acceptodds
Under review as a conference paper at ICLR 2027

GaussTwin: A Unified Scene Representation for Wireless Channel Prediction with Surface Recovery

Abstract

Given channel measurements at a set of transmitter (Tx)-receiver (Rx) links, site-specific channel prediction aims to capture the effects of the propagation environment on angular and delay power distributions and received power at unmeasured Tx-Rx links. Existing methods either take scene geometry as prior information or learn radio representations that do not recover surfaces. We propose GaussTwin, a site-specific wireless channel model that represents the scene with oriented 2D Gaussian surface primitives, predicts channel characteristics at unseen links, and recovers explicit environment geometry from channel measurements. A physical propagation operator generates these characteristics for different Tx-Rx pairs from the primitives and models both specular reflection and diffuse scattering. To the best of our knowledge, GaussTwin is the first method that predicts channel characteristics and can recover explicit environment geometry from channel measurements at the same time without geometric supervision or prior information. Results across five indoor scenes and one outdoor city block show that GaussTwin achieves high-quality predictions while retaining strong surface recovery capability. On public datasets with real measurements, GaussTwin outperforms state-of-the-art baselines on spatial spectrum prediction and achieves comparable performance on received power prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Wireless Digital Twin, Surface Reconstruction, Spatial Representation Learning, Differentiable Propagation, Inverse Rendering

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_WC9EOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GaussTwin for jointly predicting wireless channel characteristics and recovering scene surfaces from channel measurements, without external geometry or geometric supervision. It represents the environment with oriented 2D Gaussian primitives and predicts power distributions using a physical model of direct propagation, reflection, scattering, and occlusion. The primitives are initialized by RF backprojection and refined through channel prediction losses, replacement, and splitting. Experiments evaluate channel prediction and surface recovery in six simulated environments. Two public measurement datasets are used to evaluate spatial spectrum and RSSI prediction.

Critical strengths
  1. The same surface primitives support channel prediction and geometry recovery. They model reflection, scattering, and occlusion across multiple links, allowing different channel observations to constrain a shared scene representation. This provides an explicit connection between the learned geometry and the propagation model.

  2. The initialization and structural refinement form a plausible solution to a difficult nonconvex problem. RF backprojection provides spatial support. Replacement guided by residuals and splitting based on visibility conflicts address limitations of local gradient updates.

  3. The reported results provide useful evidence of feasibility in controlled environments. Table 1 shows substantial indoor channel prediction gains over the adapted baselines, and the appendix examines initialization, geometry learning, propagation mechanisms, primitive capacity, and measurement sparsity.

Critical weaknesses
  1. Initialization in Appendices C.1 and C.6.1 requires complex channels and subcarrier frequencies, which are not provided by RFID spectrum images or BLE RSSI measurements. These inputs lack the required phase and wideband delay information, and the paper does not describe how initialization and surface refinement are adapted to them. For the RFID experiment, Equation 13 uses a complex response without explaining its relation to phase averaging and diffuse power in the main model.

  2. In Section 4.1, existence probability and reflectance enter the reflection and scattering power factors through their product piϱip_i\varrho_i. For a primitive observed only through reflection or scattering, without independent attenuation constraints, different parameter pairs with the same product yield identical predictions when geometry and other parameters are fixed. These observations cannot uniquely determine both parameters.

  3. Section 3 uses one channel realization per link to estimate mean power, but Equations 7 and 8 average over random phases. The power in a single realization can differ substantially from the predicted mean because of interference between paths. The experiments do not assess how this discrepancy affects prediction and geometry recovery.

Questions for authors
  1. What measurements and procedures are used to initialize and refine the RFID and BLE models? How does the complex response in Equation 13 handle random phases and diffuse scattering?

  2. How is the ambiguity between pip_i and ϱi\varrho_i handled without independent attenuation constraints? Can each parameter be validated against known values?

  3. Why is power from one channel realization suitable for fitting the means in Equations 7 and 8? Can the authors quantify how path interference affects prediction and geometry recovery?

3 comments
AuthorsOpenReview ↗

Response to Weakness 3 & Question 3

We thank Reviewer WC9E for the careful review and the positive assessment of our work.

Weakness 3 & Question 3. GaussTwin predicts only power and averages over the reflection phases in Equations 7–8. In the marginal spectra used for training, single-realization power stays close to this phase average, as quantified in General Response G1 and the new Appendix D.7.

