{"id":"9d6a71db-4260-45e9-b416-f578142f3fb8","arxiv_id":"2509.06672","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A two-segment optimization maps per-path delays and angles to equivalent reflection points, yielding 3D object point clouds in 6.75 GHz multi-bounce ray-tracing simulations.","lead":"This paper shows how the delays and angles of reflected wireless signals can be turned into 3D point clouds of objects such as trees and cars, handling reflections that bounce several times. The method is demonstrated in 6.75 GHz ray-tracing simulations, targeting a future 6G capability where base stations map their environment using ordinary data signals.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multi-bounce ERPs may not lie on the object surface; the claimed 'first demonstration of multi-bounce ISAC imaging' rests on an unverified equivalence.","rationale":"The reader's identified weakest assumption—perfect CSI with no noise or bandwidth constraints—is a real limitation for practical deployment, and the paper's self-disclosure in Section II-B and Table I is clear. However, the more load-bearing concern for the paper's central claim is upstream: even with perfect path knowledge, the ERP computed from Eq. (21)-(22) is not guaranteed to lie on the object surface for multi-bounce paths, because the optimization lacks a surface constraint and the delay constraint forces a shift relative to the true bounce points. Since the claimed novelty is specifically multi-bounce ISAC imaging, this unverified equivalence is what the title and abstract contributions rest on. The perfect-CSI issue affects robustness; the surface-proximity issue affects whether the multi-bounce mapping is meaningful at all. Both support the same CONDITIONAL verdict: the paper should add the proposed validation or soften the multi-bounce claim. I therefore keep the reader's verdict unchanged while adding a distinct condition that should be checked before the multi-bounce novelty is accepted.","tokens_in":23679,"tokens_out":7061,"duration_ms":88542,"concrete_test":"Use the NYURay pipeline to export, for every path, the ordered list of interaction points (or at least first and last bounce points) and the object mesh. Compute each ERP via Eq. (21)-(22). (a) Report the distribution of distance from each ERP to the nearest surface point, separately for single-bounce and multi-bounce paths, for the 1 m cube and the standard tree. (b) Build the MVF point cloud using only multi-bounce paths and compute Chamfer distance to the reference mesh. If the multi-bounce-only CD is comparable to the single-bounce-only CD and the median ERP-to-surface distance is small (e.g., <0.1 m for the 1 m cube), the abstraction is validated; if not, the central claim must be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that every multipath component, including multi-bounce paths, can be reduced to an ERP on or near the reflecting object. The optimization in Section III-B does not enforce this. Problem (9) minimizes ||a(alpha)-b(beta)||^2 subject only to alpha+beta=L and box constraints; there is no surface constraint. For a true two-bounce path with internal path length s=|A-B|, the first and last bounce points A, B satisfy alpha0+beta0 = L-s. Enforcing alpha+beta=L forces the minimizer to shift the TX- and RX-segment points beyond A and B by a total of s, so the ERP in Eq. (22) can lie inside or outside the object rather than on its surface. The gamma filter (Eq. 23) only rejects cases where the two segments are far apart; it does not enforce proximity to the known surface. The paper never compares ERPs to NYURay's ground-truth bounce locations or to the Blender mesh. The figures and Chamfer curves could be dominated by single-bounce paths; the unique multi-bounce capability is asserted but not separately validated. If this equivalence fails, the novelty claim 'first demonstration of multi-bounce ISAC imaging using wireless ray tracing' is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 6G ISAC imaging pipeline that maps per-path multipath parameters (AoD, ZoD, AoA, ZoA, ToA, gain) from the NYURay ray tracer into 3D points called Equivalent Reflection Points (ERPs). For each path, the method solves a two-segment optimization problem: it estimates the transmitter-side distance alpha and receiver-side distance beta by minimizing the distance between the two corresponding lines subject to alpha+beta = c*tau (Eq. 9), yielding the closed-form solution in Eq. (21) and the ERP in Eq. (22). ERPs from multiple TX-RX positions are aggregated and filtered by a chord-length threshold gamma (Eq. 23) to produce dense point clouds. The method is demonstrated on six Blender-rendered objects (tree, cubes, triangle, circle, Tesla) at 6.75 GHz, with Chamfer distance evaluated versus number of TX-RX pairs. The paper claims the first demonstration of multi-bounce ISAC imaging using wireless ray tracing.","tokens_in":23996,"tokens_out":4303,"duration_ms":51277,"significance":"The closed-form ERP computation is clean, mathematically correct, and has O(1) complexity per path, giving a useful