{"id":"6bb68554-f991-418b-85f2-355e0c6823e4","arxiv_id":"1908.07370","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MuDLoc combines amplitude and phase Wi-Fi channel features from multiple access points with a discriminative multi-view subspace projection, and reports mean localization errors of 0.24 m in a lab and 0.15 m in a corridor.","lead":"This paper proposes MuDLoc, a Wi-Fi based indoor localization system that tracks a person without requiring them to carry a device, by combining amplitude and phase measurements from several access points. A smart generalist might read it because the method reports sub-25-centimeter localization errors in cluttered rooms using off-the-shelf Wi-Fi hardware, which could matter for elder care, security, and smart building services.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The outperformance claim is unsupported because the baselines are not given the same bimodal multi-AP features as MuDLoc, so the reported gains may come from the richer input rather than the GI2DCA formulation.","rationale":"After reading the full manuscript, the strongest claim is indeed the empirical outperformance stated in the abstract and Section 5.2. The mathematics of GI2DCA is a standard extension of GMA with a DCCA-style cross-view term; I did not find a fatal flaw in the derivation, and the generalized eigenproblem formulation is internally consistent. The load-bearing weakness is the evaluation protocol. Section 5.2 explicitly acknowledges that the baselines use only amplitude or RSS and do not jointly exploit multi-AP features, while MuDLoc fuses amplitude and phase from multiple APs. This confound is decisive: the reported accuracy gain cannot be assigned to the GI2DCA formulation without an ablation that holds the feature representation fixed. The ambiguity over whether the laboratory results used 5 APs (Section 5.1) or 3 APs (Section 5.5) further weakens the control. A single controlled re-run of MCCA and PWCCA on the same bimodal multi-AP features would settle the question. Because the reader's weakest_assumption already identifies this same concern, I agree with the CONDITIONAL verdict and see no reason to change it.","tokens_in":17645,"tokens_out":9374,"duration_ms":88914,"concrete_test":"Re-run PWCCA and MCCA with the exact feature set used by MuDLoc: for each of the 3 APs, use both the amplitude feature image and the phase-difference feature image as input views, with the same training/validation/test split (6:2:2) and the same beta selection procedure. If MCCA's mean error in the laboratory scenario falls from 0.7032 m toward MuDLoc's 0.2449 m (or PWCCA's from 0.8665 m accordingly), then the reported outperformance reflects the richer input rather than the GI2DCA method. If the errors remain above 0.5 m, the discriminative objective contributes independently.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.2 compares MuDLoc against PWCCA, MCCA, Pilot, and PC-DfL, and concludes a 65% improvement over MCCA. The same paragraph states that \"all other methods use only the amplitude feature of CSI or RSS value\" and are \"either designed to work with single AP or consider the average value for multiple APs.\" This means the comparison simultaneously varies the recognition algorithm and the input representation: MuDLoc uses amplitude plus phase-difference features from multiple APs, while the baselines use amplitude-only or RSS from one AP or from averaged AP values. Consequently, the reported mean errors (0.2449 m in the laboratory, 0.1500 m in the corridor) cannot be attributed to the GI2DCA discriminant objective; they may be substantially due to the additional phase information and multi-AP diversity, which Section 5.3 shows improve MuDLoc's own performance. The AP count is also ambiguous: Section 5.1 places 5 APs in the laboratory, but Section 5.5 states the system \"considers using 3 AP,\" leaving unclear which configuration produced Tables 1 and 2. A controlled comparison, in which PWCCA and MCCA receive the same bimodal multi-AP features and the same AP count, is needed before the central claim is supported. No code or data artifacts are provided to verify this.