acceptodds
Under review as a conference paper at ICLR 2027

Radon Implicit Field Transform: Fully Radar-Native Novel-View Synthesis and 3D Reconstruction

Abstract

We introduce the Radon Implicit Field Transform (RIFT), a fully radar-native framework for complex-valued novel-view synthesis and 3D reconstruction that combines an adaptive point-scatterer representation with a radar forward model. Both initialization and optimization use only complex radar measurements and calibrated sensor poses, without visual or LiDAR-derived priors, initialization, or supervision. RIFT represents the scene with point scatterers at adaptive spatial resolutions, with learnable positions and direction-dependent complex reflectance parameterized by spherical harmonics. We incorporate the generalized Radon transform as a forward model and optimize point-scatterer positions and reflectance by matching predicted and measured complex radar signals. During optimization, signal-loss gradients guide local increases in spatial resolution and spherical-harmonic expansion degree; these capacity updates leave the current signal predictions unchanged, except for the coarse-to-fine densification we use on lower-resolution measured radar data. We evaluate complex-valued novel-view synthesis by comparing forward-model predictions at unseen sensor poses with held-out complex measurements, and evaluate 3D reconstruction by comparing geometry extracted from the learned point scatterers with ground-truth scene geometry. On our simulated dataset, RIFT achieves a mean held-out complex relative mean-squared error of and a mean symmetric squared Chamfer distance of ;on a real-world dataset, it reduces the symmetric squared Chamfer distance by up to relative to the strongest baseline, reaching .

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Reconstruction, Scene Representations, Inverse Problem, Radar Novel-View Synthesis (NVS)

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_XE1vOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents RIFT, a framework for radar novel-view synthesis and 3D reconstruction using only radar measurements and calibrated sensor poses. It represents a scene as learnable scattering points with direction-dependent complex responses. A physics-based radar model predicts measurements, and the scene representation is optimized by matching these predictions to the observed signals. The method gradually increases spatial and directional detail during training. The same learned representation supports both signal prediction at new viewpoints and geometry reconstruction. Experiments use simulated data and real measurements from the GOTCHA dataset.

Critical strengths
  1. The adaptive representation is a useful design. The method uses training gradients to decide where to add spatial detail and more complex directional responses. This provides a concrete way to adjust the representation to the observed measurements.
  2. The experimental evaluation is extensive. The paper includes experiments on both simulated and real radar data, comparisons with multiple baselines, and a broad set of metrics covering signal prediction and 3D reconstruction. The detailed results and additional visualizations in the appendix provide a thorough assessment of the method.
Critical weaknesses
  1. The novelty needs clearer positioning against related work. The paper should cite and discuss RFDT [1], mmIR [2], which address closely related problems involving differentiable RF rendering, complex radar synthesis, and scene recovery. The general idea of recovering scene parameters by matching simulated and measured RF signals is already established. RIFT’s contribution therefore appears to lie primarily in its adaptive spatial and angular representation and its specific optimization strategy. However, the individual benefits of these designs are not sufficiently demonstrated through controlled ablations.
  2. The main paper contains too few visual results. Many useful visualizations are placed in the appendix, making it difficult to understand the method’s benefits and assess the results from the main text alone. The authors should move more representative reconstruction and novel-view synthesis comparisons into the main paper, including comparisons with existing methods. This would make the paper easier to follow and its improvements more convincing.
  3. The fairness of the qualitative comparisons needs clarification. Several baselines produce very poor visual results, raising concerns about whether they were reproduced correctly. Radar methods can be sensitive to implementation details and data processing, so these results may reflect reproduction issues rather than limitations of the original methods. I recommend restricting the main comparison to methods with official implementations that support the same dataset, using a consistent evaluation protocol. Any adapted or reimplemented baselines should be clearly identified and validated before their visual results are used to support claims of superiority.

