acceptodds
Withdrawn from ICLR 2027

CoRF: Cross-Scene RF Synthesis by Learning Propagation and Preserving Array Physics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing radio-frequency (RF) neural fields fit each scene separately, making new-scene deployment measurement- and optimization-intensive. This work studies amortized cross-scene spatial spectrum synthesis, where a shared model learns propagation across scenes and instantiates an unseen scene from sparse target-scene measurements without scene-specific training. To achieve this, CoRF separates learned scene-dependent propagation from the known receiver-array observation model. An unordered set of spectrum-only references conditions a canonical anchor field, producing arrival directions and query-dependent component powers for each query. Analytic array physics maps these components to the receiver covariance and then to the spatial spectrum. This factorization keeps the pretrained propagation model frozen, enabling synthesis at arbitrary query locations after a single reference-conditioning pass. A necessary local reference-capacity bound and an error decomposition further characterize the formulation. Across 35 simulated scenes spanning seven categories, CoRF outperforms the strongest baseline by 7.09 dB PSNR on unseen variants of represented scene categories and by 6.97 dB on entirely unseen scene categories.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Radio-frequency (RF) neural fields, spatial spectrum synthesis, cross-scene generalization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_f8FjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes CoRF for cross-scene RF spatial spectrum synthesis from sparse target-scene measurements. The method constructs a reusable scene representation from reference spectra, predicts propagation components, and renders spatial spectra through an analytic receiver-array model. Optional power refinement and receiver calibration support target-domain adaptation. Experiments cover 35 simulated scenes and one publicly available real-measurement scene.

Critical strengths

Separating learned propagation from the known receiver-array response is well motivated. The ablation results support the usefulness of the analytic renderer. The method achieves substantial improvements over the evaluated baselines under sparse-reference settings. The paper provides extensive supplementary analyses, including reference dependence, receiver calibration, rendering approximations, and computational efficiency.

Critical weaknesses
  1. The paper is very hard to follow because of the lack of a clear explanatory narrative for its method. The presentation moves quickly from the stated challenges to a sequence of equations. The motivation for combining canonical anchors, bearing-aligned sampling, and query-conditioned emission is difficult to follow by using the equations. Implementation-specific terminology, such as the “zeroth-order” emission component in Eq. (7) , is not clearly explained when introduced. Nowadays, AI can generate lengthy and seemingly impressive mathematical arguments. Even if it could produce proofs worthy of a Fields Medal, making the mathematics understandable to human readers would remain essential for a conference paper. A clearer progression from the problem to model design is much better than a long and math equations that beyond human understanding capability.
  2. Real world validation is weak, leaving concerns about overfitting to the benchmark settings. The real-data evaluation uses only one publicly available measured scene from NeRF^2, without independent measurements collected in the authors’ own laboratory. As coding agents make implementation and benchmark optimization increasingly accessible, I expect a convincing wireless systems paper to demonstrate practical effectiveness through an independently operated testbed. Improving prediction metrics on a limited dataset does not establish that the method will work reliably in new deployments. Additional experiments in independently measured environments are needed to assess whether the reported gains extend beyond dataset-specific regularities.
  3. The practical communication benefits remain insufficiently validated. It does not evaluate communication performance through metrics such as CQI, throughput, BER/BLER, or end-to-end communication latency. Consequently, it remains unclear whether the improvements in spectrum reconstruction translate into better link quality. A link-level or testbed evaluation would provide stronger evidence for the claimed practical benefits.
Questions for authors

I have no specific questions for the rebuttal. My concerns are outlined in the weaknesses above.

Reviewer_E6jGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CoRF for cross-scene RF spatial spectrum synthesis from sparse target-scene measurement. CoRF predicts an RF spatial spectrum and its uncertainty for a transmitter–receiver configuration query. It proposes to learn scene-dependent propagation through canonical spatial anchors while preserving receiver-array physics through an analytical covariance and Bartlett-spectrum renderer. Experiments on 35 simulated scenes and one real RFID dataset show the results of spectrum reconstruction, angle-of-arrival estimation, and beam selection.

Critical strengths
  • The paper presents a method to generate an RF spatial spectrum and its uncertainty for a transmitter–receiver configuration query. Importantly, the separation of the scene-dependent propagation and the receiver-array physics makes it possible to adapt to a new scene without retraining.

  • The uncertainty-aware design is appreciated. CoRF learns complex scene-dependent propagation but preserves receiver-array physics analytically, enabling adaptation to different arrays and producing uncertainty estimates useful for judging prediction reliability.

Critical weaknesses
  • Limited realism of the simulated evaluation and potentially mismatched real-world validation.
  • Narrow frequency and array coverage.
  • Marginal improvement on real measurements.
  • Lack of evaluation in dynamic environments.
Questions for authors

Most experiments are conducted on 35 ray-traced scenes generated using Sionna RT. Although ray tracing provides more realistic propagation than simplified channel models, a substantial sim-to-real gap remains. The Sionna channel generations use simplified material and reflection models, omit several hardware impairments and measurement noise sources, and do not capture environmental dynamics. Therefore, strong performance on these simulated scenes does not necessarily translate to real-world RF systems.

The real-world evaluation uses only one scene from the NeRF2 RFID dataset. This dataset represents a RFID backscatter setup involving a reader/exciter, a passive RFID tag, and a separate USRP receiver array, whereas the paper models a more conventional transmitter–receiver link. Treating the tag as the effective transmitter may be a useful approximation for angular-spectrum prediction, but the received signal also depends on the reader-to-tag channel, tag response, and backscatter-specific noise and attenuation. The paper should better justify this modeling choice and validate the method on conventional communication or sensing measurements.