low-cost geometric fusion of angle and delay measurements for single-bounce-dominant scenarios. The simulation campaign covers diverse object geometries, and the MVF strategy is sensible. However, the paper's central claims are not yet fully supported: the validation bypasses actual CSI estimation, uses perfect path knowledge, and never isolates or validates the multi-bounce behavior that is the main claimed novelty. If the multi-bounce equivalence is rigorously tested and the sensitivity to estimation error is quantified, this could be a valuable step toward ray-tracing-based ISAC imaging. In its current form, the contribution is better described as an idealized geometric mapping algorithm than as a validated CSI-based ISAC imaging framework.","major_comments":[{"comment":"The multi-bounce ERP claim is not validated. Problem (9) minimizes the distance between the TX and RX lines subject only to alpha+beta=L and box constraints; there is no constraint that the ERP lie on the object surface S. For a two-bounce path with internal segment length s, the true first/last bounce points satisfy alpha0+beta0=L-s, so enforcing alpha+beta=L can move the fitted point inside or outside the object. The gamma filter in Eq. (23) only bounds the chord between the two segment endpoints; it does not enforce proximity to the surface. The paper never compares ERPs to NYURay's ground-truth bounce locations or to the Blender mesh, nor does it report Chamfer distance separately for single-bounce and multi-bounce paths. The 'first demonstration of multi-bounce ISAC imaging' is therefore unsupported. Please provide a separate validation on multi-bounce-only paths, reporting ERP-to-t","section":"Section III-B, Eqs. (9) and (22); Section IV"},{"comment":"The central reconstruction claim is made under perfect path knowledge. Section II-B states that 'ground truth paths generated in Procedure 2 are fed directly into Procedure 5, thereby assuming perfect path knowledge,' and Table I confirms 'Noise and Bandwidth Constraints: Not included (idealized CIR used).' Thus the experiments do not exercise the CSI estimation chain implied by the title. The impact of realistic angle/delay estimation errors on ERP positions and on the final Chamfer distance is not analyzed. Section V acknowledges the limitation, but the abstract and conclusion still claim that the framework 'accurately reconstructs object surfaces.' Please add a sensitivity analysis (e.g., perturb angles and delays with realistic error levels, or simulate a standard estimator such as matrix pencil) and show CD degradation; otherwise temper the claims to an idealized geometric mapping s","section":"Section II-A, Section II-B, Table I"},{"comment":"The quantitative evaluation lacks a baseline and an adequately specified reference cloud. The Chamfer distance in Eq. (25) is computed against a 'reference cloud accurately describing the object image,' but the sampling of S from the Blender mesh is not described (density, surface coverage, whether edges are included). Without a baseline method (e.g., single-bounce ellipsoid intersection, backprojection, or a least-squares triangulation), the CD curves in Fig. 12 only show that the method improves with more views; they do not demonstrate that the reconstruction is accurate relative to existing approaches. Please specify the reference sampling and add at least one baseline under identical simulation conditions.","section":"Section IV, Eq. (25), Fig. 12"},{"comment":"The geometric filter threshold gamma=10 is a free parameter, chosen as the longest edge-to-edge dimension of any target. For the 1 m cube, this permits ERPs whose incoming and outgoing segment endpoints are up to 10 m apart, which is far larger than the object; the filter may therefore admit many off-surface points. No sensitivity analysis with respect to gamma is reported, so it is unclear how much of the reconstruction quality depends on this arbitrary value. Please provide a gamma sweep, or derive the threshold from physical path-length statistics, and report CD or outlier fraction versus gamma.","section":"Section III-C and Table I"}],"minor_comments":[{"comment":"Several typos: 'metal triangular' should be 'metal triangle', 'vehicule' should be 'vehicle', 'silhouete' should be 'silhouette'. Also, 'Mx=My=1 isotropic antennas' is inconsistent with the steering-vector model in Eqs. (2)-(4); clarify that per-path angles are supplied by the ray tracer rather than estimated from a physical array.","section":"Section IV"},{"comment":"The matrices defining Txpos and Rxpos have repeated rows (e.g., rows 3 and 4 of Txpos are identical). If this is intentional, say so; otherwise correct the listed TX-RX placements.","section":"Eq. (24)"},{"comment":"The y-axis is labeled log(ChamferDistance) but the text says 'log2 scale' and later quotes reductions as '2^8.6'. Use consistent notation, preferably log2, and state units in the axis label.","section":"Fig. 12"},{"comment":"The text refers to 'Procedure 2' and 'Procedure 5', but Fig. 2 labels steps as numbered boxes. Renumber or cross-reference consistently to avoid