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MuDLoc, a device-free indoor localization system that uses CSI amplitude and phase-difference features collected from multiple access points. The core algorithmic contribution is GI2DCA, which extends generalized multiview analysis (GMA) by adding an inter-view discriminant term based on DCCA, yielding a generalized eigenvalue problem solved for view-specific linear projections. Offline, these projections map multi-view amplitude and phase features to a common space whose averaged canonical variates are stacked into a Multi-view Discriminant Feature Image; online, a test sample is matched to the nearest training cell centroid by Euclidean distance. Experiments in a laboratory and a corridor compare MuDLoc with PC-DfL, Pilot, PWCCA, and MCCA, reporting mean distance errors of 0.2449 m and 0.1500 m and claiming about 65% improvement over MCCA.","tokens_in":17928,"tokens_out":6300,"duration_ms":56173,"significance":"If the reported accuracy is reproducible, sub-0.25 m device-free localization with commodity 802.11n hardware would be a practically valuable advance, and the paper's combination of inter-view and intra-view discriminant structure is a natural and readable extension of existing CCA/GMA machinery. The derivation of the generalized eigenvalue problem is standard and appears internally consistent, and the authors include useful ablations on feature modality and AP count. However, the central outperformance claim is currently supported only by comparisons in which MuDLoc differs from every baseline in both the recognition algorithm and the input features (bimodal multi-AP CSI vs. amplitude-only or RSS from a single or averaged AP), so the magnitude of the claimed gain cannot be attributed to GI2DCA. The absence of significance tests, confidence intervals, and released code or data further limits the strength of the empirical claims.","major_comments":[{"comment":"The benchmark comparison does not control for input features. The text in §5.2 states that 'all other methods use only the amplitude feature of CSI or RSS value' and that these methods 'are either designed to work with single AP or consider the average value for multiple APs,' whereas MuDLoc uses amplitude plus phase-difference features from multiple APs. Therefore the reported gains (e.g., 0.2449 m vs. 0.7032 m for MCCA in the laboratory) could result largely from the richer bimodal multi-AP representation rather than from the GI2DCA objective. To support the central claim that the GI2DCA formulation outperforms benchmark approaches, the authors should re-run PWCCA, MCCA, and (where possible) Pilot and PC-DfL on the same bimodal multi-AP feature set and the same AP count, or add an ablation in which GI2DCA is restricted to the same features as each baseline.","section":"§5.2, Tables 1–2"},{"comment":"The number of APs used for the headline results is ambiguous. Section 5.1 describes the laboratory experiment with 5 APs and the corridor experiment with 3 APs, yet Section 5.5 concludes that 'this work considers using 3 AP' for both deployments, and Fig. 12 shows results for 2–5 APs. It is therefore unclear whether Tables 1 and 2 report 5-AP laboratory results, 3-AP results, or a mix; this ambiguity matters because Fig. 12 shows that AP count changes mean error (e.g., the laboratory error decreases as APs increase from 2 to 5). The authors should state explicitly which AP count produced each reported table and figure, and should present the main comparison consistently for the selected configuration.","section":"§5.1, §5.5, Tables 1–2"},{"comment":"The statistical evidence for the headline comparison is incomplete. Section 5.1 says 10 independent measurements were taken on 10 different days and that 'the mean value' was used for performance evaluation, but Tables 1–2 report only a single mean and standard deviation per method, without stating whether the standard deviation is over test locations, over the 10 days, or over some other partition. No confidence intervals or significance tests are provided, so the reader cannot judge whether the reported gaps (e.g., 0.1500 m vs. 0.6888 m in the corridor) are stable across days or runs. The authors should report per-day or per-run errors and provide a paired significance test across the 10 daily measurements, or otherwise justify that the reported differences exceed experimental variability.","section":"§5.1, Tables 1–2"}],"minor_comments":[{"comment":"The notation for the number of views is inconsistent: the optimization formulations in Eqs. (13) and (20) use N while the surrounding text and Eq. (26) use M; please unify this notation.","section":"§4.1.2, Eqs. (13) and (20)"},{"comment":"The