trial, that would indicate the time-invariant direction assumption is the source of the gain.","tokens_in":7027,"feed_emoji":"📡","tokens_out":4029,"duration_ms":40850,"temperature":0.7,"pith_summary":"This paper tries to establish that Wi-Fi signals alone can support a view-invariant, three-dimensional description of human motion, and that this description improves cross-user activity recognition. The authors reconstruct a 3D velocity field from one-dimensional Doppler projections extracted from Wi-Fi channel state information, then re-sample that field onto a uniform grid of directions over a sphere, calling the result a Doppler radiance field (DoRF). On a four-gesture, six-person hand-motion dataset, they report leave-one-subject-out accuracy of 54.8% with standard deviation 7.2%, beating the MORIC baseline at 51.5% and all compared methods. The significance would be that Wi-Fi sensing could become more reliable for practical deployment across new users without retraining.","feed_headline":"Doppler radiance fields lift Wi-Fi gesture recognition to 54.8%","feed_subtitle":"Reconstructing a 3D velocity field from Wi-Fi Doppler shifts beats prior baselines at recognizing gestures from unseen users.","key_machinery":"The load-bearing object is the Doppler radiance field P(s, m, n) = v(s)^T d_mn, where v(s) is a reconstructed 3D velocity vector and d_mn are uniformly spaced unit directions on the sphere. The DoRF is obtained by solving a non-convex factorization Vr = V R^T + N via alternating least squares with unit-norm direction constraints, Procrustes alignment, and a DTW-based loss, then projecting each recovered velocity vector onto the full spherical grid. This converts an unordered, sparse set of one-dimensional Doppler observations into a dense, ordered, view-invariant tensor that is fed to the MORIC classifier, which applies random convolution kernels and max-pooling across the directional axis.","core_discovery":"The central claim is that a uniform Doppler radiance field, built by estimating the 3D center-of-mass velocity of motion from sparse Doppler projections and sampling it over a sphere, is a comprehensive and view-invariant motion representation. Rather than relying on whatever multipath reflection directions happen to be present, the method reconstructs a latent velocity trajectory V and a set of observation directions R from the observed projections Vr, then computes radial projections onto a fixed, dense spherical grid. Because the representation does not depend on the random, time-varying viewing geometry of the Wi-Fi access points, the authors argue it generalizes better to unseen users than representations built directly from raw Doppler projections. The reported 54.8% mean accuracy on UTHAMO, with lower variance than MORIC's 8.7%, is offered as evidence.","pith_inferences":["If the recovered velocity v(s) truly approximates the center-of-mass velocity, DoRF could serve as a bridge between Wi-Fi CSI and body-joint or hand-trajectory estimation, not just gesture classification.","The time-invariant direction assumption is the natural failure point: a natural extension is a time-varying or smoothly evolving set of directions R, which would test whether the remaining accuracy gap is caused by this approximation.","The uniform sampling grid size M acts as a resolution knob; one could test whether accuracy saturates as M increases and whether adaptive, information-driven grids perform better.","Applying DoRF to larger, multi-environment datasets with more subjects would clarify whether the 3.3-point gain holds beyond the single office setting and six users of UTHAMO."],"forward_implications":["Cross-user generalization of Wi-Fi human activity recognition improves, reducing the need for per-user calibration or retraining before deployment.","The same DoRF pipeline can fuse data from multiple antennas or access points by constructing a field per antenna and merging them before classification, making it robust to AP configuration.","Because DoRF encodes directional motion, it may support finer-grained tasks than classification, such as motion reconstruction or velocity estimation of the moving body.","The representation is designed to be insensitive to environmental multipath geometry, since the spherical sampling removes dependence on actual reflection directions.","The reported gain of 3.3 percentage points over MORIC, with a