Both the simulated and measured datasets use similar sub-GHz carrier frequencies (915/920 MHz) and a 4x4 antenna array. This configuration is not representative of many modern communication and sensing systems. The limited bandwidth constrains range resolution, while half-wavelength antenna spacing at this frequency results in a relatively large physical array. On 915MHz with 4 antenna elements, it is likely impossible to generate a meaningful 64-direction beam codebook. The beam-management study in Section 4.7 therefore remains questionable, and the evaluation should include more representative bands and configurations, such as 2.4 GHz, 5 GHz, 28 GHz, or 60 GHz.

The gains on the real dataset are considerably smaller than those reported in simulation. In Table 6, the strongest CoRF variant improves PSNR over the strongest baseline by only about 0.5 dB. Because the real-world evaluation contains only one scene and the best variant uses target-scene calibration and power refinement, the evidence is insufficient to establish consistent superiority over SOTA in realistic wireless environments.

The evaluation considers only static scenes and does not include people or other moving objects. Because such dynamics are common in practical communication and sensing environments and can substantially alter RF propagation, it is important to evaluate the model’s robustness to moving objects and time-varying channels.

Reviewer_2rPZOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CoRF for cross-scene spatial spectrum synthesis from sparse target-scene measurements. The main idea is to use an unordered set of reference spectra to condition a canonical anchor field. The conditioned anchors determine the arrival directions, while the query geometry controls the component powers and power variances. These components are converted to receiver covariance and spatial spectra using an analytic array model and Bartlett rendering. The paper also introduces optional component-power refinement and receiver calibration. Experiments on 35 simulated scenes and one measured scene evaluate unseen scene variants, unseen scene categories, receiver-array adaptation, beam management, localization, uncertainty, and simulation-to-real transfer.

Critical strengths

The problem formulation and physical factorization are well motivated. Learning scene-dependent propagation while keeping the known receiver-array response analytic is a reasonable design for cross-scene RF synthesis. The analytic renderer also provides a clear way to change receiver configurations and propagate component-power uncertainty. Replacing it with a learned renderer reduces PSNR by about 6.8 dB in both generalization settings, which provides strong evidence for this design choice.

The evaluation is broad and comprehensive. Feed-forward CoRF improves PSNR over the strongest baseline by 7.09 dB on unseen scene variants and 6.97 dB on unseen scene categories. The paper further evaluates reference replacement and sampling, random seeds, material shifts, receiver calibration, uncertainty, downstream beam management and localization, efficiency, failure cases, and one measured scene. These analyses make the empirical results more convincing and reproducible.

Critical weaknesses

The claim that CoRF learns physical propagation components is not directly validated. The paper treats the learned anchors, arrival directions, and powers as physical propagation components, but evaluates them only after rendering the final spectra. The simulated data provide ground-truth ray paths, and Appendix E.8.4 already ray-traces test queries for oracle rendering, yet the predicted components are not compared with these paths. The AoA metric evaluates only the strongest peak, while Appendix E.8.3 reports about 244 effective anchors per query, making it unclear whether the secondary components correspond to real reflected paths or form a diffuse latent basis. Reference replacement shows that the model uses scene-specific spectra, and the learned-renderer ablation supports the analytic observation model, but neither establishes the physical meaning of the learned components. This is important because learning propagation, rather than only predicting spectra through a physics-based decoder, is a central claim of the paper.

The evidence for generalization beyond the simulated distribution is still limited. All 35 simulated scenes use the same ray tracer, material table, carrier, and specular reflection/refraction model. The only measured evaluation uses one scene and one receiver, and the best result requires 96 target-domain spectra together with flexible receiver calibration and component-power refinement. Uncalibrated feed-forward CoRF obtains 14.79 dB PSNR, below WRF-GS+ at 16.96 dB, while CoRF with calibration reaches 16.74 dB and the full adapted model reaches 17.45 dB. Since the real-data calibration can also absorb some propagation mismatch, this experiment does not cleanly isolate the transfer of the learned propagation model. Therefore, the strong feed-forward gains are established within the shared simulation pipeline, but cross-regime and simulation-to-real generalization remain less conclusive.

Questions for authors
  1. Can the authors directly validate the learned physical propagation components using the Sionna ray paths already generated for the oracle-renderer experiment in Appendix E.8.4? Please match the predicted and ground-truth components and report angular and relative-power errors for both dominant and secondary paths. Could you also visualize the active anchors, arrival directions, and powers together with the ray-traced paths and scene reflectors for representative and failure cases? This evidence is important for distinguishing physical propagation learning from an effective latent decomposition that only reproduces the output spectrum.
  2. Could you provide additional simulation-to-real evidence across different measured scenes or receivers, if available? For S36, please add the power-refinement-only result and repeat the evaluation over different reference and calibration splits. Could you also report the fitted calibration ranges or an identity-control experiment on simulated data to clarify whether the flexible receiver model is absorbing propagation errors?
  3. Since GRaF is the only baseline using the same feed-forward reference-conditioned setting, what neighbor count LL is used in Tables 2 and 3? Appendix E.2.2 shows that GRaF decreases from 22.00 to 13.85 dB when moving to 32 references, while CoRFPR decreases from 24.33 to 22.99 dB. Could you compare the methods across matched reference budgets and spatially stratified reference sets? This would help distinguish gains attributable to the proposed propagation model from gains due to robustness under sparse local coverage.

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

Related papers

Discussion (0)

Sign in to comment.