{"id":"7e2c7683-d710-4d3a-94ef-62787cbf20c1","arxiv_id":"2412.13601","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"Turns Wi-Fi channel measurements into a map image and uses a CNN-LSTM plus particle filter to locate pedestrians with sub-meter claimed accuracy.","lead":"Wi-Fi signals can be turned into image-like maps and fed to a neural network to track a person walking indoors, with claimed errors below half a meter. The idea is attractive, but the paper's experiments are inconsistent and the key numbers cannot be tied to the reported tables.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline RMSE figures (0.36 m dynamic, 0.17 m static) are not reproducible from the reported experiments; no static test protocol is described and no table row or aggregation yields these values.","rationale":"I agree with the reader's rejection of the paper, but I identify a more basic and more load-bearing problem than the temporal CSI-map ordering. The central claim is the specific accuracy improvement stated in the abstract, yet the experimental record does not contain the experiments that would produce those numbers. Table 3 reports per-configuration mean errors before and after PF-HSM; 0.36 m and 0.17 m do not appear, no static protocol is described, and the paper does not explain how the abstract's averages were computed. The temporal-map concern is also serious and may explain why the methodology is questionable, but the missing numbers are sufficient on their own to invalidate the central claim. Because the reader's weakest assumption was about the temporal ordering rather than the unreported aggregate results, my agreement is partial: both concerns undermine the paper, but the missing numbers are the most decisive.","tokens_in":12832,"tokens_out":12722,"duration_ms":111538,"concrete_test":"Paper-only check: parse Table 3 into its twelve parameter rows and compute the mean of the after-PF-HSM column. Then identify, from the text, which rows correspond to the 'moderately dynamic' and 'static' conditions named in the abstract. With the current text, no static test protocol exists and no row-to-condition mapping is given, so the check stops at 'not reproducible'. To resolve the concern fully, the authors would need to supply the raw test logs and the exact aggregation rule that yields 0.36 m and 0.17 m; recomputing the RMSE from those logs would either reproduce the abstract numbers or show that the abstract is inconsistent with the reported experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the abstract's average RMSE of 0.36 m in a moderately dynamic and 0.17 m in a static environment. The results section reports only a 'moderately dynamic' lunch-hour test: Table 2 lists mean errors of 2.20, 1.99, 1.43, and 0.79 m for LSTM, DeepFi, ConFi, and the CNN-LSTM classifier, and Table 3 lists per-configuration errors before and after particle-filter HSM that range from 0.07 to 0.75 m after PF. The values 0.36 and 0.17 appear in no table row, no static test protocol is described anywhere, and the paper gives no mapping from the Table 3 rows to the abstract's 'moderately dynamic' and 'static' conditions. The simple mean of the twelve reported after-PF entries is about 0.46 m, not 0.36 m; 0.17 m is not among the reported values. Even the stated 38% improvement from PF would take the Table 2 CNN-LSTM value of 0.79 m to about 0.49 m, not 0.36 m. Additionally, the abstract claims RMSE while the results report mean distance error, so the headline metric is not even the same as the measured quantity. The headline accuracy is therefore not derivable from the paper's own experimental record. This is the load-bearing gap: a reader cannot verify that the claimed improvement exists, independent of whether the temporal CSI-map construction is internally consistent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Wi-Fi CSI fingerprinting localization system built around a \"CSI Fingerprint Map\" (a 2D+channel image formed by re-indexing CSI phase observations by map location), a hybrid CNN-LSTM classifier, and a particle-filter hypothesis-selection module. The authors report comparisons against DeepFi, ConFi, and a temporal LSTM baseline, and claim average RMSE of 0.36 m in a moderately dynamic environment and 0.17 m in a static environment.","tokens_in":13143,"tokens_out":4225,"duration_ms":45677,"significance":"If the claimed sub-meter accuracy were supported by the experiments, this would be a meaningful step toward fine-grained Wi-Fi localization with sparse infrastructure. The paper contributes a two-month dataset, a spatial-map representation, and an evaluation against established baselines. However, the authors do not provide code or data, and as detailed below the headline results are not reproducible from the reported experiments, which substantially limits the significance of the contribution.","major_comments":[{"comment":"The abstract's RMSE values (0.36 m moderately dynamic, 0.17 m static) are not reproducible from the experimental record. Table 3 reports mean distance error before and after PF-HSM for 12 parameter combinations, with after-PF values from 0.07 m to 0.75 m; no row or aggregation yields 0.36 m, 0.17 m is not among the reported values, and no static test protocol is described anywhere in §5. The paper also alternates between 'mean error' (Tables 2 and 3) and 'RMSE' (abstract), so the headline metric is not the measured quantity. The stated 38% improvement would turn the Table 