acceptodds
Under review as a conference paper at ICLR 2027

Physics-grounded Neural Radar Fields for High-fidelity Novel View Synthesis

Abstract

Radar-based novel view synthesis (NVS) supports robust scene understanding and radar simulation for autonomous driving, particularly under poor illumination and adverse weather. However, most existing methods simplify radar measurement formation and primarily model scene responses, leaving radar-specific artifacts insufficiently represented. This limits the fidelity of synthesized data to real radar measurements. The resulting synthetic-to-real gap can hinder the use of synthesized data in downstream tasks. To narrow this gap, we propose PhyRF, a physics-grounded neural radar field framework for high-fidelity radar-based NVS. For points sampled along each radar beam, the neural radar field predicts scene occupancy, radar reflectivity, and artifact contribution. We propose a differentiable renderer grounded in radar measurement formation that combines these predictions according to their distinct physical roles to synthesize measurements. To help disentangle scene geometry from artifacts during training, we further introduce LiDAR measurements as additional supervision by optimizing a shared geometry field. Extensive experiments demonstrate that PhyRF synthesizes radar measurements with higher fidelity to real observations than existing methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Novel View Synthesis, Neural Radiance Fields (NeRF)

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_L83cOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PhyRF, a neural radar field framework for novel view synthesis of spinning FMCW radar. For points sampled along each radar beam, the field predicts occupancy, reflectivity, and a radar-specific artifact coefficient. A differentiable spectral renderer combines these predictions following FMCW measurement formation, summing the windowed DFT responses of all sampled points to form the range profile, instead of rendering each measurement point-wise. A LiDAR branch shares the geometry field to help disentangle scene structure from radar artifacts. Experiments on Boreas scenes under sunny, rainy, night, and snowy conditions show improvements over RadarFields, RF4D, and RadarSplat in rendering fidelity and occupancy estimation.

Critical strengths
  1. The motivation is clear and well illustrated: modeling radar-specific artifacts and the spectral formation of radar measurements, rather than producing overly clean point-wise renderings, addresses a real limitation of existing methods.
  2. The spectral renderer is grounded in FMCW measurement formation, and the ablation shows that it is the main source of improvement.
  3. The method consistently outperforms strong baselines across four weather and lighting conditions, and the qualitative results clearly show the roles of the artifact field and LiDAR guidance.
  4. The paper is well written and easy to follow.
Critical weaknesses
  1. Unfair evaluation and unvalidated LiDAR design. PhyRF uses LiDAR during training, while occupancy is evaluated against LiDAR and all baselines use radar only, which favors the proposed method. The claim that joint optimization is better than direct LiDAR supervision is not validated, and snowy scenes are handled by manually reducing the LiDAR weight to 0.2 rather than automatically. Moreover, the occupancy evaluation excludes snowy scenes, where this design matters most.
  2. Main motivation is not validated. The paper argues that preserving radar-specific artifacts reduces the synthetic-to-real gap for downstream tasks, but no downstream experiment (e.g., detection or odometry) is provided. In addition, PSNR and SSIM may reward reproducing random noise, so they may not fully reflect fidelity.
Questions for authors
  1. Are the LiDAR frames used for occupancy evaluation excluded from training? Can the authors report occupancy results without LiDAR supervision for a fair comparison with radar-only baselines?
  2. How does joint optimization compare with direct LiDAR occupancy supervision, especially in snowy scenes, and why is a manually reduced LiDAR weight still needed?
  3. Can the authors provide evidence that the synthesized measurements reduce the synthetic-to-real gap in a downstream task?
4 comments

We sincerely thank you for the thorough evaluation and constructive feedback. Below, we address each concern with technical clarifications, experimental analysis, and necessary acknowledgments.

