acceptodds
Under review as a conference paper at ICLR 2027

SphGaS: Spherical Gaussian Splatting for RF Spectrum Synthesis under Varying Transmitter–Receiver Configurations

Abstract

This paper proposes SphGaS, a novel Gaussian splatting (GS)-based framework for rendering transmitter–receiver-conditioned (TX-RX) RF spectra directly on the spherical manifold. Existing GS-based methods typically synthesize planar spectra. Geometrically, this is not consistent with the practical spherical structure of the RF spectra, consequently leading to distorted Gaussian responses. Meanwhile, some methods model spectral variations at one transceiver endpoint only, neglecting the coupled variations induced by the TX and/or RX movement. To the best of our knowledge, SphGaS is the first framework that jointly addresses geometric misalignment and TX-RX mobility. It introduces spherical Gaussian primitives that deform with the TX-RX configuration and are rasterized directly in the spherical domain. This design preserves the geometry of Gaussian responses while efficiently capturing the coupled spectrum variations induced by TX–RX movement. With a customized CUDA rasterizer, SphGaS keeps fast rendering while substantially improving the modeling fidelity. Experiments show that SphGaS improves PSNR by 1.49 dB and achieves up to 40X faster rendering than state-of-the-art GS-based methods on real-world RFID datasets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Gaussian splatting, RF spectrum, Spherical geometry

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_fZfiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SphGaS for RF spatial-spectrum synthesis under changing TX and RX positions. It represents spectra using spherical Gaussian primitives, adjusts their geometry and complex attributes based on the TX-RX configuration, and renders the result with a custom CUDA rasterizer. The method also uses bidirectional link pairs during training. Experiments cover measured RFID spectra, simulated indoor and outdoor scenes, and downstream RF tasks.

Critical strengths
  1. The spherical Gaussian formulation addresses distortions and boundary artifacts introduced by planar spectrum rendering. The custom rasterizer also keeps spectrum synthesis efficient.

  2. SphGaS improves reconstruction over the reported baselines on several datasets. The paper includes useful ablations of the spherical operations, deformation components, and training choices.

Critical weaknesses
  1. The novelty over prior RF Gaussian methods is not clearly distinguished. Earlier work has already modeled complex RF responses over spherical directions, while other Gaussian-based methods have considered receiver-dependent synthesis with changing TX and RX positions. SphGaS adds intrinsic spherical evaluation and geodesic deformation, but its ablations suggest that much of the improvement comes from spherical parameterization itself. The paper also lacks a simpler baseline with fixed spherical primitive geometry and TX-RX-conditioned complex attributes. Without this comparison, it is difficult to tell what the full geometric deformation model adds beyond the existing ideas.

  2. The generalization evaluation does not test receiver positions entirely absent from training. The standard split holds out TX-RX links, while the unseen-TX experiment removes transmitter positions but retains receiver locations represented during training. The paper does not test new receiver positions or links where both endpoints are unseen. This leaves an important part of the claimed TX-RX configuration generalization unexamined.

Questions for authors
  1. What is the main difference between SphGaS and earlier approaches to spherical RF modeling and receiver-conditioned Gaussian-based synthesis, beyond the intrinsic spherical operations?

  2. Can the authors compare with spherical Gaussians whose centers, scales, and orientations are fixed but whose complex attributes depend on TX and RX positions, under a similar training and parameter budget?

  3. Have the authors tested receivers entirely excluded from training, including links where both TX and RX positions are unseen?

  4. How are Gaussian supports assigned to CUDA tiles near the azimuth seam and pole, and how is the custom backward pass implemented? Could the authors share anonymous rasterizer and training code, or provide sufficient implementation details to reproduce the results?