[1] Xingyu Chen, Xinyu Zhang, Kai Zheng, Xinmin Fang, Tzu-Mao Li, Chris Xiaoxuan Lu, and Zhengxiong Li. “Physically Accurate Differentiable Inverse Rendering for Radio Frequency Digital Twin.” In MobiCom 2026. [2] Adnan Armouti, Yixuan Gao, and Rajalakshmi Nandakumar. “mmIR: Frequency-Space Inverse Rendering for 3D Millimeter-Wave Radar ADC Synthesis.” In ECCV 2026

Questions for authors

What checks were performed to validate the adapted baseline implementations? In particular, please clarify the validation of the PyTorch ports of Radar Fields and RadarSplat, and the decision to keep GeRaF’s model step at zero, which prevents its scheduled updates from progressing.

4 comments
AuthorsOpenReview ↗

Official Comments to Reviewer XE1v (Part 1 of 2, Round 1)

Dear Reviewer XE1v,

We appreciate your constructive feedback on our work. Here is a point-by-point response to the weaknesses and the question you brought up, posted in parts.

Weakness 1: We thank the reviewer for the two references. We will cite and discuss RFDT (Chen et al., MobiCom 2026) and mmIR (Armouti et al., ECCV 2026) in the related work of the revised paper, and position RIFT relative to them. For the individual benefits of the adaptive spatial and angular representation and of the optimization, please see the ablation study in the new Appendix H (Tables 10-15, summarized in Section 5.2), which changes one choice of the main recipe at a time on A320, B787 and the race car. In short, the signal depends mainly on the angular unlocks (3.3-3.9 times worse without them), the group-sparse prior (38-47 times) and the backprojection start (6-9 times on A320 and the race car).

Weakness 2: We agree that the visual comparison should be visible in the main text. The main text is at its page limit, so the 3D reconstruction and view comparisons with the baselines are in the appendix (Figs. 8-10 for the RIFT dataset, Figs. 12a-14c for GOTCHA). In the revised paper we will add a compact reconstruction comparison with the baselines to the main text within the page limit.

Weakness 3: We agree that adapted or re-implemented baselines must be identified, and that their visual results can reflect the adaptation and not the original method. Backprojection is standard; Radar Fields and RadarSplat use their released code with PyTorch ports of their CUDA-only operations (validation details below); for SpINRv1, GeRaF 1.0 and Sugavanam-Ertin we did not locate official implementations, and we implemented SpINR-style from the paper, GeRaF 1.0 from the released stage-1 components with the settings of the v1 paper, and Sugavanam-Ertin from the equations. Section 4 and Appendix F list what we retain and change for each method, and Section 5 states that the comparison evaluates the methods as adapted to our setting. No baseline has an official implementation that supports our data, the full-sphere multi-frequency responses of an isolated object, so none runs unchanged and a restriction to such methods would leave only backprojection. RIFT and the four learned baselines follow one protocol (the same 2,400 training views, 360,000 updates, the minimum-validation checkpoint and three training seeds) and are scored on the same reserved-test views with the same metrics.

Our response continues in Part 2 of 2.

AuthorsOpenReview ↗

Official Comments to Reviewer XE1v (Part 2 of 2, Round 1)

Question (validation of the adapted baselines). Radar Fields: tiny-cuda-nn is CUDA-only, so we run an fp32 PyTorch re-implementation of its encodings and networks, with the same layout and parameter counts as the original. Checked against the real tiny-cuda-nn on an H100, its encodings agree with the fp32 encodings to within 3.0e-5 and the outputs of the full model to within 5.1e-4 (relative); its initialization and gradient reduction still differ from the original backend. RadarSplat: its rasterizer and fused-SSIM operations are CUDA-only, so we ported them to PyTorch with their backward passes. On saved inputs of a real B787 target, against the fp32 CUDA reference, every operation is identical or within 1.5e-7, the rasterized products are within 2.9e-5, the loss within 3.4e-7 (relative) and the parameter gradients within 7e-4 (relative L2). We did not compare training trajectories with the original CUDA implementations or reproduce the original papers' results on their own datasets.