(W1&Q2) Comparison with direct LiDAR supervision

To validate our joint radar-LiDAR optimization strategy, we compare it with direct LiDAR occupancy supervision, where LiDAR measurements are used to generate occupancy map as directly supervison to the occupancy prediction. As shown in Table 1, while direct supervision also outperforms the radar-only variant (w/o LGL), our joint optimization achieves better performance across all scenes, demonstrating the benefit of learning shared geometry through joint radar-LiDAR optimization. Figure 11 in Appendix further shows that direct LiDAR supervision produces reasonable occupancy predictions in sunny conditions but introduces substantial noise and erroneous occupancy in snowy scenes due to unreliable LiDAR measurements. In contrast, our method produces cleaner and more accurate occupancy predictions under both conditions.

Table 1. Comparison between direct LiDAR occupancy supervision and our joint optimization strategy.

MethodSunnySnowRainNight
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
w/o LGL26.410.500325.800.504726.820.487127.400.5192
Direct26.580.504026.020.509226.980.494727.430.5241
Ours26.650.510526.770.522327.070.502827.510.5273

We have further included this analysis in Appendix A.8.

(W1) Reason for excluding snowy scenes

Following RadarSplat's evaluation protocol, we exclude snowy scenes from quantitative occupancy evaluation because LiDAR measurements are heavily affected by snow, making the constructed reference point clouds unreliable for assessing scene geometry.

To provide additional evidence under snowy conditions, we include qualitative comparisons with LiDAR-only reconstruction in Appendix Figure 14. The results show that our joint radar-LiDAR optimization produces cleaner geometry, suppressing snow-induced noise while preserving the main scene structures, whereas LiDAR-only reconstruction retains substantial noise.

(W1&Q2) Manual reduction of LiDAR weight in snow scenes

We clarify that manually reducing the LiDAR reconstruction weight is not required for our method to benefit from LGL under snowy conditions.

As shown in Table 2, using the same weight as in other weather conditions (λlidar=1\lambda_{\mathrm{lidar}}=1) already improves PSNR from 25.80 to 26.65 compared with the radar-only variant (w/o LGL). Reducing the weight to 0.2 further improves PSNR to 26.77, mitigating the influence of unreliable LiDAR measurements in snowy conditions.

Therefore, the reduced weight provides an additional performance improvement rather than being necessary for LGL to work under snowy conditions.

Table 2. Ablation study of LiDAR-guided geometry learning (LGL) on the snow scenes.

MethodPSNR ↑SSIM ↑
w/o LGL25.800.5047
λ_lidar = 126.650.5132
λ_lidar = 0.226.770.5223

We have further included this analysis in Appendix A.10.

(Q1) Are the LiDAR frames used for occupancy evaluation excluded from training?

Yes, the LiDAR frames used for occupancy evaluation are excluded during training. We follow the baseline method (RadarSplat) to construct LiDAR point cloud reference for occupancy evaluation and provide more details in Appendix A.12.

(Q1) Occupancy results without LiDAR supervision

Without LGL, our radar-only variant suffers from a degenerate solution where the predicted occupancy collapses to zero, as the artifact branch absorbs most radar information, as illustrated in Figs. 6 and 12. Consequently, meaningful occupancy metrics cannot be obtained for this configuration.

We acknowledge that this radar-only variant cannot reliably recover scene geometry. This observation further motivates LGL, which provides additional geometric constraints to prevent such degenerate solutions and improve the disentanglement of occupancy, reflectivity, and artifacts.

(W2&Q3) Downstream task

To further evaluate the fidelity of our synthesized radar measurements, we use RLPR[1] and RaLF[2], which focus on place recognition. We select these works because both provide publicly available pretrained models evaluated on the Boreas dataset, enabling direct evaluation without retraining or finetuning. We use test frames from all 13 scenes as radar measurement queries and boreas-2020-12-18-13-44 as the LiDAR map reference. Note that boreas-2020-12-18-13-44 covers the same locations as the 13 test scenes but was collected on a different date. We follow each method's official preprocessing and evaluation protocols and report recall@1, defined as the percentage of queries whose top-ranked LiDAR database match falls within the distance threshold (3m) of the GT query location. To meet the data requirements of each method, we extend the maximum rendering range from 50 m to 90 m and retrain all baselines and our method using this new range. Table 3 shows the NVS results and downstream place recognition performance on RLPR and RaLF. Our method achieves the best performance across all metrics, followed by our radar-only variant (w/o LGL). In particular, for downstream task evaluation, our method achieves performance much closer to that obtained using GT radar measurements.