statement that 'the constraints are coupled with γ = trace ratio' is unclear; please specify how γ_i are chosen and what 'trace ratio' means in the constraint of Eq. (20).","section":"§4.1.2, Eq. (20)"},{"comment":"The construction of the phase-difference feature image Y_i is underspecified: the exact dimension d_{Y_i} and the arrangement of pairwise phase differences across subcarriers and antenna pairs into rows and columns are not described, which would hinder reproducibility.","section":"§4.1.1"},{"comment":"There are several typographical errors, including 'Morover' (§5.2), 'Simialrly' (§4.2), 'Distannce' (Fig. 9 caption), 'architechture' (§4), and 'discriminnat' (§5.2); a careful proofread is needed.","section":"Throughout"},{"comment":"The claim that MuDLoc is 'the first multi-view discriminant learning approach' for device-free localization is stronger than the related-work discussion supports, since GMA [33] already performs multi-view discriminant analysis; please soften the claim and explicitly contrast MuDLoc with the authors' earlier DCCA-based method [17].","section":"Abstract and §1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the proposed GI2DCA objective in Eq. (22) is a straightforward fusion of the GMA objective with the DCCA class-block weighting, and the paper derives the generalized eigenvalue solution cleanly. The application to multi-AP CSI with amplitude and phase-difference features is legitimate, and the experimental setup (10 days, 3000 samples per cell, cross-validated hyperparameters) is more careful than much of the CSI localization literature. But the paper's central claim—that it beats the benchmarks—rests on a comparison that changes two things at once. Section 5.2 explicitly says the baselines get only amplitude or RSS, and either one AP or averaged APs, while MuDLoc gets amplitude plus phase difference from multiple APs. So the 65% improvement over MCCA could come from the richer input representation, not from the GI2DCA discriminant term. That's not a small caveat; it directly hits the main conclusion. The AP count is also muddy: Section 5.1 says 5 APs are placed in the lab, while Section 5.5 says the system settles on 3 APs, and the tables don't say which configuration was used. Since Fig. 12 shows error dropping with more APs, that ambiguity matters.\n\nThe math itself looks fine. The constraints are handled with the same relaxation as GMA, and the resulting generalized eigenvalue problem is standard. The complexity claim O(d^3 M) is reasonable. No circularity: test cells are held out and hyperparameters are chosen by cross-validation. The paper is honest about the phase randomness and motivates the phase-difference feature well.\n\nWhat's missing: a controlled experiment where PWCCA and MCCA receive exactly the same bimodal, multi-AP input as MuDLoc, plus a clear statement of the AP count used for each table. Significance tests or confidence intervals would help, but the controlled comparison is the essential fix. Code or data would be even better.\n\nThis is not a paper to reject outright; it's a paper that needs a solid revision. A good referee could push the authors to re-run the baselines with matching features, and if the numbers still hold, the contribution would be real. I'd send it to peer review, with the evaluation concern as the main point to resolve.","headline":"A clean derivation of a plausible method, but the empirical comparison changes both algorithm and input features at once, so the headline accuracy numbers are not yet substantiated.","tokens_in":18473,"tokens_out":2937,"would_cite":true,"duration_ms":28768,"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 multi-view discriminant projection over Wi-Fi channel data can locate a person indoors to about 0.15–0.25 m mean error, without requiring the person to carry any device.","keywords":["indoor localization","device-free localization","channel state information","multi-view learning","discriminant correlation analysis","CSI phase difference","Wi-Fi fingerprinting","MIMO-OFDM"],"falsifier":"Re-run the comparison with MCCA and PWCCA receiving exactly the same features MuDLoc uses, namely amplitude and phase-difference from the same 3 APs, and check whether MuDLoc still shows a 65% mean-error improvement; if the gap nearly disappears, the claim that inter-view and intra-view discriminant structure drives the accuracy is