7 comments
AuthorsOpenReview ↗

Response to reviewer fZfi (1/5)

Dear Reviewer fZfi, thank you for highlighting the spherical formulation, efficient rasterizer, and component ablations. Your review asks (1) what distinguishes SphGaS from earlier spherical and receiver-conditioned RF methods, (2) whether geometric deformation improves on fixed spherical geometry with conditioned attributes, (3) whether RX positions and both endpoints can be unseen, and (4) how tile assignment, seam/pole handling, backward computation, and code availability work. We answer these in order: we compare the representation space and joint reciprocal conditioning; report the requested fixed-geometry control alongside the existing ablations; separate the main, seen/unseen-TX, and RX-extrapolation splits; and provide CUDA implementation details and anonymous code. The relevant evidence is in main Table 2 and Fig. 3; Appendix Figs. 7–9, Table 9, and Section F.3.

AuthorsOpenReview ↗

Response to reviewer fZfi (2/5)

Q1. What is the main difference from prior RF Gaussian methods?

What is the main difference between SphGaS and earlier approaches to spherical RF modeling and receiver-conditioned Gaussian-based synthesis, beyond the intrinsic spherical operations?

The key difference beyond intrinsic spherical operations is how joint TX–RX variation is organized. Prior spherical RF synthesis still derives the received spectrum from 3D scene Gaussians, while receiver-conditioned Gaussian methods express link changes through scene-space or planar representations. SphGaS instead uses the RX sphere as a shared space for arrival clusters, coupling their changes across both endpoints through canonical deformation and reciprocal-pair training. We explain this through representation space and cross-link coupling.

(i) Representation space. Scene-space RF Gaussian methods such as WRF-GS+, GAI-GS, and GSpaRC project 3D primitives to angular outputs. GSRF also uses 3D Gaussians and an ERP receiving plane; SwiftWRF uses a planar angular map. In SphGaS, the primitives instead represent arrival clusters on the RX sphere itself (Appendix Fig. 9 and Table 8). GSBF uses spherical kernels for a different task, beamforming.

(ii) Cross-link coupling. In scene-space synthesis, RX motion changes how the field is observed, while TX motion changes the field itself; these effects are handled through the field and its rendering. On the RX sphere, both appear as changes in arriving signal clusters. SphGaS therefore deforms the same canonical supports for both endpoints, modeling the effect of six-dimensional TX–RX configurations as structured variation on the two-dimensional receiving manifold. Forward and reverse links train this shared network as reciprocal pairs (main Fig. 3; Appendix Figs. 7–8 and Table 11). This couples predictions across links without assuming that each primitive is a persistent 3D path. We invite reviewers to examine the visual motivation in Appendix Figs. 7–8 and the cross-method primitive-response shapes on the receiver sphere in Appendix Fig. 9.

AuthorsOpenReview ↗

Response to reviewer fZfi (3/5)

W1 / Q2. What does spherical deformation add?

Can the authors compare with spherical Gaussians whose centers, scales, and orientations are fixed but whose complex attributes depend on TX and RX positions, under a similar training and parameter budget?

(i) Role of deformation. As TX/RX positions change, the directions and extents of arriving signal clusters can change; center, scale, and orientation deformation let the spherical supports follow those changes. Centers can move beyond the observed hemisphere, allowing a cluster to disappear from the received spectrum after blockage (Appendix Section B, especially Fig. 8).

(ii) Interpreting the existing ablation. The modest gains in Appendix Table 13 reflect controlled changes to individual intrinsic operations within an otherwise spherical model. Spherical representation and manifold-consistent deformation work together, so removing their shared geometry would change the representation itself. Main Table 4 separately tests deformation components.

(iii) Fixed-geometry control. To test the requested simpler baseline directly, we optimized shared spherical Gaussian centers and covariances, kept them fixed across links, and let the MLP predict TX–RX-conditioned complex attributes absolutely on RFID S23. This additional response experiment uses the same primitive count and training updates as the reference:

RFID S23 variantMean PSNR (dB)Median SSIM
Canonical + geometric/attribute deformation26.240.9266
Fixed optimized centers/covariances + absolute attributes21.590.8082

The 4.65 dB difference shows that fixed geometry with conditioned complex attributes does not match the full model in this matched response-only run; the reference score is not the main paper's headline RFID number. Because the controls also differ in attribute parameterization, this is evidence for the full construction rather than an isolated geometric-deformation effect; exact parameter-count parity has not been established.