Table 3. Radar synthesis quality and downstream place recognition performance on test frames from 13 Boreas scenes.

MethodPSNRSSIMRLPR Recall@1RaLF Recall@1
GT measurements----76.9258.65
RadarFields23.100.311360.5834.62
RF4D24.300.197460.5827.88
RadarSplat26.150.493669.2339.42
Ours (w/c LGL)27.050.505173.0849.04
Ours (full)27.200.515274.0454.81

We have further added this analysis as a subsection in Sec. 5.5.

Reference:

[1] Qi, Z., et al. "RLPR: Radar-to-LiDAR Place Recognition via Two-Stage Asymmetric Cross-Modal Alignment for Autonomous Driving," in IEEE Robotics and Automation Letters, vol. 11, no. 8, pp. 9676–9683, 2026.

[2] A. Nayak, D. Cattaneo, A. Valada, "RaLF: Flow-based global and metric radar localization in LiDAR maps," in 2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 5097–5103.

The author clearly addressed my concerns, so I have raised my score to Clear acceptance.

Thank you so much for recognizing our work and increasing your rating! We sincerely appreciate your insightful suggestions for strengthening our work. Thank you again for your time and effort in reviewing our paper.

Reviewer_MLg4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PhyRF, a physics-grounded neural radar field for radar-based NVS. For each 3D coordinate, a neural field predicts its occupancy, radar reflectivity, and artifact contribution, which define a harmonic function. Then a radar’s ray sums over its associated responses to determine its final measurement. They further introduce LiDAR as an additional supervision signal. Experiments show the contribution of each part of PhyRF and its advantages over previous methods.

Critical strengths
  • Modeling the physical response with a differentiable spectral renderer is an interesting and effective approach. It is consequential to my overall assessment because it makes the results more likely to generalize.

  • Artifact handling seems very effective. It is consequential to my overall assessment because artifacts are a main obstacle in radar rendering.

Critical weaknesses
  • If I understand correctly, the paper only shows comparisons to methods that used radar as inputs with no LiDAR use; is this because there are no such methods? If so, it would be better to show the “no LiDAR ablation” in Table 1 as well and to have apples-to-apples comparisons. In addition, the effect of the input LiDAR is not clear. As Figure 2 (a) shows, LiDAR is less reliable in snowy conditions. Yet in Table 2, without LiDAR guidance, “snow” gets 25.80 PSNR compared to 26.77 with LiDAR guidance, an almost 1dB improvement. In contrast, “sunny” gets only a 0.24 dB improvement. It is consequential to my overall assessment because it raises doubts about how the LiDAR guidance works.

  • In Table3, the paper shows a geometric evaluation compared to methods with only radar supervision. If so, it should also include methods that use only LiDAR as inputs, like neural lidar fields, to see if there is any advantage in geometric accuracy compared to methods that use only LiDAR. It is consequential to my overall assessment because the added value of radar for geometry stays unclear.

Questions for authors

Can the authors report geometry metrics (Table 3) in the ablation study (Table 2)?

7 comments

We sincerely thank you for the insightful feedback and comments. Below, we address each concern with technical clarifications, experimental analysis, and necessary acknowledgments.

(W1) Comparison to methods that used radar as inputs with no LiDAR.

Yes, to our knowledge, there is no method for radar-based NVS that use both radar and LiDAR.

There is only one work, NeuRadar, that render radar, LiDAR and RGB data. But we do not include it as a baseline because they use radar point clouds, while we use range-azimuth maps. Range-azimuth maps preserve more geometric information than spares radar point clouds extracted from them via filtering, so the inputs are not directly comparable.

(W1) Effect of LiDAR-guided geometry learning (LGL).

We clarify that the larger improvement from LGL is not specific to snowy conditions. The magnitude of improvement varies across scenes, and the larger gain observed in the selected snowy scenes reflects scene-specific differences rather than a general trend associated with snowy weather.