falsified.","tokens_in":17450,"feed_emoji":"📍","tokens_out":6792,"duration_ms":63049,"temperature":0.7,"pith_summary":"The paper proposes MuDLoc, a device-free indoor localization method that locates a person by matching Wi-Fi channel measurements taken when the person stands in a known cell. It claims that combining two CSI modalities (amplitude and phase) from several access points and projecting them into one shared discriminant space makes the pattern match accurate enough for sub-0.25 m mean distance error in a cluttered lab and 0.15 m in a corridor. The core idea is that each access point is a separate view of the same location, and a joint projection that preserves both within-view class separation and between-view class association can expose the location-specific common structure. If the claim holds, commodity Wi-Fi without any device on the subject can compete with device-based localization.","feed_headline":"Wi-Fi locates a person to 15 cm without a wearable","feed_subtitle":"Fusing both Wi-Fi signal properties from several access points locates a person with no device on them.","key_machinery":"The load-bearing object is the GI2DCA optimization, a generalized eigenvalue problem that maximizes a weighted sum of intra-view between-class scatter (from LDA-style terms) and inter-view discriminant correlation (a DCCA-style term with the class-block matrix G), subject to a trace-ratio constraint over within-view covariance. It generalizes MCCA by injecting class labels across views and generalizes GMA by adding inter-view class association. Solving it yields one projection per access point; projections are averaged per modality, and the amplitude and phase results are stacked to form the final feature image used with Euclidean-distance matching. The paper notes the cost is $O(d^3 M)$ for $d$ the largest feature dimension and $M$ views.","core_discovery":"On the paper's own terms, the discovery is that a generalized discriminant correlation analysis over multi-view CSI, called GI2DCA, produces a common feature space in which different access points' amplitude and phase features agree for the same cell and differ across cells. The method stacks per-view projected amplitude features and per-view projected phase features into a Multi-view Discriminant Feature Image, then labels a test point by nearest Euclidean distance to the training cell average. In the reported experiments this yields mean distance errors of 0.2449 m in the laboratory and 0.1500 m in the corridor, with roughly 90% of test locations within 1 m, beating RSS-based and single-AP CSI baselines. The paper attributes the gain to exploiting inter-view and intra-view class structure jointly, not just correlations.","pith_inferences":["An extension beyond the paper: because the baselines were evaluated with amplitude-only or RSS-only features, a controlled ablation feeding MCCA and PWCCA the same amplitude-plus-phase multi-AP features is needed to isolate whether GI2DCA's inter-view and intra-view terms, rather than the extra input information, produce the reported gap.","An extension beyond the paper: the phase-difference stabilisation relies on the receiver antennas sharing a clock, a property the paper observes on Intel 5300 hardware; re-testing on other NICs and antenna spacings would show whether sub-0.25 m accuracy is portable.","An extension beyond the paper: the same joint projection could be tested on other device-free sensing tasks such as fall detection, gait recognition, or occupancy counting, where multiple access points and dual CSI modalities form natural views."],"forward_implications":["A device-free system can reach mean errors of 0.2449 m in a cluttered lab and 0.1500 m in a corridor using 3 to 5 commodity Wi-Fi access points.","Adding more access points improves accuracy, but the marginal gain beyond 3 APs is small, so the paper recommends 3 APs as the cost-accuracy trade-off point.","Using amplitude and phase together is better than either alone; the reported bimodal fusion reaches the lowest errors.","The multi-view discriminant projection outperforms MCCA and pairwise CCA on the same raw dataset, with a reported 65% improvement in mean distance error over MCCA.","Increasing the number of online CSI packets from 300 to 600 reduces error only mildly, so 300 packets are treated as sufficient for the test phase."],"supporting_citations":[{"why":"Supplies canonical correlation analysis, the foundation the paper extends to a multi-view