Reviewer_AXJDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a method for synthesizing RF angular spectra conditioned on transmitter and receiver positions. It learns canonical Gaussian primitives on the receiver's sphere and uses a deformation network to adjust their angular centers, shapes, and complex amplitudes for each TX–RX configuration. Evaluation includes measured RFID data and ray-traced environments, with additional experiments on RSSI prediction, transmitter placement, and directional propagation analysis. The results show that SphGaS improved reconstruction by 1.49 dB over several baselines.

Critical strengths

The paper implements this representation efficiently and evaluates it across both measured and simulated data. The evaluation provides useful evidence about reconstruction quality and generalization within static scenes.

Critical weaknesses

1. Contributions are not sufficiently distinguished from existing techniques.

The paper motivates SphGaS through two limitations: geometric distortion from planar rendering and restricted TX–RX mobility. For the first, spherical and panoramic Gaussian rendering have already been explored in methods such as 360-GS and UniTriSplat, yet neither is cited, discussed, or compared with. SphGaS's directly learned angular primitives differ from these methods’ projected 3D scene primitives, but the paper does not adequately establish the technical advance beyond existing approaches to spherical rendering.

For joint TX–RX mobility, the main idea is to conditioned the network on TX-RX positions. This extension seems quite straightforward, and can be easily applied to prior methods that condition only on RX position. Conditioning on RX position has much stronger physical explanation, as the network is learning a static field while encoding multi-view consistency. SphGaS instead directly predicts changes in angular primitives as both endpoints move, without an equivalent explicit connection to shared physical geometry. Appendix B provides a useful local physical motivation for Gaussian-shaped angular responses, but this does not establish the physical consistency of the learned deformation across configurations. The canonical primitives need not correspond to persistent paths or scatterers, and their centers, shapes, and complex amplitudes can all change through the network.

2. The evaluation does not sufficiently establish generalization and practical value under realistic deployment conditions.

The method learns a scene-specific mapping from TX-RX positions to spectra without explicitly enforcing consistency with a shared 3D propagation scene. The dedicated indoor experiment uses 7,200 training links per scene; without the spatial distribution and separation of training and test locations, it remains unclear whether the model captures propagation structure that generalizes beyond nearby configurations or primarily interpolates densely sampled measurements.

The position-only conditioning also leaves important physical variables unspecified. TX orientation can change the radiation pattern. RX orientation affects the relationship between local angular coordinates and world directions. What defines zero azimuth, elevation, and the array's reference axes? If these quantities vary independently of position, the same model input can correspond to different spectra.

Finally, an approximately 1.5 dB improvement in reconstruction PSNR does not show a consequential application benefit. For online communication applications, scene dynamics may invalidate a model trained for a fixed environment. Collecting thousands of spectra also raises questions about acquisition time.

Questions for authors
  1. How are training and test TX/RX positions distributed, and how far are test configurations from the nearest training configurations? What do the existing experiments establish beyond local interpolation, and what measurement density is required to achieve useful accuracy?

  2. What are the TX/RX orientations and how would their changes affect the model? Is the reception sphere expressed in a world frame or an RX-local frame, and what defines its zero azimuth and elevation axes?

  3. How sensitive is the learned field to the dynamics in the scene? or in the other words, how frequently does the scene need to be re-measured?

  4. Since the primitives are defined directly on the angular sphere, is SphGaS best understood as a TX–RX-conditioned mixture of spherical Gaussian kernels? What additional constraint or modeling benefit does the canonical-template-plus-deformation construction provide over directly predicting the centers, shapes, and complex coefficients of such a mixture?

6 comments
AuthorsOpenReview ↗

Response to reviewer AXJD (1/5)

Dear Reviewer AXJD, thank you for recognizing the value of the implementation and the measured and simulated evaluation. You raise four connected questions. (1) Sampling and generalization: we distinguish the main new-link interpolation split from an RX-region extrapolation test, and report training densities, nearest-link distances, and training-size results. (2) Orientation: we specify the world-to-local input transform and RX-local angular axes, then test held-out RX orientations. (3) Scene dynamics and practical use: we measure indoor changes and a quasi-static outdoor vehicle, test model updates, and state what these results cannot establish about re-measurement frequency or online throughput. (4) Model construction: we explain why a shared spherical template with TX–RX-conditioned residual deformation links spectra across configurations, and compare it with fixed geometry and predicted complex attributes. The tables and figures cited under each question locate the evidence.