Our motivation for introducing LiDAR is to constrain geometry learning and prevent degenerate solutions where the artifact branch absorbs most radar information, resulting in poor disentanglement of occupancy, reflectivity, and artifacts.

As shown in Appendix Figure 12, we provide an example of the same environment under sunny and snowy conditions, where the radar-only variant (w/o LGL) achieves comparable or even higher PSNR and SSIM (+0.01 PSNR, +0.003 SSIM)than the full model on individual test frames. However, predicted occupancy and reflectivity using the variant (w/o LGL) fail to preserve meaningful scene properties, as the artifact branch absorbs most radar information. In contrast, ours with LGL helps preserve meaningful occupancy and reflectivity predictions under both weather conditions.

We have further included this analysis in Appendix A.11.

(W2) LiDAR-only methods for geometry evaluation.

Table 1 compares our method with a LiDAR-only method (LiDAR-NeRF) on non-snowy scenes.

For a fair comparison, we evaluate our rendered LiDAR point clouds against those of LiDAR-NeRF, as both follow the same LiDAR sampling pattern and resolution. Our predicted radar point clouds are included only for reference due to differences in sensor sampling and resolution. As shown in Table 1, our rendered LiDAR point clouds outperform LiDAR-NeRF across all six metrics, demonstrating that incorporating radar measurements into joint optimization improves geometry estimation beyond LiDAR-only supervision.

Furthermore, since LiDAR measurements are less reliable in snowy conditions, we provide qualitative comparisons in Appendix Figure 14. Compared with LiDAR-only reconstruction, our method produces cleaner geometry by suppressing snow-induced noise while preserving the main scene structures.

Table 1. Occupancy estimation averaged over non-snowy scenes.

MethodRMSE (m) ↓RCD ↓ACC ↑Pre ↑Rec ↑F1 ↑
LiDAR-only0.78510.021985.3186.8084.8385.66
Ours (radar)0.84300.037682.6071.5785.4777.28
Ours (LiDAR)0.77630.020285.8787.7585.2486.34

We have further included this analysis in Appendix A.12.

(Q1) Ablation study on occupancy estimation

We provide an ablation study on occupancy estimation using non-snowy scenes in Table 2.

We evaluate the effects of the spectral renderer (SR), LiDAR-guided geometry learning (LGL), artifact branch (AF), and rendered-LiDAR filtering (RLF). While SR, LGL, and AF are components of our framework, RLF is applied only during evaluation.

Since our predicted occupancy represents the geometry of the entire scene, it may include regions not visible from the current viewpoint, resulting in substantially broader spatial coverage than the single-scan LiDAR reference point cloud, as illustrated in Appendix Figure 13(a). Therefore, these two point clouds are not directly comparable as shown in Table 2 Row 1-5. To address this mismatch, we apply RLF to retain only predicted radar points whose nearest rendered LiDAR point is within 1 m, aligning the evaluation coverage with the single-scan reference.

With RLF enabled (row 6-8), the three configurations achieve comparable occupancy estimation performance. However, each configuration uses its own rendered LiDAR point cloud for filtering, so the observed differences reflect both occupancy prediction and filtering quality and cannot be attributed solely to the ablated components.

Table 2. Occupancy estimation ablations averaged over non-snowy scenes, grouped by whether RLF is disabled (top) or enabled (bottom). (Without LGL, the predicted occupancy collapses to zero, as the artifact branch absorbs the occupancy information, making these metrics inapplicable.)

SRLGLAFRLFRMSE (m) ↓RCD ↓ACC ↑Pre ↑Rec ↑F1 ↑
✓✓8.98480.821749.8824.2294.0937.40
✓3.02580.218561.4954.5367.0557.45
✓✓//////
✓✓5.51180.304366.7737.0091.9652.09
✓✓✓4.98710.252666.1635.8789.4150.59
✓✓✓0.82290.045383.5371.7287.4378.24
✓✓✓0.79410.029984.4272.5287.8878.95
✓✓✓✓0.84300.037682.6071.5785.4777.28

We have further included this analysis in Appendix A.12

Thanks for the rebuttal, it partially addressed my concerns.