GeRaF's model step. In the submission we kept the model step at zero, as the released runner does, so its annealing and sharpness schedules stayed at their initial values. In the updated experiments the step advances: GeRaF is trained with the schedules stretched to our 360,000 updates (annealing ends at 360,000, sharpness unfreezes at 72,000; the release uses 50,000 and 10,000) on all six objects and three training seeds, 18 runs, under one declared variant (gradient clip 3, a zero gradient for the matched-filter magnitude where the prediction is exactly zero, and a non-finite guard):

GeRaF, mean over the six objectsComplex RelMSE (%)Range-power RelMSE (%)
Step held at zero (released runner), seed 42148.018.91
Step advancing, seed 42152.923.34
Step advancing, mean of three seeds (GeRaF row of Table 1)167.418.24

Advancing the step does not change the picture: the complex RelMSE is above 100% in all 18 runs (123.9% to 215.6%), and these runs replace the earlier GeRaF row of Table 1.

What these checks mean. They validate the operations and losses of the ports, not the end-to-end behavior of the adapted methods, and none of this diminishes the contribution of the works we compare to. GeRaF, Radar Fields and RadarSplat were designed mainly for azimuthal views, with native observables organized in range and azimuth, and are trained on matched-filtered power or amplitude; our comparison evaluates them as adapted to our setting (Section 5), and their entries characterize our adaptation under this paper's budget and radar presentation, not the methods in their own settings (Appendix F). Our claim is limited to one problem: solving 3D reconstruction and radar novel-view synthesis together, from the same radar responses and within one representation. On that problem RIFT scores best in every column of Table 1 on the RIFT dataset and in every column but IoU of Table 2 on GOTCHA, against the baselines as adapted here.

We are grateful for your thoughtful suggestions, which helped us strengthen the manuscript during revision.

Best, Authors

The authors have addressed the main concerns about the method, baseline comparisons, and evaluation. I recommend accepting the paper.

AuthorsOpenReview ↗

Response to Reviewer XE1v

Dear Reviewer XE1v,

Thank you for taking the time to read our response. We’re glad we addressed your main concerns, and we appreciate your recommendation. Thanks to your review, we were able to explain the method more clearly and give a fuller account of the baseline comparisons and evaluation. We believe these revisions strengthened the paper.

Best, Authors

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

This paper proposes Radon Implicit Field Transform (RIFT), a radar-native method for complex-valued radar rendering and 3D reconstruction. The scene is a set of points with learnable positions and view-dependent reflectance (spherical harmonics), and adaptive point allocation. Because the scene and sensor are represented explicitly, the method supports novel view synthesis (NVS). The method is trained on two radar datasets, one public (GOTCHA) and one synthetic generated using Ansys software by the authors (RIFT-Dataset). NVS is compared against SpINR, RadarSplat, RadarFields and others using Complex RelMSE and common range-power RelMSE. Geometry is compared against SpINR and backprojection using Chamfer, max Hausdorff, HD95, IoU and F-score.

