{"id":"e8562217-2c2f-457b-a4fe-7f91242d379b","arxiv_id":"2507.07331","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"mmFlux estimates a crowd flow field from commodity mmWave radar point clouds, reconstructs a directed graph of splits and merges, and uses divergence and curl to identify gatherings, dispersions, and turns.","lead":"A new radar-based system reads crowd movement without cameras, building a map of how people flow, split, and merge from a single affordable sensor. The authors test it on 21 real crowds of up to 20 people and report accurate flow maps in 17 of 18 structured scenarios.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quantitative support for topology reconstruction is weaker than reported: the one-sided Chamfer distance and metrics computed only on recovered graphs cannot detect missing or spurious edges, and the single failed case is excluded.","rationale":"I read the paper's central claim as: a single commodity mmWave radar can recover aggregate crowd flow topology and key semantic regions without tracking individuals. The pipeline is described in detail, the experimental campaign is real, and the authors openly discuss the one failed C4 case. I do not think the central claim should be rejected. However, the strongest quantitative evidence for it is less secure than the abstract suggests. The reader's weakest assumption concerns the Lucas-Kanade pairwise flow step in Eq. (7), which is a reasonable upstream risk: occlusion, multipath, and missed detections can break point associations. My concern is directly connected to that risk but located one step downstream: even if flow estimation is imperfect, the current evaluation metrics would not reveal a systematic failure that omits or misconnects edges. A one-sided Chamfer distance cannot penalize a missing edge, and the orientation MAE is computed only for edges that exist in the recovered graph. The failed C4 experiment is precisely a missed-edge case, and it is excluded from those metrics. Thus the reported 0.45 m Chamfer and 8.8 degree MAE support the claim that recovered edges are reasonably placed, but they do not support the claim that the full topology is recovered. The concrete test I propose would settle this by measuring topological completeness directly. I partially agree with the reader because the pairwise-flow assumption is plausible as the origin of the problem, but I would place the load-bearing weakness in the evaluation's insensitivity to that failure mode. The verdict stays CONDITIONAL: the paper's conclusions are plausible and worth publishing conditionally, but the quantitative case needs a topology-complete evaluation rather than the current selective, one-sided metrics.","tokens_in":22123,"tokens_out":6864,"duration_ms":90604,"concrete_test":"Re-run the graph evaluation on all 18 structured experiments, including the failed C4 case, and replace the one-sided Chamfer distance with topology-aware metrics: (1) compute the bidirectional Chamfer distance, averaging both directions between estimated and true graph embeddings; (2) compute skeleton precision and recall, counting estimated skeleton pixels within 0.76 m of the true skeleton and true skeleton pixels within 0.76 m of the estimated skeleton; (3) compute graph edit distance or an equivalent vertex/edge match rate that counts missed vertices, missed edges, extra vertices, and extra edges. If recall or missing-edge rates are substantial, or the bidirectional distance degrades materially, the claim of high-fidelity reconstruction should be qualified to partial reconstruction in favorable cases.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on high-fidelity graph reconstruction, but the reported evaluation cannot measure topological completeness. In Sec. 5.2, the one-sided Chamfer distance d_avg = 0.45 m is computed for each estimated point as the distance to the nearest point on the true graph, and the edge orientation MAE is computed only for the 62 edges \"present in the successfully reconstructed graphs.\" A one-sided distance penalizes estimated edges that stray from true edges, but it does not penalize missing true edges at all: a reconstruction that drops an entire branch can still score near zero if the remaining estimated edges lie on true edges. Similarly, orientation error is undefined for missed edges and does not account for spurious or incorrectly connected edges. This is not a hypothetical concern: the one failed experiment in C4 is described in Sec. 6 as a missed edge caused by occlusion and tree clutter, and it is excluded from the Chamfer and orientation statistics. The split-ratio MAE of 0.1 also depends on the recovered graph, and the effect of the failed C4 experiment on that number is not reported. The paper is transparent about the failed case, which is to its credit, but the headline numbers therefore characterize only the cases where reconstruction already succeeded, not the overall fidelity of the method. A related limitation appears in Sec. 5.3: for diffuse crowds the KS threshold is set to p_th = 1, meaning denoising is effectively disabled, and Sec. 6 says one can easily determine which threshold to use but