  • I'm not convinced by the explanation for the large LGL gain in snow. There are only two snow scenes, so calling it scene-specific is hard to verify without per-scene results.

  • For geometry, the gain over LiDAR-only in your Table 1 here is very small. In your Table 2 here, the radar-only version gives no geometry at all (row 3, Table 2 here). Without the rendered-LiDAR filter (row 5 Table 2, here), the full model is worse than the baselines in Table 3 of the paper , and the baselines don't get this filter. So I'm still not sure radar adds much to the geometry.

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

This paper proposes PhyRF, a physics-grounded neural radar field framework for radar novel view synthesis. The authors argue that existing methods primarily model radar responses but may not explicitly account for radar-specific artifacts, which can limit the realism of synthesized data and its usefulness for downstream tasks. PhyRF includes: (1) a neural radar field that predicts scene occupancy, radar reflectivity, and radar-specific artifacts; (2) a differentiable spectral/IF-domain renderer that simulates the radar measurement process; and (3) LiDAR-supervised shared geometry to provide richer geometric constraints. The paper reports results on 13 scenes under diverse weather conditions, where PhyRF is state-of-the-art.

Critical strengths
  1. The motivation is clear and relevant. Modeling radar-specific artifacts, rather than only radar responses, is a reasonable direction for improving synthesis realism.
  2. The design is sensible. Jointly predicting scene properties and radar artifacts, and separating reflectivity from artifacts, provides a useful structure for interpreting the model's predictions.
  3. The experiments are reasonably thorough. The method is evaluated on 13 scenes under different weather conditions, and the visual results suggest that artifacts are at least partially modeled.
Critical weaknesses
  1. The novelty relative to prior frequency-space methods needs clearer explanation. Prior work already uses frequency-domain analysis for radar synthesis(RadarSplat). The paper should more clearly state what is fundamentally new beyond existing methods.
  2. Dynamic scenes are under-evaluated. The experiments mainly cover static or quasi-static settings, so it is unclear how the method behaves with moving objects or time-varying effects. Prior work such as RF4D has already considered dynamic radar scenes.
  3. The downstream value is motivated but not demonstrated. The paper suggests that more realistic radar synthesis could benefit downstream perception, but there is no evaluation on detection, tracking, or other perception tasks. The corresponding claims should be qualified
  4. The LGL-disabled comparison needs clearer interpretation. Since LiDAR supervision adds extra geometric information, the LGL-disabled variant provides a useful controlled comparison. In this setting, PhyRF achieves higher PSNR in all four conditions, while its SSIM is equal in sunny scenes, lower in snow and rain, and higher at night, which should be discussed. Separately, the occupancy evaluation applies additional rendered-LiDAR filtering to PhyRF. Reporting the results before and after filtering would clarify how much of the geometric improvement comes from the learned occupancy field itself.
  5. The physical interpretation of the artifact field is not fully validated. A better reconstruction and visually plausible decomposition do not establish that this branch captures physical artifacts rather than errors in geometry or reflectivity. A controlled example with known geometry and artifact locations, or a more limited interpretation of the branch, would help address this concern.

1, 4, and 5 are more important.

Questions for authors
  1. Clarify whether PhyRF can handle dynamic scenes. The experiments focus mainly on static settings, so the paper should state whether PhyRF targets static-only synthesis or can extend to dynamic scenarios such as moving objects and time-varying multipath.
  2. Is the downstream benefit intended as a demonstrated result or a potential application? The current evidence (Table 3) supports geometric consistency but not the higher-level downstream benefit the introduction motivates.
  3. Does disabling LGL remove all LiDAR-derived supervision and processing? Please discuss the consistent PSNR advantage and the mixed SSIM results relative to RadarSplat. Separately, how do the occupancy metrics change when the rendered-LiDAR filtering is removed?
  4. What evidence distinguishes the artifact field from a residual branch that also absorbs geometry or reflectivity errors? Is there a controlled example or independent reference that validates the decomposition?
3 comments

We sincerely thank you for the thorough evaluation and constructive feedback. Below, we address each concern with technical clarifications, experimental analysis, and necessary acknowledgments.