AuthorsOpenReview ↗

Response to reviewer AXJD (2/5)

W2 / Q1. Training/test distribution, nearest configurations, and measurement density

How are training and test TX/RX positions distributed, and how far are test configurations from the nearest training configurations? What do the existing experiments establish beyond local interpolation, and what measurement density is required to achieve useful accuracy?

(i) Setting, density, and nearest links. Each main scene has 4,800 training and 1,200 test TX–RX links. TX and RX positions vary within the same sampled region; every test pair is new. In the available coordinate splits, every test RX position is unseen, while a few indoor test links reuse a TX position. Thus the main split establishes new-link in-region interpolation. Appendix Table 9 separately tests new RX positions at seen TXs and links from unseen TXs; its 7,200 training links are an auxiliary protocol.

The following table reports training links per sampled room volume (indoor) or RX footprint (outdoor). Nearest-link distance is the Euclidean distance in joint TX–RX coordinates from each test pair to its closest training pair:

SceneTraining-link densityNearest distance: median / 90th percentile (m)
Bedroom3.08 links/ft³0.68 / 0.88†
Conference room12.42 links/ft³0.43 / 0.55†
Office1.58 links/ft³0.85 / 1.11†
Munich0.01546 links/ft²8.33 / 12.00
Etoile0.01214 links/ft²≈8.9 / ≈12.8‡
Florence0.01201 links/ft²≈8.9 / ≈12.8‡

† The indoor distances come from available filtered splits with 4,386–4,788 training links and are indicative. Munich uses the matching 4,800/1,200 split. ‡ Etoile and Florence are rough estimates scaled from Munich by the fourth root of their RX-footprint area ratio, assuming similar planar sampling; their split-level distances have not been measured. The outdoor TX footprint corresponds to 0.04463 training links/ft² in each city. For scale, the public real-world RFID study GSRF reports about 15.5 training measurements/ft³; our separate 4,898-link RFID training split corresponds to about 17.71 links/ft³ under the same volume normalization. These densities describe paired links per region, not uniform spatial spacing or identical measurement difficulty.

(ii) Training-set size. Main Fig. 6A compares RFID training sizes. We also retrained SphGaS and two baselines on four nested, spatially balanced indoor subsets, testing on the same 1,200 links (mean PSNR in dB / median SSIM):

Training linksSphGaSNeRF²WRF-GS
80019.13 / 0.62415.02 / 0.49412.42 / 0.380
1,60021.05 / 0.72017.26 / 0.54013.61 / 0.441
2,40021.63 / 0.74319.06 / 0.58713.96 / 0.475
4,80023.83 / 0.79819.11 / 0.59315.35 / 0.534

This gives a practical budget–accuracy tradeoff: in this indoor test, 2,400 training links yield 21.63 dB / 0.743 median SSIM, while 4,800 yield 23.83 dB / 0.798. A user can choose a sampling budget against the accuracy needed for their task; we do not claim that one link density guarantees useful accuracy in every scene.

(iii) Extrapolation. Although the main train and test TX–RX configurations are disjoint, they occupy the same spatial region and therefore test interpolation. For the added test, we start from uniformly sampled TX–RX links and partition them at an RX x-coordinate boundary: 4,800 links have training-side RX positions, while 1,200 have test-side RX positions. This makes the RX regions disjoint, with every test RX outside the training RX convex hull. TX positions still vary in both groups and are not spatially held out; the test isolates RX-region extrapolation while TX also varies across links. We evaluate the fixed final training step without test-based checkpoint selection (main Table 2):

MethodPSNR (dB)Median SSIMS-SSIM
SphGaS19.440.67560.6801
NeRF²16.930.54110.5395
WRF-GS12.230.44220.4227

SphGaS gains 2.51 dB over NeRF² and 7.21 dB over WRF-GS on this RX holdout.