Critical strengths
  1. On RIFT-Dataset, Complex RelMSE improves from 46.12% to 0.3909% and Chamfer from 0.8179 to 0.4813 (×10⁻⁴ m²). On GOTCHA, validation Complex RelMSE improves from 128.5% to 54.33% and F-score from 0.4416 to 0.5513. Qualitative results in the main paper and appendix are convincing.
  2. The formulation is clean. A single explicit representation addresses both radar NVS and geometric reconstruction, which prior work usually treats separately.
Critical weaknesses
  1. The paper states that the scalar single-scattering renderer omits the waveform and antenna pattern responses, and visibility and occlusion are also handled implicitly. View-dependent SH reflectance can absorb beam pattern, occlusion and polarization effects, which can help fit the data but entangles with material response. This limits interpretability.
  2. The near-zero error on RIFT-Dataset raises the question of whether the Ansys settings (solver, bounce count, polarization) match the renderer’s own single-scattering, scalar assumptions. If so, the synthetic results partly measure self-consistency instead of robustness to model mismatch, and could explain the difference compared to the results obtained on real data in GOTCHA dataset.
  3. Training uses roughly 2400 views (Appendix E), about an order of magnitude more than typical scene reconstruction benchmarks. Radar returns are famously specular and sparse, and many commodity arrays resolve only one angular dimension, so geometry recovery from few views is severely ill-posed. This might explain why RIFT requires thousands of training views. In few view settings, radar-native methods are most likely to need LiDAR, so the view requirement should appear clearly in the main paper and the limitations.
  4. Evaluation is limited to two datasets only, where one is synthetic and produced by the authors. This makes it hard to judge generality across sensors and capture geometries.
  5. Ablations, runtime and memory are not reported. Please provide ablations on (a) the number of training views, with NVS and geometry metrics reported on both datasets; (b) the angular distribution of views, comparing full circular coverage against a limited arc or a linear trajectory at a matched view count; (c) Spherical Harmonics order; (d) adaptive point allocation versus a fixed point budget; and (e) loss terms.
Questions for authors
  1. Does implicit visibility cause failures, for example in occluded or concave regions? An ablation or failure case would help
  2. What is the minimum number of views for which geometry recovery remains reliable? A view-count sweep on both datasets would help
  3. Can you please explain the choice of implicit over explicit modeling for: (a) RCS via SH rather than a material decomposition, (b) visibility, and (c) antenna beam patterns?
  4. Can you please provide all settings used in Ansys AVxcelerate RSS Ray Tracer that generated RIFT-Dataset (bounce count, polarization, multipath)?
  5. Why is common range-power RelMSE omitted for GOTCHA? Please report it if possible.
  6. Can you please provide an explanation for the gap between synthetic and real results, and also for the train/validation gap on GOTCHA (41% vs 54%)?
  7. The renderer predicts calibrated frequency-domain samples. Could the authors clarify which effects are excluded when they state that it “omits explicit waveform and antenna pattern responses”?
13 comments
AuthorsOpenReview ↗

Official Comments to Reviewer omzY (Part 1 of 7, Round 1)

Dear Reviewer omzY,

Thank you for your careful review. We address your comments point by point across seven posts. We begin with two questions you raise: what the learned scattering responses reveal about self-shadowing, and how closely the simulator matches RIFT’s modeling assumptions.

Weakness 1: We agree that there are effects the reviewer points out are handled implicitly by design. The direction-dependent SH response of each point is an effective aspect-dependent response that absorbs the antenna pattern, polarization, visibility and occlusion, and material effects together, so they are fitted by one explicit model without ray casting; the cost is the interpretability the reviewer points out, and the coefficients should not be read as material parameters or as a visibility estimate (Limitations, "Modeling scope"). In our experiments occlusion between objects does not arise, because RIFT reconstructs one object at a time (a single object in an empty scene for the RIFT dataset, each vehicle in its own 6 m region of interest for GOTCHA); what remains is occlusion within an object (self-shadowing and concave parts). To see whether the learned field is visibility-aware there, we analyzed the trained fields of the main runs (seed 42) on B787, whose interior is thinner than one lattice cell (at most 4.5 mm), and on the fire truck, whose interior is deep (up to 18.9 mm, about three lattice cells) and whose body has pronounced concave regions, against the ground-truth meshes. A point's energy is its squared response magnitude averaged over 2,048 look directions. We report, averaged by energy over the points that can be judged, the share of a point's energy in directions hidden from it by the mesh divided by the share of hidden directions (1: visibility ignored; 0: no energy in hidden directions), and the fraction of the largest reduction that a degree-3 response allows that is reached (1: the best possible).

B787Fire truck
Energy in hidden directions relative to their share (share of the energy that can be judged)0.965 (85.5%)0.265 (48.3%)
Fraction of the largest possible reduction reached0.0350.74

The response is visibility-aware for the part of the fire truck's energy that can be judged (48%), and not on B787. The B787 points with the most hidden-direction energy are points at the engine nacelles (six of the seven largest), which put more energy into hidden directions than an isotropic response would (fraction of the possible reduction reached: -0.09 to -0.36); on the fire truck, the 20 points with the most hidden-direction energy (wheel and body points, 14 of them in concave regions) are suppressed only partly (0.18 to 0.60). Most of the response energy lies within the radar's localization limit of the outer surface (λ/4\lambda/4 = 7.5 mm at λ\lambda = 30 mm): 87% on B787 and 71% on the fire truck; on the fire truck, which has an interior deep enough to test this, only 0.8% lies deeper than λ/2\lambda/2 inside the hull. The representation can therefore learn self-shadowing, as on the fire truck, but it did not on B787.