AuthorsOpenReview ↗

Response to Weakness 1-2 & Question 1-2

Weakness 1 & Question 1. Appendices B.3, B.4 and C.6.1 now describe both adaptations. Initialization keeps the placement procedure of Appendix C.6.1, using only training observations with their Tx and Rx positions. RFID builds its support from spectrum residuals, BLE from excess received power, and each refines the surfaces through errors in its own observable. This keeps both models as close as possible to the main method.

For RFID, Equation 14 writes out the complex response of Equation 13. Its reflection phase is learned instead of drawn at random, and it omits the diffuse term. BLE keeps both the phase average and the diffuse term of Equations 7–8.

Weakness 2 & Question 2. Reflection and scattering constrain only the product αi=piϱi\alpha_i = p_i\varrho_i, while blockage separates the two parameters through the transmittance ti=1−piwit_i = 1-p_iw_i, which involves pip_i without ϱi\varrho_i. Appendix C.2 now verifies this numerically. When pip_i varies with αi\alpha_i held fixed, predictions barely change for a primitive that blocks no direct path, but change substantially for one that does.

The learned reflectances on the known surfaces preserve the concrete–wood ordering of the simulated materials (Appendix C.2). The existence probability has no direct reference, and the reported surface metrics evaluate geometry independently of it.

Thank you for your response. I have no further questions. I think the topic of this paper is very interesting, and the overall quality is good. Since I am not very familiar with this topic, I will keep my rating at weak accept.

Reviewer_inn7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents GaussTwin, a unified scene representation for site-specific wireless channel prediction and surface recovery from position-labeled channel measurements, without geometric supervision. It represents the environment using oriented 2D Gaussian surface primitives and a differentiable propagation model incorporating blockage, specular reflection, and diffuse scattering. The primitives refined using channel prediction losses. Experiments across simulated indoor and outdoor environments demonstrate joint channel prediction and surface recovery.

Critical strengths
  1. Jointly predicting wireless channel characteristics and recovering explicit surfaces without geometric supervision is a meaningful research objective.
  2. The surface representation and differentiable propagation model provide a physically interpretable connection between scene geometry and wireless observations.
Critical weaknesses
  1. The benefit of the 2D Gaussian surface representation is not sufficiently isolated. Explicit surface normals make this representation well suited to the proposed reflection and scattering model, but the paper does not directly compare it with a suitably adapted 3D Gaussian representation under comparable training conditions and model capacity. Existing baseline comparisons also differ in their propagation models, making the contribution of the representation itself difficult to assess.
  2. The geometry-recovery component lacks real-world validation. While the RFID and BLE experiments demonstrate spatial spectrum and received power prediction on real measurements, neither evaluates the accuracy of recovered surfaces. The paper’s central contribution is a unified representation supporting both channel prediction and geometry recovery, yet this joint capability is only validated in simulation. Although the authors acknowledge the lack of geometric ground truth in public datasets, it remains unclear whether the learned primitives correspond to physically meaningful surfaces in real environments, especially when only limited channel observables are available.
  3. Channel prediction accuracy does not necessarily imply geometric fidelity. Table D.1 shows that optimizing primitives initialized from ground-truth geometry improves PDP NMSE from −5.59 to −9.00 dB, but worsens Chamfer-L1 distance from 3.56 to 13.85 cm. Although the authors acknowledge that learned geometry can compensate for unmodeled propagation effects, this finding raises a fundamental concern about the identifiability and physical reliability of the recovered surfaces. The paper would benefit from further analysis of when the optimization preserves true geometry and when it introduces geometrically inaccurate primitives to improve channel fitting.
  4. The phase-marginalization assumption requires further empirical justification. GaussTwin marginalizes independent uniformly distributed reflection phases, suppressing interference terms between most distinct reflection paths. However, the observations consist of a single coherent channel realization per Tx–Rx link, whose angle–delay power may differ from the phase-marginalized prediction. Although the mathematical derivation is provided in Appendix C.4, the paper does not quantify the impact of this modeling mismatch, particularly when multipath components are poorly resolved in angle or delay. An analysis of its effect on channel prediction and geometry recovery would strengthen the justification for this assumption.
  5. Robustness to practical measurement imperfections remains underexplored. The indoor simulations use noise-free channel tensors without phase distortion, together with a 1 GHz bandwidth and relatively large antenna arrays. Although Appendix D.5 examines reduced observation density, it considers only the parallel-wall scene. The proposed initialization and geometry optimization rely heavily on accurate delay, angular, and transceiver-position information. It would therefore be valuable to evaluate sensitivity to position errors, antenna calibration errors, phase distortion, and reduced measurement resolution, particularly in terms of the recovered geometry rather than channel prediction alone.
  6. The reconstruction advantage is not sufficiently established against independent geometry-recovery methods. Although the paper compares channel prediction with adapted baselines and includes useful geometry-learning ablations, the surface recovery results are primarily evaluated against the method's own variants. Since back-projection is already used for primitive initialization, a comparison with a standalone back-projection or migration-based reconstruction method using the same channel measurements would provide a meaningful reference. Such a comparison would help distinguish the benefits of the proposed joint optimization from those of the initial geometric estimation procedure.
  7. The real-world evaluation lacks sufficient context about the measurement environments and reference data. Although Appendix B reports the sampling counts, data splits, and basic radio configurations of the public RFID and BLE datasets, it provides limited information about the measurement hardware, scene layouts, and multipath conditions. The construction of the measured RFID spatial spectra is also insufficiently explained; the appendix mainly describes the predicted spectra and their normalization. A brief description of the hardware and reference-data acquisition procedures, together with scene layouts showing transceiver placement, would help readers assess how challenging these benchmarks are and whether they adequately test the proposed multipath modeling capabilities.
Questions for authors
  1. Could the authors provide a small real-world experiment with calibrated channel measurements and reference geometry used only for evaluation, reporting both channel prediction and surface recovery on the same scene?
  2. Under what conditions should the recovered primitives be interpreted as physical surfaces rather than effective reflectors that compensate for model mismatch? Could the authors analyze where the geometry departs from ground truth in Table D.1 and how these deviations relate to unmodeled propagation effects?
  3. What justifies using a single channel realization as the target for phase-averaged predictions? Could the authors quantify the approximation error, and compare against a variant that retains or learns reflection phases?
  4. How sensitive are channel prediction and geometry recovery to measurement noise, and transceiver position or calibration errors? Could the measurement-sparsity experiment also be extended to the six-room scene?
4 comments
AuthorsOpenReview ↗