confusion.","section":"Section II-B, Fig. 2"},{"comment":"The phrase 'first demonstration of multi-bounce ISAC imaging using wireless ray tracing' should be qualified as a simulation demonstration, since the paper uses synthetic ray-tracing output rather than over-the-air measurements.","section":"Abstract/Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The core closed-form derivation is sound and the paper is likely to be of interest to the ISAC and ray-tracing community. However, the 'first demonstration of multi-bounce ISAC imaging' claim rests on an unverified equivalence between multi-bounce paths and a single ERP on the object surface. The authors should either rigorously validate that equivalence with ground-truth bounce comparisons, or reframe the contribution as a single-bounce-dominant imaging method with multi-bounce paths handled heuristically. In addition, the absence of any CSI estimation step makes the title's promise stronger than the actual experiments. I would support publication after these points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent, clearly written paper whose real contribution is a closed-form, two-segment equivalent reflection point (ERP) optimization (Eq. 21) plus multi-vantage fusion. The math is sound. The validation, however, is entirely simulation under perfect CSI, no noise, no bandwidth limits, with no baseline comparison. The 'first demonstration of multi-bounce ISAC imaging' claim is not yet supported because the paper never validates that multi-bounce ERPs land on or even near the object surface.\n\nWhat's new: the idea of replacing a multi-bounce trajectory with a single ERP by minimizing the distance between the TX-segment and RX-segment under the alpha+beta=L delay constraint is a neat, simple construction. The closed-form solution is correct; complexity is linear. The fusion step is straightforward but does produce denser point clouds in the qualitative figures, and Chamfer distance decays with more TX-RX pairs for all six objects. Credit also for using a calibrated ray tracer and for being explicit about what is assumed.\n\nWhere it's soft: The weakest point is the surface claim. For a two-bounce path with internal length s, imposing alpha+beta=L forces the ERP to shift by s relative to the actual first/last bounce points, so the ERP can sit inside or outside the object. The gamma filter does not fix this; it only rejects large segment gaps. The paper never compares ERPs to NYURay's ground-truth interaction points, and the good-looking point clouds could be dominated by single-bounce paths. The abstract's phrasing 'on the object surface' is therefore misleading, even if the body carefully calls it an equivalent point. Also, the gamma threshold is fitted to the test objects, and there is no sensitivity analysis, no error bars, and no baseline comparison against single-bounce-only imaging or other multipath exploitation methods. The paper itself acknowledges the perfect-CSI limitation and says real experiments are future work, which is honest but doesn't change the conditional status.\n\nBottom line: This is a solid algorithmic contribution that deserves to be published after revision, but it is not a breakthrough. I would send it to peer review and ask the reviewers to demand: (1) an ablation separating single- and multi-bounce paths, (2) sensitivity to angle/delay errors, (3) a single-bounce baseline, and (4) code/data release. The authors should also soften the 'first demonstration' claim unless they can show multi-bounce ERPs actually trace the object surface.","headline":"Clean closed-form ERP mapping with honest limitations, but the multi-bounce imaging claim is only as strong as the untested equivalence — conditional accept.","tokens_in":24442,"tokens_out":4172,"would_cite":true,"duration_ms":45160,"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":"Every multipath echo collapses into one 3D reflection point.","keywords":["integrated sensing and communications","equivalent reflection point","multipath imaging","channel state information","ray tracing","6G","point cloud reconstruction","6.75 GHz"],"falsifier":"Run the same imaging pipeline on CSI whose per-path parameters carry realistic estimation errors—for example, add measurement noise to the ray-tracer angles and delays consistent with the bandwidth and SNR of the table, or use real measured CSI at 6.75 GHz—and compare the resulting point clouds to the known object geometry; if the Chamfer distance grows proportionally with the parameter errors or the surfaces and edges no longer register, the central reconstruction claim is falsified. A simpler check: displace a single angle by one standard deviation of a typical high-resolution estimator and","tokens_in":23590,"feed_emoji":"📡","tokens_out":7093,"duration_ms":65883,"temperature":0.7,"pith_summary":"The paper aims to show that the per-path channel state information already present in a wireless link—each path's angles, delay, and power—is enough to produce 3D images of objects in the environment, with no dedicated radar hardware. Its central claim is that