lower standard deviation (7.2% vs. 8.7%), suggests more stable performance across different users."],"supporting_citations":[{"why":"Supplies the MORIC baseline and classifier architecture that DoRF feeds into, and the performance numbers DoRF must beat.","marker":"[14]"},{"why":"Provides the neural radiance field concept that motivates reconstructing a 3D representation from sparse 1D projections.","marker":"[15]"},{"why":"Supplies the CSI phase sanitization method that removes hardware distortion before Doppler extraction.","marker":"[16]"},{"why":"Provides the Procrustes alignment step used to stabilize the orientation of the recovered velocity basis.","marker":"[17]"},{"why":"Supplies the UTHAMO dataset and experimental protocol used for the four-gesture, six-subject evaluation.","marker":"[18]"},{"why":"Provides the Nexmon toolkit used to acquire the CSI measurements on commodity Wi-Fi hardware.","marker":"[19]"}],"fun_headline_variants":["Wi-Fi senses motion in 3D via Doppler radiance fields","Doppler radiance fields hit 54.8% on Wi-Fi activity recognition","Uniform Doppler radiance field boosts Wi-Fi activity recognition","From Doppler shifts to 3D motion field for robust Wi-Fi sensing","Wi-Fi activity recognition gets a NeRF-inspired lift"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument assumes that the multipath reflection directions stay fixed while the person moves, so that a single set of directions can explain all Doppler projections; in real rooms the directions shift with the body, and if that shift matters the reconstructed velocity field is distorted.","fun_headline_variants_meta":{"raw":{"variants":["Wi-Fi senses motion in 3D via Doppler radiance fields","Doppler radiance fields hit 54.8% on Wi-Fi activity recognition","Uniform Doppler radiance field boosts Wi-Fi activity recognition","From Doppler shifts to 3D motion field for robust Wi-Fi sensing","Wi-Fi activity recognition gets a NeRF-inspired lift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001211,"raw_usage":{"total_tokens":4950,"prompt_tokens":875,"completion_tokens":4075,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":3986}},"tokens_in":491,"tokens_out":4075,"duration_ms":30523,"temperature":1.0,"reasoning_tokens":3986,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:52:24.735406+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place a motion-capture system beside the Wi-Fi setup and compare the recovered center-of-mass velocity v(s) with the ground-truth hand or body velocity during each gesture. If the factorization is correct, the two trajectories should align up to a fixed rotation; if the DoRF accuracy gain over MORIC disappears when the room's reflecting furniture is moved during a trial, that would indicate the time-invariant direction assumption is the source of the gain.","supporting_citations":[{"cited_title":"Wifi csi-based device-free sensing: from fresnel zone model to csi-ratio model,","cited_arxiv_id":null,"evidence_quote":"Supplies the MORIC baseline and classifier architecture that DoRF feeds into, and the performance numbers DoRF must beat."},{"cited_title":"Fewsense, towards a scalable and cross-domain wi-fi sensing system using few-shot learning,","cited_arxiv_id":null,"evidence_quote":"Provides the neural radiance field concept that motivates reconstructing a 3D representation from sparse 1D projections."},{"cited_title":"Exposing data leakage in wi-fi csi-based human action recognition: A critical analysis,","cited_arxiv_id":null,"evidence_quote":"Supplies the CSI phase sanitization method that removes hardware distortion before Doppler extraction."},{"cited_title":"Sharp: Environment and person independent activity recognition with commod- ity ieee 802.11 access points,","cited_arxiv_id":null,"evidence_quote":"Provides the Procrustes alignment step used to stabilize the orientation of the recovered velocity basis."},{"cited_title":"Widar3. 0: Zero-effort cross-domain gesture recognition with wi-fi,","cited_arxiv_id":null,"evidence_quote":"Supplies the UTHAMO dataset and experimental protocol used for the four-gesture, six-subject evaluation."},{"cited_title":"MORIC: CSI Delay-Doppler Decomposition for Robust Wi-Fi-based Human Activity Recognition","cited_arxiv_id":"2506.12997","evidence_quote":"Provides the Nexmon toolkit used to acquire the CSI measurements on commodity Wi-Fi hardware."}],"review_version":1}