2 CNN-LSTM value of 0.79 m into approximately 0.49 m, not 0.36 m. The central claim of the paper is therefore not derivable from the paper's own experimental results.","section":"Abstract and §5.4 / Table 3"},{"comment":"The 'time-ordered' property of the CSI Fingerprint Maps is a construction, not a physical temporal sequence. A single device cannot collect simultaneous observations at 36 locations; the text itself acknowledges that 'one must gather observations at each location at different times.' The re-indexing that creates M_i therefore does not provide a real temporal step between M_i and M_{i+1}, and the LSTM's temporal modeling lacks a physical basis. The paper must either demonstrate that this re-indexing preserves meaningful temporal structure or abandon the LSTM temporal claim.","section":"§4.2.1, Eq. (3)"},{"comment":"Configuration selection is performed on the reported test results. The best after-PF error (0.07 m) is achieved for grid size 1 m, 3×3 sliding window, 1 m/s, while other configurations give errors up to 0.75 m; no a priori selection rule or cross-validation procedure is stated. Reporting the minimum of a grid search over grid size, window size, and walking speed as the method's accuracy is circular unless the evaluation protocol separates model selection from testing.","section":"Table 3"},{"comment":"The paper claims in the conclusion that the method 'delivers estimated pedestrian heading without the use of any additional sensors,' but the heading accuracy evaluation in §5.4 is performed only with a simulated human walk motion model, not with real walking data. No heading ground truth from real pedestrian trajectories is presented. This overclaim is not supported by the reported experiments.","section":"§5.4 and §6"}],"minor_comments":[{"comment":"The particle-filter measurement model equations contain apparent errors: d0 uses (rx - lx) for both the cos and sin terms, omitting (ry - ly); the normalization factor 1/(√(2π)(1 - c)) in p_L and p_R is not derived. These issues hinder reproducibility.","section":"§4.2.4"},{"comment":"The manuscript has numerous presentation issues: Table 2 is referred to as Table 1 in the text ('The mean errors of all compared methods is presented in Table 1'), Figure 3 is mislabeled as 'Figure 3 Trajectory to grid-cell mapping,' equations are unnumbered, and the formatting of Table 3 is garbled.","section":"Throughout"},{"comment":"The methodology does not state how many test trajectories were collected, their lengths, or how the reported mean errors are aggregated over trajectories and parameter combinations. Without this information, the comparison in Table 2 cannot be independently reproduced.","section":"§5.2 and §5.3"},{"comment":"The denoising rule (removing observations with CSI phase values beyond 2σ for any subcarrier) is described only qualitatively; the fraction of removed observations, their spatial/temporal distribution, and the effect on the resulting fingerprints are not reported, so the impact of denoising on the final accuracy cannot be assessed.","section":"§3.3"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early draft with substantial internal inconsistencies (abstract vs. tables, metric mismatch, mislabeled figures). The gap between the abstract's headline accuracy and the experimental results is not a presentation issue; it would require new experiments and a clear evaluation protocol. The time-ordered map construction is acknowledged by the authors as a re-indexing, which undermines the core justification for using an LSTM. These concerns are load-bearing for the paper's main claim, so rejection is recommended, though the underlying idea of map-based CSI representation may be worth pursuing with a rigorous evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: the map-based CSI fingerprint idea is genuinely new relative to DeepFi and ConFi, but the paper's headline numbers (0.36 m dynamic, 0.17 m static) do not follow from its own tables. Table 3 reports mean distance error before/after particle filtering for various grid sizes, window sizes, and speeds; no row or aggregation gives 0.36 or 0.17. The abstract also says RMSE while the results report mean distance error, and no static experiment is described anywhere. The 38% improvement from the particle filter is based on simulated trajectories, not the real test set.\n\nThe time-ordered map construction is the other trouble spot. The paper says all pixels in M_i are taken one time step before M_{i+1}, but the data were collected by walking to 36 locations and taking measurements at each. Those measurements are spread over minutes, not one time instant, so the \"temporal sequence\" is a construction. The LSTM's temporal modeling therefore has no physical grounding unless the authors can show the ordering preserves something real about the channel.\n\nWhat the paper does well: the representation of CSI as a 2D map aligned to physical coordinates is a real idea, and the paper actually collects a 44-day dataset, which is more than most fingerprinting papers do. The comparison against DeepFi, ConFi, and an LSTM baseline is reasonable, and the sliding-window notion is explained clearly. The particle filter model itself follows prior work and is not the issue.\n\nThe softer spots: there are no error bars, no code or data release, and a grid search over grid size, window size, speed, and regularization without a stated selection rule. That makes the claimed accuracy a fitted value. The particle filter validation on simulated data is fine as a sanity check, but citing it as evidence for the real system's 38% improvement is not.