(W1) Novelty relative to existing frequency-space methods

Our main novelty lies in deriving the spectral renderer from the FMCW radar measurement formation process, rather than approximating frequency-domain effects on rendered radar measurements.

RadarSplat directly predicts power at each location and convolves the predicted power with a Gaussian kernel along the range dimension to approximate DFT-induced spectral leakage.

In contrast, our spectral renderer follows the FMCW radar measurement formation process, from transmitted and received signals to IF signal generation and DFT processing. Specifically, we calculate the received signal amplitude according to the radar equation and derive a closed-form DFT response for each reflector, explicitly modeling its contribution to the final radar measurements.

To validate this difference, we replace our spectral renderer with RadarSplat's renderer while keeping the remaining framework unchanged. As shown in Table 1 (below), our spectral renderer consistently achieves higher PSNR and SSIM across all four weather conditions, demonstrating the effectiveness of our physics-grounded formulation.

Table 1 Comparison between RadarSplat renderer and our spectral renderer across different weather conditions.

MethodSunnySnowRainNight
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
RadarSplat renderer25.890.458326.090.473126.450.447026.870.4825
Ours26.650.510526.770.522327.070.502827.510.5273

We have further included this analysis in Appendix A.9.

(W2&Q1) Dynamic scenes

We acknowledge our method is limited to static scenes, assuming that scene geometry remains unchanged over time. While our spectral renderer is grounded in the radar measurement formation process and remains applicable to dynamic scenes, moving objects introduce time-varying geometry that cannot be captures by our current neural radar fields. Future work will incorporate temporal modeling to capture dynamic geometry, extending our framework to dynamic driving environments.

We have further included a subsection of "Limitation and future work" in Appendix A.13.

(W3&Q2) Downstream value

To further evaluate the fidelity of our synthesized radar measurements, we use RLPR[1] and RaLF[2], which focus on place recognition. We select these works because both provide publicly available pretrained models evaluated on the Boreas dataset, enabling direct evaluation without retraining or finetuning. We use test frames from all 13 scenes as radar measurement queries and boreas-2020-12-18-13-44 as the LiDAR map reference. Note that boreas-2020-12-18-13-44 covers the same locations as the 13 test scenes but was collected on a different date. We follow each method's official preprocessing and evaluation protocols and report recall@1, defined as the percentage of queries whose top-ranked LiDAR database match falls within the distance threshold (3m) of the GT query location. To meet the data requirements of each method, we extend the maximum rendering range from 50 m to 90 m and retrain all baselines and our method using this new range. Table 2 shows the NVS results and downstream place recognition performance on RLPR and RaLF. Our method achieves the best performance across all metrics, followed by our radar-only variant (w/o LGL). In particular, for downstream task evaluation, our method achieves performance much closer to that obtained using GT radar measurements. Our radar-only variant (w/o LGL) also substantially outperforms other baseline methods.

Table 2 Radar synthesis quality and downstream place recognition performance on test frames from 13 Boreas scenes.

MethodPSNRSSIMRLPR Recall@1RaLF Recall@1
GT measurements----76.9258.65
RadarFields23.100.311360.5834.62
RF4D24.300.197460.5827.88
RadarSplat26.150.493669.2339.42
Ours (w/c LGL)27.050.505173.0849.04
Ours (full)27.200.515274.0454.81

We have further added this analysis as a subsection in Sec. 5.5.

Reference:

[1] Qi, Z., et al. "RLPR: Radar-to-LiDAR Place Recognition via Two-Stage Asymmetric Cross-Modal Alignment for Autonomous Driving," in IEEE Robotics and Automation Letters, vol. 11, no. 8, pp. 9676–9683, 2026.

[2] A. Nayak, D. Cattaneo, A. Valada, "RaLF: Flow-based global and metric radar localization in LiDAR maps," in 2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 5097–5103.

(W4&Q3) PSNR advantage and mixed SSIM results

Compared with RadarSplat, our LGL-disabled variant achieves mixed SSIM results while consistently achieving higher PSNR across all four conditions.