provides no decision rule. This weakens the semantic claims, though the graph-reconstruction evidence is the more load-bearing issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"mmFlux proposes a privacy-preserving crowd flow analytics pipeline using a single commodity mmWave MIMO radar. The pipeline converts raw FMCW data into point clouds via range-only peak detection and targeted AoA, estimates pairwise flows with Lucas-Kanade optical flow on binary occupancy maps, denoises the resulting flow field with a Kolmogorov-Smirnov directional-consistency test, morphological pruning, and median filtering, then skeletonizes the field into a directed geometric graph whose vertices are split/merge points and whose edges are dominant flow currents. Flow split ratios are estimated by counting radar returns inside edge buffers, and curl/divergence fields computed from a least-squares Jacobian estimate are used to identify semantic events such as turns, splits, merges, dispersions, and gatherings. The paper reports experimental results from 21 experiments (18 structured, plus diffuse configurations) with crowds up to 20 people in three outdoor areas, claiming 17/18 successful topology reconstructions with mean one-sided Chamfer distance 0.45 m, edge orientation MAE 8.8 degrees, split-ratio MAE 0.1, and qualitatively validated semantic detections.","tokens_in":22392,"tokens_out":5542,"duration_ms":63249,"significance":"If the claimed results hold, mmFlux would be a useful step toward aggregate crowd-flow analytics that avoids individual tracking and preserves privacy, using off-the-shelf radar hardware. The graph-reconstruction formulation (skeletonization plus directional validation) and the Jacobian-based semantic analysis are well-motivated and are, to my knowledge, novel in the mmWave crowd-sensing literature. The authors are to be credited for running a nontrivial real-world data collection (up to 20 people, three areas, foliage/multipath) and for transparently discussing the one failed reconstruction. At the same time, the quantitative evidence is weaker than the abstract implies: the headline metrics exclude the failed case, use a one-sided distance, and lack baseline comparisons and error bars. The semantic claims are currently qualitative. These gaps are fixable with additional analysis, and the underlying methodology appears sound enough to warrant a revision rather than rejection.","major_comments":[{"comment":"The headline graph-reconstruction metrics are computed only on the 17 successfully reconstructed graphs, and the one-sided Chamfer distance d_avg averages over estimated points only. A one-sided distance does not penalize a missing ground-truth branch, so the reported 0.45 m and 8.8 degrees do not quantify completeness of the recovered topology. Because the paper's central claim is high-fidelity graph reconstruction, the evaluation should report a two-sided Chamfer distance or an equivalent precision/recall measure on all 18 structured experiments, including the C4 case with the missed upper-route edge.","section":"Sections 5.2 and 6"},{"comment":"The edge orientation MAE is computed for the 62 edges 'present in the successfully reconstructed graphs' and there is no accounting for spurious edges or wrong connectivity. This metric is therefore not an unbiased estimate of orientation accuracy. Please supplement it with edge-level precision and recall, a graph edit distance, or a quantitative comparison of vertex positions, and report per-experiment numbers so that the variance across the 18 runs is visible.","section":"Section 5.2"},{"comment":"The split-ratio MAE of 0.1 is averaged over the same successful experiments and depends on the recovered graph; the failed C4 case is excluded and no per-vertex or per-experiment error bars are reported. Given that split-ratio estimation is said to rely on an accurate G-hat topology, the reader cannot tell how much of the 0.1 error is due to flow-field estimation versus graph mis-reconstruction. Please report the per-vertex errors, including the C4 cases, and state the effect of the failed case on the aggregate MAE.","section":"Section 5.2 and Table 1"},{"comment":"The semantic curl/divergence claims (abrupt turns, split/merge boundaries, divergence source, gathering location) are supported only by qualitative visual comparisons in Fig. 7. Since the abstract and contributions claim that these semantics are 'accurately inferred,' please provide a quantitative evaluation, such as the distance between predicted curl/divergence peaks and ground-truth event locations for each configuration, or at least a labeled localization error per experiment.","section":"Section 5.3"},{"comment":"The pairwise flow estimate assumes that a detection at x_w,i has a counterpart inside the 2 m x 2 m search neighborhood in the next point cloud. Under occlusion, multipath, and missed detections, which the paper acknowledges, this association can fail, and every downstream step inherits those errors. Since no flow-field-level validation or sensitivity analysis is provided, please add a diagnostic quantifying pairwise-flow coverage (e.g., fraction of points with no reliable flow) or