Taken together, the split and nearest-link distances characterize the tested interpolation regime; main Fig. 6A and the added training-size study provide a reference for selecting measurement density; and main Table 2 supports extrapolation over this tested RX region. They do not establish arbitrary joint TX–RX extrapolation. Physical acquisition time was not measured.

AuthorsOpenReview ↗

Response to reviewer AXJD (3/5)

W2 / Q2 / Deployment concern. Coordinate frames and orientations

What are the TX/RX orientations and how would their changes affect the model? Is the reception sphere expressed in a world frame or an RX-local frame, and what defines its zero azimuth and elevation axes?

(i) Coordinate convention. We apologize that this was unclear. Main Section 3 now defines TX/RX world positions and their separate world-to-local rotations: the MLP receives the transformed coordinates, and the spectrum's zero azimuth/elevation axes are those of the RX-local array frame. The main datasets hold orientations fixed. This input convention allows orientation changes to affect conditioning, although different physical poses could map to the same transformed coordinates; such collisions are unlikely in the sampled position sets.

(ii) Orientation test. We formed 27 training orientations from distinct yaw–pitch–roll combinations of the virtual RX array, then evaluated those 27 seen and 26 additional held-out orientations (1,200 links per pose; Appendix Table 17):

MethodSeen orientation: PSNR / median SSIMUnseen orientation: PSNR / median SSIM
SphGaS22.60 / 0.76622.31 / 0.760
NeRF²16.37 / 0.55816.34 / 0.550
WRF-GS14.08 / 0.54414.06 / 0.539

The 0.29 dB decrease supports generalization within the tested RX-pose range after multi-orientation training; it does not establish TX-orientation generalization or rotation invariance of the original model.

Reviewer_fZgiOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors propose SphGaS, a Gaussian Splatting-based framework for RF spectrum reconstruction and synthesis under varying transmitter–receiver (TX–RX) configurations. The main innovation is to represent RF spectra directly on the spherical manifold using learnable spherical Gaussian primitives, rather than relying on conventional planar representations. The framework also incorporates TX–RX-conditioned deformation to capture spectrum variations across different configurations. Experiments on real-world and simulated RF datasets demonstrate improved reconstruction quality compared with several existing methods.

Critical strengths

One strength of this paper is its extension of conventional planar Gaussian Splatting to spherical geometry, which better matches the directional nature of RF spectra and helps reduce geometric distortion. This is a well-motivated design choice for RF spectrum modeling. In addition, the paper presents strong experimental results, with SphGaS achieving better reconstruction quality than several existing baselines on both real-world and simulated datasets.

Critical weaknesses

One concern is that the proposed method is primarily evaluated in fixed environments, which raises questions about its generalization to practical wireless scenarios. Wireless propagation is highly sensitive to environmental material properties, such as permittivity and conductivity, as well as changes in scene geometry. However, the paper does not sufficiently investigate how these factors affect the learned RF spectrum representation. It would be helpful if the authors could discuss the robustness of SphGaS to such environmental variations and clarify whether changes in material properties or scene geometry would require retraining the entire model.

Questions for authors
  1. How robust is SphGaS to changes in environmental material properties, such as permittivity and conductivity, or changes in scene geometry?

  2. Would such environmental changes require retraining the entire model, or could the learned representation be efficiently updated?

3 comments
AuthorsOpenReview ↗

Reponse to reviewer fZgi (1/3)

Dear Reviewer fZgi, thank you for your positive assessment of the spherical formulation, presentation, and experimental results. You ask (1) how geometry and material changes affect the learned spectrum and (2) whether the model must be retrained after such changes. We answer the first with controlled indoor furniture and material changes, a chair-distance sweep, and five quasi-static car placements in Florence. We answer the second with network fine-tuning: it partially recovers quality after furniture movement, whereas the same update does not help under the tested material substitutions. Main Table 5, Appendix Fig. 14, and Appendix Tables 14–16 provide the corresponding results and setup.