Response to Weakness 4 & Question 3

We thank Reviewer inn7 for the detailed review and address the points one by one below.

Weakness 4 & Question 3. Please see General Response G1 for the measured interference difference, including its relation to prediction and geometry learning. The sum of traced path powers removes all cross terms, whereas the phase average of GaussTwin keeps those between paths with identical reflection counts on every primitive (Appendix C.4). With paths grouped at about the primitive scale, these carry under 1% of link energy at the 99th percentile. A variant with retained or learned reflection phases can only reduce the interference part of the error. The reference replacement in G1 puts this part at only 0.2–0.4 dB of median PDP NMSE, which approximates the largest gain of such a variant in the marginal spectra used for training. Coherent prediction at unseen links would also require path lengths accurate to a fraction of a wavelength, which the phase average avoids.

AuthorsOpenReview ↗

Response to Weakness 1 & 6

Weakness 1. A comparison with 3D Gaussian primitives would not in fact demonstrate the effectiveness of our planar 2D Gaussian primitives, although a 3D Gaussian could be adapted to our method in two ways.

  1. Direct replacement would require a new propagation mechanism. The result is a distinct method that does not permit a meaningful comparison.
  2. If the shortest axis serves as the normal and the two longer axes span the interaction plane, our reflection model is kept. This adaptation is essentially equivalent to the present method.

Our 2D primitive was not derived from 3D Gaussian methods but follows directly from modeling ray interactions, where a planar element yields a simple intersection and an elliptical shape provides flexibility.

Weakness 6. The standalone back-projection reference is the Fixed init. variant in Table D.1, which keeps the primitive geometry at its initialization throughout training. The full GaussTwin method outperforms it on both the channel and the geometry metrics. The geometric accuracy therefore comes not only from the initialization but also from learning the channel characteristics.

AuthorsOpenReview ↗

Response to Weakness 2,3,7 and Question 1-2

Weakness 2 & Question 1. Please see General Response G2, which explains why the public measured datasets, including DICHASUS and WiWorld-RealData, are not suited to evaluating geometry reconstruction.

Weakness 3 & Question 2.