This difference can be partially attributed to the reconstruction objectives. Our method adopts an MSE reconstruction loss, whose minimization directly corresponds to maximizing PSNR, whereas RadarSplat employs L1 and SSIM losses, providing more direct supervision of local structural similarity. These loss functions follow the common practices of NeRF-based and GS-based frameworks, respectively, rather than being specifically chosen to favor different evaluation metrics.

In addition, without LGL, the model may lead to a trivial solution where the predicted occupancy approaches 0 and the artifact explains most radar measurements (as discussed in the original paper Row 296-300) . Our spectral renderer is derived from the radar measurement formation process, which assumes accurate scene geometry (Appendix A.1). Therefore, this trivial solution may reduce the accuracy of this rendering process, potentially leading to a loss of local structure details and lower SSIM. This further motivates LGL, which provides additional geometric constraints to prevent such degenerate solutions.

(Q3) Does disabling LGL remove all LiDAR-derived supervision and processing?

Yes. Our radar-only variant (w/o LGL) removes all LiDAR-related components, including LiDAR inputs, the LiDAR prediction head in the neural field, and the LiDAR reconstruction loss. Therefore, this variant is trained using radar measurements only.

(W4&Q3) Effect of rendered-LiDAR filtering for occupancy estimation

We acknowledge that without the rendered-LiDAR filter (RLF), our method achieves worse geometric metrics than the radar-only baselines.

This can be attributed to that our method learns a shared occupancy field from radar and LiDAR measurements across multiple viewpoints. Consequently, the radar point clouds derived directly from predicted occupancy may contain geometry not visible from the current viewpoint, which can be penalized as false positives even if physically valid (as shown in Figure 13(a)). We therefore use the rendered LiDAR point clouds in novel viewpoints to retain only radar-derived points whose nearest rendered LiDAR point is within 1 m, obtaining the geometry associated with the current viewpoint.

We have revised Section 5.4 and Table 3 to report both filtered and unfiltered results, clarify the reason for worse occupancy metrics without RLF, and explain how RLF jointly uses rendered LiDAR point clouds and radar point clouds to estimate occupancy in novel viewpoints.

(W5&Q4) Effect of artifact branch

Our artifact branch is designed to model artifact-related signal contributions rather than simply absorb rendering errors. Specifically, both physical reflectors and artifact-related components contribute to the intermediate-frequency (IF) signal, which subsequently undergoes DFT processing to form the final radar measurements. Therefore, we explicitly predict artifact contributions and incorporate them into spectral rendering, rather than adding a residual term to the rendered measurements.

To validate this design, we replace our artifact branch with a residual branch, where radar measurements are rendered using only occupancy and reflectivity, followed by a predicted residual term to compensate for the difference between rendered and GT measurements.

The quantitative and qualitative results are presented in Table 3 (below) and Appendix Figure 10, respectively. As shown in Table 3, our artifact branch consistently outperforms residual modeling across all four weather conditions.

Furthermore, Appendix Figure 10 reveals that residual modeling leads to a trivial solution: although occupancy can still be learned through LiDAR supervision, the predicted reflectivity collapses to zero, while the residual branch absorbs nearly all radar information. Consequently, the residual model fails to meaningfully separate physical reflectivity from artifact-related contributions.

Figure 1 provides qualitative evidence of the decomposition. Our predicted 3D occupancy is similar to the 3D LiDAR reference (shown in the left corner), while different objects exhibit distinct predicted reflectivities. Meanwhile, the artifact branch primarily captures radar-specific artifacts that do not correspond to physical structures in the scene.

Table 3 Comparison between residual modeling and our artifact branch.

MethodSunnySnowRainNight
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Residual26.220.480425.870.461026.700.467327.140.5068
Ours26.650.510526.770.522327.070.502827.510.5273

We have added this analysis in Appendix A.7

Thank you for constructive feedback.

We have provided additional experiments and clarifications to address your concerns. As the discussion period is approaching its end, we would be grateful for any further feedback on whether your concerns have been addressed.

Thank you again for your time and effort in reviewing our paper.

Comments on the paper

1 comment

We thank the reviewers for their efforts and insightful comments, which have helped strengthen our work. We have uploaded a revised manuscript with all changes highlighted in blue.