AuthorsOpenReview ↗

Reponse to reviewer fZgi (2/3)

Q1. Sensitivity to geometry and material changes

How robust is SphGaS to changes in environmental material properties, such as permittivity and conductivity, or changes in scene geometry?

We tested three controlled settings with the trained model fixed: (i) furniture re-arrangement and displacement in an indoor room, (ii) a vehicle at several static positions in an outdoor scene, and (iii) indoor furniture material replacement. We evaluated each changed scene on the same held-out links as its unchanged baseline.

(i) Indoor geometry. In a 6 × 5 × 2.8 m editable room at 3.5 GHz, we relocated selected furniture and evaluated the same 1,200 test links. Each of the one-, two-, three-, and four-object conditions averages three randomly sampled layouts. The results are in main Table 5; the affected links are visualized in Appendix Fig. 14.

Change from original sceneMean PSNR (dB)Drop (dB)Links losing >1 dB
None23.88700%
Chair moved23.7420.14515.61%
Chair + table moved22.9760.91233.75%
Chair + table + cabinet moved22.2951.59342.42%
Above + bed moved21.6522.23648.42%

We also moved one chair along three directions, evaluating the same 1,043 valid links at each distance. The table reports averages across directions:

Chair displacement (m)Mean PSNR drop (dB)Links losing >1 dB
0.100.0285.50%
0.250.0327.70%
0.500.05310.77%
1.000.06412.56%
1.360.06913.61%
2.730.12115.53%

Indoor changes are link-specific and become more pronounced as furniture moves farther or more objects are rearranged. One moved chair reduces mean PSNR by only 0.145 dB, yet 15.61% of links lose over 1 dB; with four moved objects, the mean drop is 2.236 dB and 48.42% of links lose over 1 dB. The full distance sweep is in Appendix Table 14.

(ii) Outdoor vehicle. We consider a simple, localized change: one car placed at five positions in Florence. These are quasi-static snapshots, without continuous vehicle motion. With TX at 35 m and RX at 1.5 m, we froze a car-free model and evaluated the same 1,200 links at ground-level car positions 3 m apart. car_0 is the first placement; the baseline contains no car. Parentheses show PSNR drops from that baseline (negative means improvement). See Appendix Table 16 for the full setup:

Car state and ground position (m)RX 1.5 m: PSNR (drop), dB
Car-free18.1302 (0)
car_0 (2, 8)18.1246 (+0.0057)
car_1 (5, 8)18.1285 (+0.0017)
car_2 (8, 8)18.1307 (−0.0004)
car_3 (11, 8)18.1298 (+0.0005)
car_4 (14, 8)18.1314 (−0.0012)

Across the five placements, PSNR shows no clear monotonic trend and the mean drop is only 0.0013 dB. In summary, this outdoor perturbation is much less disruptive than the tested indoor re-arrangements; individual links can still be affected, with one losing 6.30 dB. A vehicle moving continuously can also introduce time-varying paths and Doppler. Modeling those effects, as well as major urban changes, is outside the present study and is subject to future work.

(iii) Indoor material properties. We kept furniture geometry fixed and replaced the surfaces of four wooden items together. Sensitivity results are in main Table 5 and material parameters in Appendix Table 15:

Replacement materialRelative permittivityConductivity (S/m)Mean PSNR drop (dB)Links losing >1 dB
Plasterboard2.730.0275790.27123.33%
Concrete5.240.1230870.83927.58%
Glass6.310.0192761.12341.17%
Metal1.0010,000,0003.34163.92%

The wooden baseline is 23.887 dB (relative permittivity 1.99; conductivity 0.017998 S/m). Some material changes affect many links, especially metal: mean PSNR falls by 3.341 dB and 63.92% of links lose over 1 dB in PSNR. A change in attenuation and reflection along many paths could explain this wider effect. Permittivity and conductivity change together here, so their separate roles are not identified. For spatially averaged tasks such as coverage estimation, a local object change might have a smaller effect than on an individual link.

AuthorsOpenReview ↗

Response to reviewer fZgi (3/3)

Q2. Updating the learned representation

Would such environmental changes require retraining the entire model, or could the learned representation be efficiently updated?