  1. From ground-truth initialization, most of the primitive area stays on the true surfaces, as the precision in Table D.1 shows, and the learned geometry falls short only of the ground truth itself. GaussTwin assumes no known geometry, unlike RF-3DGS and RF-PGS. The Ground-truth init. and Fixed ground truth variants therefore serve only as diagnostics.
  2. Geometrically inaccurate primitives arise where the geometry adapts to the mismatch between the model and the data. Please see General Response G4 for when the recovered primitives lie on physical surfaces, how primitives that compensate for the mismatch can be interpreted and where the mismatch comes from.
  3. The departures concentrate near outer walls, floors and ceilings, while the primitives that stay on true surfaces carry most of the channel advantage of Ground-truth init. over Fixed ground truth (Appendix D.2). For this response, we also related the departures to unmodeled effects in Six-room. The channel gains from geometry learning have rank correlations of at most 0.11 with the energy share of omitted mechanisms, the interference difference in General Response G1 and the spectral change when constant reflectances replace the angle-dependent ones. The departures are therefore not tied to any single unmodeled effect at the link level.

Weakness 7. Appendix B now describes the RFID and BLE measurement hardware and acquisition.

Figure B.4 shows the BLE transceiver layout in top-down view, with the original floor plan in NeRF$^2$ Figure 12. For RFID, NeRF$^2$ Figures 2 and 9 show the measurement environment.

Appendix B.3 now describes how the measured spectrum is formed from the phase differences across the array.

Appendix B.4 now states that both benchmarks test feature prediction under real indoor multipath, while path-level angle–delay prediction is evaluated on the simulated channels in Table 1.

AuthorsOpenReview ↗

Response to Weakness 5 & Question 4

Weakness 5 & Question 4.

Regarding the sensitivity to measurement noise and position errors (W5 & Q4-1), please see General Response G3.

As for extending the measurement sparsity experiment to Six-room (Q4-2), Figure D.4 now includes it as dashed curves. Coverage and Chamfer-L1L_1 stay near full-data values down to 5% of training links. Precision and channel prediction degrade as data decrease. At 50%, both channel and geometry metrics remain close to full-data values.

Reviewer_aicvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a scene-specific wireless channel prediction framework that scene representation recovers explicit environment geometry from channel measurements. A differentiable model using oriented Gaussian surface primitives to predict channel characteristics at unmeasured Tx-Rx configurations. The learned primitives are natively interpreted as a reconstruction of propagation -relevant scene geometer without a specific reconstruction loss.

Critical strengths
  1. This paper proposed a connection between wireless channel prediction and geometry recovery, utilizing the same scene representation for both objectives.
  2. GaussTwin presented a coherent pipeline: the surface primitives have explicit geometric and propagation attributes, and forward model incorporates reflection, scattering, and visibility.
Critical weaknesses
  1. The paper shows that the learned representation can support both channel prediction and geometry recovery. But what is learned from this unified representation remains unclear. In particular, does the shared representation reveal when geometry is identifiable from communication measurements, which propagation cues are most informative, or when improving channel prediction should improve geometry? The current discussion tends to emphasize that both outputs are available rather than establishing a broader relationship between them.
  2. The paper argues that channel prediction and geometry recovery reinforce each other. But in Table D.1 shows that fixed ground-truth geometry produces worse channel prediction than the learned geometry in the six-room experiment. The relationship between the prediction quality and physical geometry could be more complicated than the paper suggests.
  3. The integration of main components (2D Gaussian primitives, differentiable propagation, back-projection-based initialization, and iterative refinement) is useful for both wireless channel prediction and surface recovery, but a clear comparison between prior wireless channel modeling and RF geometry-recovery methods can help to understand which part of components contributes most.
Questions for authors

What is the most important conceptual advance over prior methods that learn wireless fields or recover geometry from RF observations? Under what conditions should optimizing channel-prediction accuracy improve the physical accuracy of the learned surfaces? How should the learned primitives be interpreted when they compensate for propagation mechanisms omitted by the forward model?

1 comment
AuthorsOpenReview ↗

Response to Weaknesses & Questions

We thank Reviewer aicv for the careful reading of our paper and hope the responses below clarify each point.

Weakness 1, Questions 2 & 3. Since Reviewers inn7 and Mmib raised related questions, we answer these points in General Response G4, covering the conditions for accurate surfaces, the informative propagation cues and the interpretation of compensating primitives.