every resolvable multipath component, including multi-bounce paths that bounce around inside or between objects, can be reduced to a single equivalent reflection point (ERP) on or near the reflecting surface. The reduction is carried by a two-segment optimization that splits the measured total path length into a transmitter-side and a receiver-side segment, solving for the closest pair of points along the two angle-defined lines in closed form, and taking their midpoint as the ERP. Aggregating such points across multiple transmitter-receiver positions yields dense point clouds that, in the paper's simulations, recover surfaces, edges, and curved features of a tree, cubes, plates, and a car model at 6.75 GHz. If the claim holds, ordinary communication links become environmental sensors for digital twins, blockage prediction, and beam management in 6G.","feed_headline":"Every multipath echo collapses to one 3D reflection point","feed_subtitle":"Multi-bounce radio paths compress into surface points, letting base stations map objects with existing CSI.","key_machinery":"The two-segment reflection point optimization. For a path with departure direction d̂_T, arrival direction d̂_R, and total length L = cτ, it places a point P = p_Tx + α d̂_T on the departure ray and P = p_Rx − β d̂_R on the arrival ray, minimizing ∥(p_Tx + α d̂_T) − (p_Rx − β d̂_R)∥² under α + β = L. The minimizer is exact and closed-form (Eq. 21), and the equivalent reflection point (Eq. 22) is the average of the two closest points. This is the mechanism that abstracts multi-bounce interactions into a single surface point, enabling single-bounce geometry to be assumed even when the true trajectory is much more complex. The multi-vantage fusion step then aggregates these points across positi","core_discovery":"For each measured path component, the departure and arrival angles define two rays in space, and the delay fixes the total distance the signal traveled. The paper shows that the point best reconciling these three pieces of information is the minimizer of the squared distance between the two rays under the constraint that the two segment lengths sum to the total path length. That minimizer has a closed-form expression: the transmitter-side length is α* = −(p_Tx − p_Rx + L d̂_R)ᵀ(d̂_T − d̂_R)/∥d̂_T − d̂_R∥², the receiver-side length is L − α*, and the ERP is the midpoint of the two segment endpoints. For multi-bounce paths this point is not an actual scatterer but an equivalent one: it collaps","pith_inferences":["The procedure effectively reframes inverse scattering as a per-path geometric fit; for strongly diffuse or penetrating materials, an uncertainty-weighted version (e.g., the power-weighted aggregate cost the paper notes as future work) would likely be needed to keep weak paths from corrupting the map.","The paper's own open-challenges section implies that angle/delay estimation error, hardware impairments, and limited bandwidth are the natural stress test; a sensitivity analysis displacing ERPs under realistic estimation noise would show how far the ideal-path results degrade.","The linear-scaling complexity and the observed convergence of Chamfer distance suggest a practical resource law: radial resolution set by bandwidth, angular resolution by array aperture, and coverage by the number of vantage points—an explicit relation the paper leaves open.","Extending the fusion to moving receivers or multiband FR3 spectrum (as the paper's SAR-inspired question suggests) could synthesize a larger effective aperture and sharpen the point cloud beyond a single snapshot."],"forward_implications":["Every resolvable path in a standard CSI snapshot becomes a 3D point, so imaging can ride on ordinary pilot signals without extra sensing hardware or spectrum.","Multi-bounce paths—historically the main obstacle to CSI-based imaging—are handled by the same closed-form formula, extending the method to foliage, vehicles, and other complex scatterers.","Per-path computation is O(1) and total work is linear in the number of TX-RX pairs, leaving room for real-time mapping in 6G beam management and blockage prediction.","Because the formula is frequency-agnostic, the same imaging pipeline transfers to other bands or to measured CSI, provided per-path delays and angles are available.","Chamfer distance falls steeply with added views—roughly to 2^-11 after six TX-RX pairs for flat or cubic objects—so a handful of vantage points can already close the surface."],"supporting_citations":[{"why":"Calibration of the ray tracer against measured channels at 28/73/142 GHz; supplies the engine whose per-path outputs drive the imaging pipeline.","marker":"[32]"},{"why":"6.75 GHz indoor propagation and material penetration measurements; the channel data the ray tracer is calibrated on at the employed frequency.","marker":"[41]"},{"why":"Upper mid-band channel measurements at 6.75/16.95 GHz in a factory scenario; extends the same calibration evidence to the working frequency.","marker":"[42]"},{"why":"LMMSE channel estimation for OFDM; the standard CSI estimation route that would provide per-path