We froze the canonical Gaussians and fine-tuned only the deformation MLP for 24,000 steps on 4,800 changed-scene links. Main Table 5 reports the paired changes:

Changed sceneOne objectTwo objectsThree objectsFour objectsPlasterboardConcreteGlassMetal
Fine-tuning PSNR change (dB)+0.042+0.176+0.198+0.242−0.297−0.252−0.278−0.069

These changes come from matched update runs. Their starting checkpoints differ from those used for the separate frozen-model zero-shot averages in main Table 5, so the changes should not be added directly to the displayed zero-shot PSNR values.

Geometry motion is partially adaptable: the fine-tuning gain rises from +0.042 to +0.176, +0.198, and +0.242 dB as more objects move. Material shifts are harder: the same MLP-only update reduces final PSNR in all four tests. Re-learning the canonical response or full retraining may help, but neither was tested. These results show a partial update for movement, not few-shot or full-accuracy recovery. The movement visualization is in Appendix Fig. 14.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Common response

We thank the reviewers and area chairs for their careful reading and constructive feedback.

Summary. SphGaS represents complex RF spectra with Gaussian primitives directly on the receiver sphere and learns how their supports change under joint TX–RX mobility. The paper combines this representation with reciprocal-pair training and a spherical CUDA rasterizer, and evaluates reconstruction on measured and simulated RF data.

The reviewers recognized several strengths:

  • Spherical modeling: a “well-motivated design choice” with “strong experimental results” (Reviewer fZgi).
  • Measured and simulated evaluation: “useful evidence about reconstruction quality and generalization within static scenes” (Reviewer AXJD).
  • Efficient implementation and ablations: the “custom rasterizer also keeps spectrum synthesis efficient,” alongside useful component studies (Reviewer fZfi).

The revision addresses the main questions with results that can be located directly in the manuscript; new material is marked in blue:

  • Scene changes and model updates (Reviewers fZgi, AXJD): indoor movement and material tests appear in main Table 5; the link-level visualization and distance, material, outdoor-car, and RX-orientation details are in Appendix Fig. 14 and Tables 14–17.
  • New configurations and sampling density (Reviewers AXJD, fZfi): main Fig. 6A studies training size; main Table 2 tests RX-region extrapolation; Appendix Table 9 separates seen-TX/new-RX from unseen-TX synthesis. The individual responses also report training densities and nearest-link distances, with estimates labeled as such.
  • Modeling contribution (Reviewers AXJD, fZfi): main Fig. 3 and Appendix Figs. 7–9 illustrate the reception-space formulation, deformation, and geometric difference from projected methods. We also report the requested fixed-geometry/conditioned-attribute control in the individual responses.
  • Implementation (Reviewer fZfi): Appendix Section F.3 and Algorithm 1 now describe disk-grid rendering, conservative tile assignment, spherical evaluation, resampling, and backward computation; the anonymous code is available.

The responses below quote each reviewer question and give the corresponding evidence and limits.

AuthorsOpenReview ↗

Common Response: Additional Experiments Following Reviewer fZfi's Feedback

Dear all,

Following the latest feedback from Reviewer fZfi, we have carefully reconsidered the remaining concerns and conducted two additional experiments. We would like to share these updates with all reviewers, as they also provide further evidence regarding the contributions and generalization capability of SphGaS.

  1. Spherical representation and deformation: We implemented an additional 3D Gaussian baseline that retains Euclidean deformation while adopting the same spherical mapping (ExactMap) as SphGaS. This experiment helps distinguish the contribution of spherical mapping from that of our full spherical formulation. We also clarified how the existing deformation ablations in Main Table 4 complement the previously requested fixed-geometry control.

  2. RX-region extrapolation: Following Reviewer fZfi's suggestion, we additionally evaluated SwiftWRF under the same held-out RX-region extrapolation protocol, providing further evidence of generalization to unseen receiver regions.

The detailed results and discussions are available in our latest response to Reviewer fZfi. We sincerely appreciate all reviewers' constructive feedback, which has helped us further strengthen our experimental evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.