an ablation that measures how much downstream graph quality depends on the Lucas-Kanade neighborhood size and the KS threshold.","section":"Section 3.1, Eq. (7)"}],"minor_comments":[{"comment":"The opening of Section 5 states that 'all 18 experiments' show strong visual alignment, while the next paragraph and Section 6 state that one C4 experiment was only partially successful; this is internally inconsistent and should be corrected.","section":"Section 5"},{"comment":"The conditional definition of v_TAF(x) should be written with a cases environment; the current inline 'If |psi(x)| > 0' is awkward and could be confused with an argument of the function.","section":"Eq. (8)"},{"comment":"The choice p_th = 1 for diffuse crowds means that the KS-based denoising rejects nothing; given that p_th is described as a trade-off parameter, a brief sensitivity discussion or an alternative way to set it for diffuse flows would improve reproducibility.","section":"Section 5.3"},{"comment":"The caption says 'nearly all of the flow topology is recovered, except for one edge' while the text calls the result 'partially successful'; this is fine, but the wording should be unified so readers are not left wondering whether the miss is one edge or something larger.","section":"Figure 9 and Section 6"},{"comment":"The Chamfer distance is defined with an integral over P-hat; after discretization it should be stated as a sum over the discretized estimated graph points so that the reader can reproduce the number exactly.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"This is a solid systems paper with a real contribution, and the authors have been transparent about the one failure. The main gap is not the methodology but the evaluation protocol: the headline metrics currently overstate the graph-reconstruction claim because they are one-sided and exclude the failed case. I would ask for a revision that adds two-sided metrics, full-sample results, error bars, and baseline comparisons; with those additions the paper could be acceptable. The semantic evaluation also needs at least a minimal quantitative element."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know the core claim before anything else: this is the first RF-based work I've seen that reconstructs aggregate crowd flow topology from a single commodity mmWave radar without tracking individuals. That claim holds up. The literature review is careful, and the identified gap is real. The pipeline—binary occupancy maps, Lucas-Kanade optical flow, KS-based denoising, skeletonization, graph extraction, then Jacobian-based curl/divergence for semantics—is a coherent and sensible combination of existing tools, even if no single component is novel in isolation. The real experiments (21 runs, up to 20 people, three sites) are legitimate evidence, and the authors are transparent about the one C4 failure and the reasons for it. That transparency earns real credit; it's more than most papers do.\n\nThe soft spots are the ones the reader flagged, and I agree with them in proportion. The one-sided Chamfer distance and the edge-orientation MAE computed only on successfully reconstructed graphs cannot detect missing edges or spurious topology. A reconstruction that drops a branch can score fine. The failed C4 case is excluded from the numbers, and the split-ratio MAE of 0.1 also depends on the recovered graph, so the failure's effect on that number is unknown. The semantic curl/divergence results are qualitative; they look plausible but are not quantified. I'd also add that the KS threshold for diffuse crowds is set to 1, effectively disabling denoising, with no decision rule for when to switch beyond \"easily determine.\" These are addressable issues rather than fatal flaws. The pairwise flow association assumption (Eq. 7) is a legitimate concern, but the experimental success rate suggests it works well enough in practice; I wouldn't call it a load-bearing contradiction.\n\nOne point where I'd push back on the skeptic: the dependency on the authors' own pretrained GBM from [70] for point-cloud binarization is not a circularity problem. It is a reused component from prior work, and the target result here is the flow topology, not the binarizer. Same with the free parameters; they are tuned with physical justifications (pedestrian lane widths, spatial footprints), which is acceptable at this stage, though ablation would strengthen it.