Weakness 2. The mutual reinforcement in Section 5.1 refers to the trained variants in Table D.1. Among them, channel prediction improves as the initial geometry becomes more accurate, from random through back-projection to ground-truth initialization. The geometry learned from the channel in turn outperforms the fixed back-projection initialization on every geometry metric. Fixed ground-truth geometry serves only as a diagnostic, since its channel errors reflect propagation effects that the model omits and that learned geometry partly absorbs (Appendix D.2).

Weakness 3 & Question 1. The most important conceptual advance is that our primitives are the physical surfaces of the propagation model. Methods that learn wireless fields, such as WRF-GS+, GSRF and nGRF, fit channels with radiance attributes on virtual sources that do not correspond to real objects. Methods that recover geometry from RF observations instead rely on dedicated radar waveforms, resolved multipath components or geometric supervision. Learned from position-labeled channel measurements alone, these surfaces serve both channel prediction and geometry recovery.

A like-for-like comparison is unfortunately not possible, because to our knowledge no prior method addresses both channel modeling and geometry recovery from channel measurements alone. Comparisons with methods dedicated to one of these tasks are possible, but they provide no evidence on which of our components contributes most. That evidence comes from the variants in Table D.1. Among them, the variant with fixed initial geometry degrades channel prediction most, while random initialization degrades the recovered geometry most. We also added a structured comparison with published wireless field methods in Appendix C.8.

Reviewer_MmibOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GaussTwin, a framework for predicting wireless channels and reconstructing scene geometry from wireless measurements. It represents the environment using oriented 2D Gaussian surface primitives and learns their geometric and electromagnetic properties through a differentiable propagation model. The model accounts for direct paths, reflections, scattering, and blockage.

The authors evaluate GaussTwin on simulated indoor and outdoor scenes and two real-world wireless datasets. The results show promising channel prediction accuracy and demonstrate that the learned representation can recover propagation-relevant surfaces in simulation.

Critical strengths

The paper addresses an interesting problem. Jointly modeling wireless propagation and scene geometry could be useful for building wireless digital twins.

The physics-informed Gaussian representation is interesting. Using the same surface primitives for reflection, scattering, blockage, and geometry reconstruction provides a more interpretable model than purely data-driven wireless representations.

The experimental results are encouraging, particularly for indoor channel prediction. The comparisons with existing methods and the reconstruction visualizations help demonstrate the potential of the approach.

Critical weaknesses
  • Why RF-only geometry reconstruction? While recovering geometry from wireless measurements is technically interesting, RGB-D, LiDAR, and existing 3D maps can often provide accurate geometric information more directly. Wireless sensing has unique strengths but is limited in spatial resolution and coverage. Combining visual and wireless sensing seems more practical in many settings. The authors should better justify the RF-only setting and compare against multimodal alternatives.

  • Novelty is not very high. The method builds on Gaussian splatting, differentiable ray tracing, and RF-based reconstruction, each of which has been considered in the past. The integration is interesting, but the novelty appears limited.

  • Insufficient real-world validation. Geometry reconstruction is evaluated only in simulation. It remains unclear how well the method handles real-world noise, calibration errors, and propagation-model mismatch. Real-world reconstruction results would substantially strengthen the paper.

  • The method requires known transceiver positions and extensive channel measurements (14,400 training links per simulated indoor scene). Its performance also degrades in more complex environments. The measurement cost, computational overhead, and robustness to imperfect measurements need further investigation.

Questions for authors

What scenarios would make most sense to reconstruct geometry from wireless measurements alone?

Could the proposed framework benefit from using visual geometry as an initialization or prior, while relying on wireless measurements to estimate electromagnetic properties? How would this compare with RF-only reconstruction?

3 comments
AuthorsOpenReview ↗

Response to Weakness 1-2 & Question 1-2

We thank Reviewer Mmib for the review and would first like to clarify the problem we address. GaussTwin predicts wireless channels at unmeasured links and recovers the surfaces that produce them, using only position-labeled channel measurements that existing wireless networks can provide.

Weakness 1, Questions 1 & 2. Regarding why RF-only reconstruction is useful and where it makes sense (W1 & Q1), we keep the RF-only setting because GaussTwin can reuse the base stations and user equipment of existing communication networks without additional sensors or configurations. This setting is central to integrated sensing and communication (ISAC), one of the six usage scenarios of 6G (IMT-2030), where reconstructing the environment from communication signals is an envisioned use case. Unlike cameras, RF measurements work without lighting, protect privacy and respond directly to the surfaces that shape propagation. RF-only reconstruction therefore suits deployed networks, particularly where cameras or LiDAR are undesirable.