discriminant setting.","marker":"[25]"},{"why":"Defines SUMCOR multi-view CCA, the unsupervised baseline that GI2DCA generalizes and compares against.","marker":"[30]"},{"why":"Provides the generalized multiview analysis framework with intra-view discriminant terms whose objective GI2DCA modifies to add inter-view class association.","marker":"[33]"},{"why":"Introduces discriminant CCA with the class-block matrix G that GI2DCA reuses for between-view class correlation.","marker":"[39]"},{"why":"Shows adjacent-antenna phase differences are stable on commodity Wi-Fi, justifying the phase feature.","marker":"[34]"},{"why":"Supplies the Linux 802.11n CSI measurement tool used to collect the experimental data.","marker":"[10]"},{"why":"Provides the RSS-based device-free localization baseline (PC-DfL) that MuDLoc is compared against.","marker":"[6]"},{"why":"Provides the single-AP CSI amplitude baseline (Pilot) used for comparison.","marker":"[18]"}],"fun_headline_variants":["Wi-Fi CSI fusion pinpoints you to 15 cm","Device-free localization: Wi-Fi pinpoints to 15 cm","No wearable needed: Wi-Fi locates you to 15 cm","Multi-view Wi-Fi signals locate you to 15 cm","AI uses Wi-Fi to find you within 15 cm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experiment assumes the benchmark methods are compared on equal footing, so the baselines use the same bimodal, multi-AP features and only the discriminant projection differs; if that is not the case, the reported accuracy gain could come from the richer input rather than from GI2DCA.","fun_headline_variants_meta":{"raw":{"variants":["Wi-Fi CSI fusion pinpoints you to 15 cm","Device-free localization: Wi-Fi pinpoints to 15 cm","No wearable needed: Wi-Fi locates you to 15 cm","Multi-view Wi-Fi signals locate you to 15 cm","AI uses Wi-Fi to find you within 15 cm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000569,"raw_usage":{"total_tokens":2686,"prompt_tokens":933,"completion_tokens":1753,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":1669}},"tokens_in":549,"tokens_out":1753,"duration_ms":13354,"temperature":1.0,"reasoning_tokens":1669,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:32:17.916792+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison with MCCA and PWCCA receiving exactly the same features MuDLoc uses, namely amplitude and phase-difference from the same 3 APs, and check whether MuDLoc still shows a 65% mean-error improvement; if the gap nearly disappears, the claim that inter-view and intra-view discriminant structure drives the accuracy is falsified.","supporting_citations":[{"cited_title":"Relations between two sets of variates,","cited_arxiv_id":null,"evidence_quote":"Supplies canonical correlation analysis, the foundation the paper extends to a multi-view discriminant setting."},{"cited_title":"Multi-view canonical corre- lation analysis,","cited_arxiv_id":null,"evidence_quote":"Defines SUMCOR multi-view CCA, the unsupervised baseline that GI2DCA generalizes and compares against."},{"cited_title":"Generalized multiview analysis: A discrimi- native latent space,","cited_arxiv_id":null,"evidence_quote":"Provides the generalized multiview analysis framework with intra-view discriminant terms whose objective GI2DCA modifies to add inter-view class association."},{"cited_title":"A novel method of combined feature extraction for recognition,","cited_arxiv_id":null,"evidence_quote":"Introduces discriminant CCA with the class-block matrix G that GI2DCA reuses for between-view class correlation."},{"cited_title":"Phaser: Enabling phased array signal processing on commodity wifi access points,","cited_arxiv_id":null,"evidence_quote":"Shows adjacent-antenna phase differences are stable on commodity Wi-Fi, justifying the phase feature."},{"cited_title":"Predictable 802.11 packet delivery from wireless channel mea- surements,","cited_arxiv_id":null,"evidence_quote":"Supplies the Linux 802.11n CSI measurement tool used to collect the experimental data."},{"cited_title":"The case for efficient and robust rf-based device-free localization,","cited_arxiv_id":null,"evidence_quote":"Provides the RSS-based device-free localization baseline (PC-DfL) that MuDLoc is compared against."},{"cited_title":"Pilot: Passive device- free indoor localization using channel state information,","cited_arxiv_id":null,"evidence_quote":"Provides the single-AP CSI amplitude baseline (Pilot) used for comparison."}],"review_version":1}