\n\nBottom line: this is a solid, honest systems paper with a new result and credible field evidence. It deserves a serious referee. The revision path is clear: report the C4 failure in the headline metrics or provide a two-sided/correspondence-aware metric, add error bars, and give at least one baseline comparison for the flow-field or graph-reconstruction quality. I'd bring it to reading group and would support peer review over desk rejection.","headline":"A genuinely new mmWave crowd-flow topology pipeline with an honest but underpowered evaluation—worth sending to review, conditional on fixing the metrics.","tokens_in":22998,"tokens_out":1034,"would_cite":false,"duration_ms":14963,"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":"One mmWave radar maps crowd flow without tracking anyone","keywords":["mmWave radar","crowd flow analytics","flow field estimation","optical flow","directed geometric graph","flow split ratio","curl and divergence","privacy-preserving sensing"],"falsifier":"Take a structured flow configuration with a deliberately extreme split ratio, such as 80 percent of the crowd on one branch and 20 percent on a branch partially occluded by foliage, and run mmFlux; the paper already reports that this setting caused a missed edge in one of three runs on configuration C4. A decisive test would repeat that scenario across many trials and check whether the one-sided Chamfer distance stays below the pedestrian-lane width of 0.76 m or whether the thin branch is consistently lost. A second, cleaner falsifier is to increase the per-frame displacement beyond 2 m, either by faster walking or a lower radar frame rate, and observe whether the reconstructed graph breaks exactly where equation (7)'s search neighborhood no longer contains the displaced detection.","tokens_in":21871,"feed_emoji":"📡","tokens_out":5500,"duration_ms":58366,"temperature":0.7,"pith_summary":"mmFlux claims that a single off-the-shelf mmWave radar can recover the aggregate structure of crowd motion, including where flows split, merge, turn, and gather, without ever tracking an individual person. The paper builds a pipeline that turns sparse radar point clouds into smooth 2D flow fields, then into directed geometric graphs whose edges are dominant flow currents and whose vertices are split or merge points. Across 21 real-world experiments with crowds of up to 20 people, the reconstructed graphs match ground-truth topologies in 17 of 18 structured cases, with an average one-sided Chamfer distance of 0.45 m, an edge-orientation mean absolute error of 8.8 degrees, and a flow split-ratio mean absolute error of 0.1. The practical stakes are privacy-preserving crowd analytics: flow patterns and crowd semantics can be sensed with radio reflections, not cameras, and without per-person tracking or identity.","feed_headline":"One mmWave radar maps crowd flow without tracking anyone","feed_subtitle":"Flow fields become graphs of splits and merges; curl and divergence reveal turns, dispersions, and gatherings.","key_machinery":"The load-bearing object is the estimated flow field $\\hat{\\mathbf v}(\\mathbf x)$, a 2D vector field built from pairwise Lucas-Kanade optical-flow matches between consecutive radar point-cloud frames. Because raw mmWave point clouds are sparse and noisy, the field is denoised by testing whether each point's flow directions are uniformly distributed around the circle, with a Kolmogorov-Smirnov test rejecting points that are, then pruned by connected-component size and smoothed by two rounds of median filtering. A normalized version $\\hat{\\mathbf v}_{\\mathrm{unit}}$ is skeletonized with the Zhang-Suen thinning algorithm, and the skeleton is converted into a directed geometric graph $\\hat G(\\hat V,\\hat E)$ whose edges carry flow-split ratios computed by counting radar returns inside polyline buffers. Semantic analysis then estimates the local Jacobian $\\mathbf J(\\mathbf x)$ by a least-squares fit in a neighborhood and reads off divergence and curl, mapping positive and negative divergence to dispersion and gathering and positive and negative curl to counterclockwise and clockwise turning.","core_discovery":"The paper's central claim is that aggregate crowd motion carries a describable topology that a single commodity mmWave MIMO radar can reconstruct. Concretely, mmFlux estimates a dense flow field from binary occupancy maps via optical-flow-style matching, cleans it with a Kolmogorov-Smirnov test for directional consistency plus morphological pruning and median filtering, skeletonizes the resulting field, and reads a directed geometric graph off the skeleton. The authors report that this graph agrees with hand-defined ground-truth flow topologies in 17 of 18 structured-crowd experiments and that flow split ratios are estimated to within a mean absolute error of 0.1. For both structured and diffuse crowds, the Jacobian of the flow field yields curl and divergence maps that localize sharp turns, direction-reversal boundaries, dispersions, and gatherings, including in experiments where no meaningful graph exists. Together these results are offered as evidence that flow-level sensing can replace the need for fragile per-person tracking in mmWave crowd analytics.","pith_inferences":["A natural testable extension is to treat the curl and divergence fields as time-series detectors: when the sign of divergence at a point flips, the pipeline could flag a transition from gathering to dispersal without any labeled event data.","If the 2 m search window is the limiting assumption, a multi-scale or coarse-to-fine variant of the optical-flow step would likely extend mmFlux to faster-moving crowds or lower radar frame rates; the paper does not test this.","The reported 0.45 m Chamfer distance sits inside a typical pedestrian lane width, suggesting the graph embedding is accurate enough for spatial analytics; a stricter test would measure how often reconstructed