Our work focuses on high-fidelity radar novel view synthesis, with two main technical contributions:

  • a neural radar field framework that jointly models scene responses and radar-specific artifacts: The neural radar fields predict occupancy, reflectivity and artifacts. To better disentangle scene geometry from artifacts, we jointly optimize the geometry field using radar and LiDAR measurements.
  • Physics-grounded spectral renderer: We derive a differentiable spectral renderer from the radar measurement formation process that directly constructs the frequency spectrum used to form radar measurements.

We appreciate the reviewers' recognition of our motivation and technical design.

Motivation and technical design

pnmg: "The motivation is clear and relevant"; "The design is sensible"

MLg4: "Modeling the physical response with a differentiable spectral renderer is an interesting and effective approach"

L83c: "The motivation is clear and well illustrated"; "The spectral renderer is grounded in FMCW measurement formation."

Performance

pnmg:"The experiments are reasonably thorough. The method is evaluated on 13 scenes under different weather conditions, and the visual results suggest that artifacts are at least partially modeled."

MLg4:"Artifact handling seems very effective"

L83c:"The method consistently outperforms strong baselines across four weather and lighting conditions, and the qualitative results clearly show the roles of the artifact field and LiDAR guidance."

The reviewers also raised valuable concerns and questions. We have addressed them in individual responses, supported by additional experiments and analysis.

Our key responses are summarized below:

Downstream value (pnmg W3 and L83c W2)

  • We evaluated the fidelity of synthesized radar measurements by using them as inputs to pretrained downstream place recognition models. Our method achieves performance closest to that obtained with GT radar measurements among the synthesis methods, demonstrating the practical value of high-fidelity radar NVS beyond reconstruction metrics. To the best of our knowledge, this is the first evaluation of radar NVS methods on downstream tasks.

LiDAR-guided geometry learning(LGL)

  • We conducted additional experiments on two new snowy scenes and clarified that the magnitude of PSNR improvement varies across scenes, rather than being consistently larger in snowy conditions. (MLg4, W1)
  • We explained the mixed SSIM results without LGL compared with RadarSplat, considering differences in reconstruction objectives and the dependence of spectral rendering on accurate geometry. Without LGL, the artifact branch may absorb most radar responses, reducing structural fidelity (pnmg, W4).
  • We demonstrated that joint radar–LiDAR geometry optimization is more effective than direct LiDAR supervision, and clarified that reducing the LiDAR reconstruction weight in snowy scenes is not necessary for LGL to provide improvements (L83c, W1).
  • We provided qualitative geometry reconstruction results for snowy scenes and quantitative evaluations for non-snowy scenes to further assess the effectiveness of joint radar–LiDAR geometry learning. (MLg4, W2)

Spectral renderer and artifact modeling

  • We conducted additional experiments demonstrating that our spectral renderer is more effective than the renderer used in prior work, which uses Gaussian convolution to approximate the DFT. (pnmg, W1).
  • We conducted additional experiments demonstrating that our artifact branch achieves better synthesis performance than a residual branch designed to compensate for rendering errors (pnmg, W5).

Occupancy estimation for novel viewpoints

  • We explained that snowy scenes are excluded from the occupancy evaluation following the baseline method's protocol, due to unreliable LiDAR measurements under snowy conditions. (L83c, W1)
  • We now report occupancy estimation results both with and without the rendered-LiDAR filter (RLF).(pnmg W4)
  • We clarified why rendered LiDAR point clouds are used to filter occupancy-thresholded point clouds (occupancy > 0.5) to estimate geometry associated with novel viewpoints. Since our occupancy field is learned from multiple viewpoints, the extracted point clouds contain more geometry not visible from the current viewpoint. These additional points are counted as false positives against the current-view LiDAR reference, leading to worse geometric metrics. (pnmg W4,MLg4)

Limitations and future work (pnmg W2)

  • Our method currently assumes static geometry. Since the spectral renderer models radar measurement formation, it can support dynamic scenes with time-varying geometry. Future work will incorporate temporal modeling into the neural radar field to capture moving objects.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.