parameters in a real system.","marker":"[55]"},{"why":"Single-snapshot joint angle/delay estimation via matrix pencil; one concrete high-resolution method for extracting the per-path six-tuples from the received signal.","marker":"[56]"},{"why":"Material electrical properties standard; supplies the wood permittivity and conductivity used in the tree test object.","marker":"[59]"},{"why":"Balanced Chamfer distance; the metric used to quantify reconstruction accuracy as a function of TX-RX pair count.","marker":"[66]"},{"why":"Site-specific ray-tracing prediction of path loss in multifloored buildings; establishes ray tracing as a validated basis for generating site-specific channel impulse responses.","marker":"[43]"}],"fun_headline_variants":["Multipath collapses to a single 3D reflection point","Every echo becomes one point: 6G imaging via CSI","Two rays + delay = one surface point for mapping","6.75 GHz: multi-bounce paths compress to one point","One point per path: ISAC imaging from CSI"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the per-path angle, delay, and gain values are known perfectly—the paper feeds the ray tracer's ground-truth paths directly into the imaging step, so any real-world channel-estimation noise or bandwidth limit that corrupts these parameters would displace every equivalent reflection point.","fun_headline_variants_meta":{"raw":{"variants":["Multipath collapses to a single 3D reflection point","Every echo becomes one point: 6G imaging via CSI","Two rays + delay = one surface point for mapping","6.75 GHz: multi-bounce paths compress to one point","One point per path: ISAC imaging from CSI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000269,"raw_usage":{"total_tokens":1479,"prompt_tokens":789,"completion_tokens":690,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":607}},"tokens_in":533,"tokens_out":690,"duration_ms":7834,"temperature":1.0,"reasoning_tokens":607,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:17:14.613790+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same imaging pipeline on CSI whose per-path parameters carry realistic estimation errors—for example, add measurement noise to the ray-tracer angles and delays consistent with the bandwidth and SNR of the table, or use real measured CSI at 6.75 GHz—and compare the resulting point clouds to the known object geometry; if the Chamfer distance grows proportionally with the parameter errors or the surfaces and edges no longer register, the central reconstruction claim is falsified. A simpler check: displace a single angle by one standard deviation of a typical high-resolution estimator and","supporting_citations":[{"cited_title":"Calibration of NYURay for Ray Tracing Using 28, 73, and 142 GHz Channel Measurements Conducted in Indoor, Outdoor, and Factory Scenarios,","cited_arxiv_id":null,"evidence_quote":"Calibration of the ray tracer against measured channels at 28/73/142 GHz; supplies the engine whose per-path outputs drive the imaging pipeline."},{"cited_title":"Comprehensive FR1(C) and FR3 Lower and Upper Mid-Band Propagation and Material Penetration Loss Measurements and Channel Models in Indoor Environment for 5G and 6G,","cited_arxiv_id":null,"evidence_quote":"6.75 GHz indoor propagation and material penetration measurements; the channel data the ray tracer is calibrated on at the employed frequency."},{"cited_title":"Upper Mid-Band Channel Measurements and Characterization at 6.75 GHz FR1(C) and 16.95 GHz FR3 in an Indoor Factory Scenario,","cited_arxiv_id":null,"evidence_quote":"Upper mid-band channel measurements at 6.75/16.95 GHz in a factory scenario; extends the same calibration evidence to the working frequency."},{"cited_title":"Channel Estimation for OFDM,","cited_arxiv_id":null,"evidence_quote":"LMMSE channel estimation for OFDM; the standard CSI estimation route that would provide per-path parameters in a real system."},{"cited_title":"Single snapshot joint estimation of angles and times of arrival: A 2D Matrix Pencil approach,","cited_arxiv_id":null,"evidence_quote":"Single-snapshot joint angle/delay estimation via matrix pencil; one concrete high-resolution method for extracting the per-path six-tuples from the received signal."},{"cited_title":"Effects of Building Materials and Structures on Radiowave Propagation Above About 100 MHz,","cited_arxiv_id":null,"evidence_quote":"Material electrical properties standard; supplies the wood permittivity and conductivity used in the tree test object."},{"cited_title":"Balanced chamfer distance as a comprehensive metric for point cloud comple- tion,","cited_arxiv_id":null,"evidence_quote":"Balanced Chamfer distance; the metric used to quantify reconstruction accuracy as a function of TX-RX pair count."},{"cited_title":"914 MHz path loss prediction models for indoor wireless communications in multifloored buildings,","cited_arxiv_id":null,"evidence_quote":"Site-specific ray-tracing prediction of path loss in multifloored buildings; establishes ray tracing as a validated basis for generating site-specific channel impulse responses."}],"review_version":1}