vertices fall inside the true split or merge zone rather than averaging edge distance.","The same flow-field representation may transfer to other sparse sensing modalities, such as automotive radar or large Wi-Fi arrays, whenever the uniform-noise assumption behind the Kolmogorov-Smirnov filter holds."],"forward_implications":["Crowd flow analytics can be done with radio signals alone, preserving privacy while still revealing how spaces are used.","Flow split ratios at junctions can be estimated accurately enough to guide retail layout, venue egress, and urban planning decisions.","Curl and divergence maps localize semantic events such as abrupt turns, gathering points, and panic-dispersal origins without requiring a graph model.","Because no per-person tracking is needed, the approach should scale to denser crowds where occlusions defeat individual-association methods.","The framework supplies a flow prior for crowd counting, replacing the uniform-spatial-usage assumption that earlier counting methods required."],"supporting_citations":[{"why":"Supplies the binary trace maps and multi-scale range detection pipeline that produce the input point clouds.","marker":"[70]"},{"why":"Provides the classical Lucas-Kanade least-squares optical-flow matching used for pairwise flow estimation between consecutive frames.","marker":"[54, 67]"},{"why":"Gives the Kolmogorov-Smirnov test used to discard flows whose direction distribution is consistent with uniform noise.","marker":"[59]"},{"why":"Provides the Zhang-Suen skeletonization algorithm that thins the denoised flow field to a single-pixel-wide structure for graph extraction.","marker":"[106]"},{"why":"Describes the commodity radar evaluation board whose FMCW MIMO measurements drive the entire pipeline.","marker":"[96]"},{"why":"Defines the one-sided Chamfer distance used to score spatial alignment between estimated and ground-truth flow graphs.","marker":"[51]"},{"why":"Supplies the polyline buffering method used to count radar returns near each graph edge and compute flow split ratios.","marker":"[85]"},{"why":"Provides a scene-flow pipeline from another mmWave domain, used to show that single-human motion flow methods do not generalize to crowd scale.","marker":"[26]"}],"fun_headline_variants":["Crowd flow mapped by single radar, no tracking needed","mmWave radar reads crowd motion as flow graphs","One radar reveals crowd splits, merges, and turns","Crowd analytics without per-person tracking","Flow fields from mmWave radar decode crowd behavior"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything downstream rests on the assumption that a person seen in one radar frame appears again one frame later inside a 2 m by 2 m search box, so the optical-flow step can find the true displacement; occlusion, multipath, and missed detections can break this association with no tracking or motion model to fall back on.","fun_headline_variants_meta":{"raw":{"variants":["Crowd flow mapped by single radar, no tracking needed","mmWave radar reads crowd motion as flow graphs","One radar reveals crowd splits, merges, and turns","Crowd analytics without per-person tracking","Flow fields from mmWave radar decode crowd behavior"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1292,"prompt_tokens":966,"completion_tokens":326,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":251}},"tokens_in":582,"tokens_out":326,"duration_ms":3682,"temperature":1.0,"reasoning_tokens":251,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:44:14.319037+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a structured flow configuration with a deliberately extreme split ratio, such as 80 percent of the crowd on one branch and 20 percent on a branch partially occluded by foliage, and run mmFlux; the paper already reports that this setting caused a missed edge in one of three runs on configuration C4. A decisive test would repeat that scenario across many trials and check whether the one-sided Chamfer distance stays below the pedestrian-lane width of 0.76 m or whether the thin branch is consistently lost. A second, cleaner falsifier is to increase the per-frame displacement beyond 2 m, either by faster walking or a lower radar frame rate, and observe whether the reconstructed graph breaks exactly where equation (7)'s search neighborhood no longer contains the displaced detection.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the binary trace maps and multi-scale range detection pipeline that produce the input point clouds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Zhang-Suen skeletonization algorithm that thins the denoised flow field to a single-pixel-wide structure for graph extraction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the commodity radar evaluation board whose FMCW MIMO measurements drive the entire pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the one-sided Chamfer distance used to score spatial alignment between estimated and ground-truth flow graphs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the polyline buffering method used to count radar returns near each graph edge and compute flow split ratios."}],"review_version":1}