As for using visual geometry as an initialization or prior (Q2), visual initialization or an existing geometric model could indeed improve GaussTwin further, as the ground-truth initialization in Table D.1 outperforms our method. Several methods already start from visual or existing geometry. For example, RF-3DGS and RF-PGS build on visual reconstruction. Differentiable ray tracing calibrates material parameters on a given scene model, while Photon Splatting attaches its primitives to a supplied scene. We deliberately depart from these methods and show that high channel prediction quality requires no visual initialization, while geometry itself can even be recovered to a certain degree from channel measurements. In every indoor scene, GaussTwin outperforms an adapted RF-3DGS initialized from ground-truth meshes on the channel metrics (Table A.1).

Weakness 2. GaussTwin differs from each of the three techniques. It is not a ray tracing method, because differentiable ray tracing calibrates the parameters of a given scene model while GaussTwin learns the scene itself from channel measurements. Gaussian splatting only inspired the geometric form of our surfel, while GaussTwin performs no graphics rasterization, deriving the channel response from wireless propagation itself. RF-based reconstruction has typically required dedicated radar waveforms, resolved multipath components or geometric supervision, whereas GaussTwin requires only position-labeled channel measurements. Its surfaces explicitly determine reflection points, path delays and blockage. Channel errors in turn can update those surfaces, which makes GaussTwin a new method.

AuthorsOpenReview ↗

Response to Weakness 4

Weakness 4. These practical aspects are important for deployment, and we address each of them below.

  • Measurement cost. The 14,400 training links are an evaluation setting, not a minimum requirement. The measurement sparsity experiment in Appendix D.5 shows that GaussTwin keeps channel prediction and coverage near full-data levels with 5% of the training links in Parallel-wall.
  • Complex scenes. Regarding the performance in more complex environments, GaussTwin indeed requires more primitives or higher-order propagation, as discussed in Section 5.1 with Appendices D.2 and D.3. However, GaussTwin outperforms all baselines on every channel metric in every indoor scene, including the most complex Six-room (Appendix A.2, Table A.1).
  • Computation costs. In Six-room with 512 primitives, the training time is about 42 minutes without dedicated performance optimization (Figure D.2(c)) on a single GPU. Inference takes only about 2.8 ms per link. Peak GPU memory is only about 7.1 GB in training and 2.5 GB in inference.
  • Robustness. To quantify the tolerance of GaussTwin to imperfect measurements, we added an experiment that trains it with UE position errors or measurement noise (Appendix D.6, Figure D.5; General Response G3). Channel prediction stays close to the clean results with a position error of 5 cm or at high SNRs.
AuthorsOpenReview ↗

Response to Weakness 3

Weakness 3. Please see General Response G2 for real-world reconstruction, G3 for the robustness to measurement noise and calibration errors and G4 for propagation-model mismatch.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response to All Reviewers (1/2)

We thank all reviewers for their feedback. Several questions were raised by more than one reviewer, so we answer them here once and refer to them from the individual responses. This first part covers G1 and G2, and the second part covers G3 and G4.

G1. Single-realization observations and phase averaging (Reviewers WC9E and inn7). GaussTwin predicts only power and averages over the reflection phases in Equations 7–8, while each training link provides one coherent realization. The new Appendix D.7 compares the coherent observations with the power sum of the traced paths, which contains no interference. Their median PDP difference is about −34 dB, far below the median PDP NMSE of GaussTwin in Table 1, and only about −27 dB even on poorly resolved Six-room links, where interference is strongest. When this power sum replaces the observations as the evaluation reference, the median PDP NMSE of the trained models changes by only 0.2–0.4 dB. In addition, the channel gains from geometry learning in Six-room are nearly uncorrelated with this difference (rank correlation 0.07), indicating that interference does not drive these gains.

G2. Real-world validation of surface recovery (Reviewers inn7 and Mmib). We agree that evaluating the reconstruction on real-world data would further increase the credibility of our method, and this is a direction we hope to help advance. Such an evaluation requires calibrated wideband multi-antenna channel measurements with reference scene geometry, which the public datasets known to us do not provide. Ray-traced data are standard practice in wireless communications research because calibrated channel measurements are costly and difficult to collect, and the difficulty grows further when reference geometry is needed. Some measured datasets pair wireless observations with geometric information, such as the DICHASUS series and the recent WiWorld-RealData dataset, but their radio configurations are not suited to geometry reconstruction.