\n\nOverall: this is a proof-of-concept with a plausible idea and sloppy reporting. It should not be accepted as is, but I'd send it to reviewers—the core idea deserves scrutiny and a request for a clean experiment. A serious referee would need to see the real test protocol, a static test, error bars, and a stated model selection procedure.","headline":"The map-based CSI fingerprint idea is new, but the headline accuracy numbers are not reproducible from the paper's own experiments.","tokens_in":13717,"tokens_out":2803,"would_cite":false,"duration_ms":24204,"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 spatial Wi-Fi phase maps, a hybrid CNN-LSTM, and a particle filter localize pedestrians to 0.36 m in dynamic settings and 0.17 m in static ones.","keywords":["indoor localization","Wi-Fi fingerprinting","channel state information (CSI)","CSI phase sanitization","hybrid CNN-LSTM","particle filter","pedestrian trajectory","OFDM subcarriers"],"falsifier":"One decisive check is to train the same model on the same CSI Fingerprint Maps but with the index order of the maps randomly shuffled; if the localization error stays the same, the LSTM is not using temporal structure and the central mechanism is a spatial prior. A corroborating check is to rerun the evaluation on longer test walks measured against independent motion-capture ground truth instead of sequential walk labels.","tokens_in":12561,"feed_emoji":"📍","tokens_out":8981,"duration_ms":81036,"temperature":0.7,"pith_summary":"This paper is trying to establish that ordinary Wi-Fi channel measurements, which usually fluctuate too much for reliable positioning, can be rearranged into a spatial image and learned from as a sequence. The image, called a CSI Fingerprint Map, places 270 sanitized phase values from a single observation at the pixel corresponding to its physical location, and the hybrid convolutional-and-recurrent network learns both spatial fingerprint structure and short-term temporal dependencies between consecutive map slices. A particle filter then enforces a human walk model by generating many trajectory hypotheses and keeping the one a real pedestrian could have walked. On a 140 square meter corridor with one access point and one receiver, the reported average errors are 0.36 meters in a moderately dynamic environment and 0.17 meters in a static one, well below the 1 to 4 meter range typical of earlier Wi-Fi fingerprinting. If this holds, sub-meter indoor pedestrian localization becomes possible with existing Wi-Fi infrastructure and without cameras, IMUs, or magnetometers.","feed_headline":"Wi-Fi maps localize pedestrians to 0.36 meters","feed_subtitle":"A hybrid CNN-LSTM reads Wi-Fi phase data as a spatial map, then a particle filter rejects paths that humans cannot walk.","key_machinery":"The central object is the CSI Fingerprint Map: a 2D spatial grid whose pixel at each reference location is a vector of 270 sanitized CSI phase values from 30 OFDM subcarriers across 9 MIMO antenna pairs. Successive maps are indexed so that all pixels in map i are treated as one time step before map i+1, which is what lets the recurrent network see the trajectory as a sequence. The argument is carried by three supporting mechanisms: a phase-sanitization and denoising step that makes fingerprints temporally stable, an Information Adaptive Sliding Window that feeds the CNN spatial patches of increasing neighborhood size, and a particle-filter hypothesis-selection module that enforces a human gait model and discards location sequences that violate realistic walking speed, stride, or heading constraints.","core_discovery":"The central claim is that treating CSI phase observations as pixels in a time-ordered 2D map turns Wi-Fi fingerprinting from a per-location classification problem into a spatial-sequence learning problem. Each map pixel carries 270 sanitized phase values from nine MIMO antenna pairs, and a sliding window of pixels (1x1, 2x2, or 3x3) forms proposals whose size adapts to how much surrounding spatial information is available. A CNN extracts spatial features from these map patches, an LSTM models short-term dependencies between consecutive maps, and the output is a 37-entry confidence vector over grid locations. A multi-hypothesis particle filter, driven by a foot-position human walk model, then weights candidate trajectories and selects the most likely one, reducing mean distance error by about 38 percent on average and down to as low as 0.07 meters for some parameter combinations. The author frames the result as a proof of concept that sparse Wi-Fi observations, one access point, and moderate noise are enough for fine-grained pedestrian localization and heading estimation.","pith_inferences":["An extension the paper leaves implicit is that the same spatial map representation could be trained with RSSI values instead of CSI phase; holding the architecture fixed would isolate how much of the gain comes from the spatial arrangement and how much from the richer phase feature.","A decisive ablation the paper does not report is shuffling the temporal order of the map slices during training; if localization accuracy does not degrade, the LSTM is effectively learning a spatial prior while the claimed temporal dependency is an artifact of the collection design.","Because the data come from a single corridor and a single walk pattern, whether the learned spatial-temporal weights transfer to other floor plans, device placements, and walking styles is an untested consequence of treating fingerprints as maps."],"forward_implications":["The reported error bars put Wi-Fi-only localization in the sub-meter range previously associated with camera-based or sensor-fused systems.","Because the pipeline needs only one access point and one receiver, deployment could piggyback on existing Wi-Fi infrastructure rather than requiring a dense grid of anchors.","The particle-filter step not only removes gross misclassifications but also produces pedestrian heading estimates from Wi-Fi alone, removing the need for IMU or magnetometer input.","The post-filtering accuracy improvement implies that most residual CNN-LSTM errors are implausible walking jumps, which is why enforcing a human walk model can cut mean distance error by roughly 38 percent on average."],"supporting_citations":[{"why":"Supplies the evidence that 5 GHz OFDM CSI phase differences are more stable than RSSI, motivating the choice of phase features for fingerprinting.","marker":"[4]"},{"why":"Provides the 1-4 m typical accuracy range for Wi-Fi positioning and the Gaussian noise-bound reasoning used to justify removing CSI observations beyond 2 sigma.","marker":"[5]"},{"why":"An earlier deep-learning CSI fingerprinting baseline whose reported accuracy the proposed method is compared against.","marker":"[15]"},{"why":"A CNN-based Wi-Fi localization method using CSI feature images, used as the closest comparison baseline.","marker":"[17]"},{"why":"Supplies the linear-transform phase sanitization that removes random CSI phase offsets before fingerprint construction.","marker":"[18]"},{"why":"Provides the foot-position human walk model and particle-filter update equations used by the hypothesis-selection module.","marker":"[20]"},{"why":"Supplies the simulated human walking trajectories used to evaluate particle-filter tracking accuracy against ground truth.","marker":"[21]"}],"fun_headline_variants":["CSI maps plus hybrid deep net hit 0.36 m accuracy","Wi-Fi CSI becomes a map: 0.36 m pedestrian tracking","DL hybrid maps CSI to sub-meter localization","CSI fingerprint maps cut Wi-Fi location error to 0.36 m"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on treating map slices built from observations taken at 36 locations at different times as consecutive moments of one walk; if that artificial time order is not physically meaningful, the temporal modeling that the whole hybrid architecture depends on has nothing real to learn.","fun_headline_variants_meta":{"raw":{"variants":["CSI maps plus hybrid deep net hit 0.36 m accuracy","Wi-Fi CSI becomes a map: 0.36 m pedestrian tracking","DL hybrid maps CSI to sub-meter localization","CSI fingerprint maps cut Wi-Fi location error to 0.36 m"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000428,"raw_usage":{"total_tokens":2215,"prompt_tokens":999,"completion_tokens":1216,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":1144}},"tokens_in":615,"tokens_out":1216,"duration_ms":8696,"temperature":1.0,"reasoning_tokens":1144,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:57:57.376109+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One decisive check is to train the same model on the same CSI Fingerprint Maps but with the index order of the maps randomly shuffled; if the localization error stays the same, the LSTM is not using temporal structure and the central mechanism is a spatial prior. A corroborating check is to rerun the evaluation on longer test walks measured against independent motion-capture ground truth instead of sequential walk labels.","supporting_citations":[{"cited_title":"time-ordered maps","cited_arxiv_id":null,"evidence_quote":"Supplies the evidence that 5 GHz OFDM CSI phase differences are more stable than RSSI, motivating the choice of phase features for fingerprinting."},{"cited_title":"Experiment Setup We use a low-power pocket held embedded PC as a Wi-Fi device for collecting test data set","cited_arxiv_id":null,"evidence_quote":"Provides the 1-4 m typical accuracy range for Wi-Fi positioning and the Gaussian noise-bound reasoning used to justify removing CSI observations beyond 2 sigma."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"An earlier deep-learning CSI fingerprinting baseline whose reported accuracy the proposed method is compared against."},{"cited_title":"Wi -Fi Crowdsourced Fingerprinting Dataset for Indoor Positioning","cited_arxiv_id":null,"evidence_quote":"A CNN-based Wi-Fi localization method using CSI feature images, used as the closest comparison baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the linear-transform phase sanitization that removes random CSI phase offsets before fingerprint construction."},{"cited_title":"Ambiguity function analysis of wireless LAN transmissions for passive radar","cited_arxiv_id":null,"evidence_quote":"Provides the foot-position human walk model and particle-filter update equations used by the hypothesis-selection module."},{"cited_title":"This motion model is easily configurable, and bounds can be set on normally distributed pedestrian walk phase, speed, stride and heading direction","cited_arxiv_id":null,"evidence_quote":"Supplies the simulated human walking trajectories used to evaluate particle-filter tracking accuracy against ground truth."}],"review_version":1}