  1. At half-wavelength spacing, the number of antennas sets the array length, which limits the angular resolution of propagation paths. WiWorld-RealData uses a single antenna at each end, which resolves no path angles, whereas our scenes use arrays at both ends.
  2. In the same way, bandwidth limits delay resolution, and paths separated by much less than its inverse cannot be resolved. The inverse of the 50 MHz in DICHASUS corresponds to about 6 m of path length, compared with about 0.3 m at our 1 GHz. At about 6 m, reflections within a room may fall into one delay bin.
  3. Both datasets operate below 7 GHz, whereas our scenes use 28 GHz. A reflection at 28 GHz therefore constrains the surface position within a smaller region, as the first Fresnel zone of a 10 m path reflected at its midpoint has a radius of about 0.16 m, compared with about 0.8 m at 1.272 GHz.

As is standard practice, we therefore evaluate surface recovery on ray-traced scenes with reference geometry (Table 1), while the measured RFID spectra and BLE received power test channel prediction (Section 5.2).

AuthorsOpenReview ↗

General Response to All Reviewers (2/2)

G3. Robustness to measurement imperfections (Reviewers inn7 and Mmib). We added an experiment that trains GaussTwin with UE position errors or measurement noise and evaluates it on clean test links (Appendix D.6, Figure D.5). Small position errors or high SNRs keep channel prediction close to the clean results. At low SNRs, the recovered geometry gains coverage but loses precision. For initialization, static element phase errors up to 15° or a bandwidth of 500 MHz have little effect, while narrower bandwidths or smaller arrays lower geometric accuracy.

G4. When the recovered primitives lie on physical surfaces (Reviewers aicv, inn7 and Mmib). Methods such as WRF-GS+, GSRF and nGRF treat their Gaussians as virtual sources, while GaussTwin takes the first step from such virtual sources to physical surfaces that channel observations can refine.

  • Conditions for accurate surfaces. The surfaces improve when the channel fit starts from a reasonable initialization such as back-projection, which requires only the channel data. From random initialization, the surfaces remain much less accurate, since the training channels underdetermine the geometry (Table D.1). In future work, we plan to refine and complete the reconstructed geometry with 3D generative models, which may in turn improve channel prediction further.
  • Informative propagation cues. First-order reflection may already constrain most planar surfaces, while diffuse scattering still improves surface recovery despite its small energy share. Second-order reflection matters mainly for channel prediction (Appendix D.2).
  • Compensating primitives. Without reference geometry, primitives that compensate for omitted mechanisms cannot be distinguished from physical surfaces, since both explain the observed channels through the same forward model. The learned primitives are therefore best interpreted as the surfaces most consistent with the channel observations under the modeled mechanisms. In our simulated scenes, such primitives form a minority near outer walls, floors and ceilings (Appendix D.2). In real measurements, the mismatch comes from effects that no model reproduces exactly, such as thermal noise, clock offsets or interference. Our simulated data are noise-free, yet Sionna RT 2.0.1 still gives different channels in repeated traces because of random path sampling. The mismatch also comes from the limited number of primitives and propagation mechanisms, which keeps the model tractable.

Dear reviewers,

Thank you for your reviews and constructive feedback.

The authors have submitted their rebuttal. Please take a moment to review it and consider whether your concerns have been adequately addressed. It would be helpful if you could briefly indicate:

  • Whether the rebuttal addresses your main concerns;

  • Whether any significant issues remain unresolved;

  • Whether your overall recommendation has changed.

If appropriate, please also update your review and recommendation in the system.

Thank you for your time and valuable contributions.

Best regards,

AC

Area_Chair_veKCOpenReview ↗

AC update

Dear authors,

Thank you for your time and contributions to the discussion.

I understand that you may be concerned if some reviewers have not provided follow-up comments or explicitly acknowledged how their concerns have been addressed in the rebuttal and revised manuscript. Please be assured that the AC will carefully consider all available information, including the original reviews, author rebuttal, revised manuscript, and discussion comments, as part of the final assessment.

The absence of further reviewer comments does not mean that the authors’ responses and revisions have been overlooked. I will carefully examine the issues raised, assess the authors’ clarifications and revisions, and take the discussion into account when forming my recommendation. I will also consult with the reviewers as appropriate to ensure that the relevant issues are fully considered.

Thank you again for your constructive engagement and contributions to the review process.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.