Weakness 2: We agree that the synthetic setting is closely matched to RIFT's assumptions. All six objects of the RIFT dataset are generated with the same code and settings in the Ansys AVxcelerate RSS ray tracer (a maximum of one reflection; full settings in our answer to Q4). Two points qualify the concern. First, the data are not generated by RIFT's own renderer: the simulator is polarization-aware (we store one co-polarized, vertical-vertical complex channel) and models a 10-degree antenna beam, whereas RIFT is a scalar sum over point scatterers (therefore a wide beam). Second, the match does not by itself produce a low error: in Appendix H, storing the same complex SH field in a hash grid and an MLP is 3.4-20 times worse, and removing the group-sparse prior 38-47 times worse. The difference to GOTCHA is consistent with effects the renderer omits and the simulator lacks (clutter, car-ground multipath, calibration errors), but we have not isolated their contribution. To quantify the simulator-side mismatch, we regenerated the same viewpoints with up to 2, 3, 4 and 5 reflections: the complex response changes by 3.5% of the one-reflection energy (pooled over all viewpoints) and the total power by about -0.29 dB, so the one-reflection data are close to the multi-reflection data.

Our response continues in Part 2 of 7.

AuthorsOpenReview ↗

Official Comments to Reviewer omzY (Part 2 of 7, Round 1)

Weakness 3: We agree that the number of training views should be stated: 2,400 for each RIFT-dataset object and 2,410 azimuth units for each GOTCHA vehicle (Appendix E). We would add that this number is a property of radar-native 3D reconstruction itself, set by the radar measurement, rather than a limitation of RIFT in particular. Two observations support this view.

First, RIFT does not need 2,400 views. We retrained it on B787 on nested subsets of 10% to 90% of the 2,400 training views (seed 42, 150 epochs; six of the ten subsets shown) and scored every run on the same 1,000 reserved-test views:

Training views (share of 2,400)Complex RelMSE (%)Range-power RelMSE (%)Chamfer (1e-4 m^2)F-score
240 (10%)88.4237.1223.850.2035
480 (20%)13.984.5332.2980.6394
720 (30%)3.1591.8240.54790.7849
960 (40%)1.0301.0400.51020.8136
1,440 (60%)0.43980.68080.49580.8245
2,400 (100%)0.33610.62160.43850.8686

With 480 views the complex RelMSE (13.98%) is already below that of SpINR-style (78.11%), GeRaF (174.1%) and Sugavanam-Ertin (19.98%) trained on all 2,400 views (Table 6a).

Second, the number of views a radar needs is set by the object's size in wavelengths. For a monostatic sensor, the two-way phase of a scatterer at cross-range offset xx changes with the look direction by 4πxΔθ/λ4\pi x \Delta\theta/\lambda, so the echo of an object with cross-range extent L decorrelates after an aspect change of θc=λ/(2L)\theta_c = \lambda/(2L): about 8.6 degrees for the longest cross-range extent of the 0.10 m objects (3.3 wavelengths at 10 GHz) and about 0.2 degrees for the GOTCHA vehicles (150 wavelengths at 9.6 GHz). Sampling the sphere at 8.6 degrees takes about 560 views, so 2,400 is roughly four times that count, and the error is already 1% at 960. The GOTCHA counts reflect the vehicles' electrical size and say nothing about the minimum; we have not run the view-count sweep on GOTCHA. We will state the view requirement in the main paper and in the limitations.

Weakness 4: We agree that two datasets limit what we can say about generality, and we make no claim beyond them. We chose them to differ as much as available data allow. The RIFT dataset is simulated with a commercial ray tracer (Ansys AVxcelerate RSS), not a simulator of our own. GOTCHA is measured by a real airborne radar: monostatic circular-SAR passes at about 10 km slant range, 9.6 GHz with 640 MHz of bandwidth, HH polarization (4 polarization are present in the dataset and we only used 1), eight passes at different elevations, and three vehicles about 150 wavelengths long in a cluttered scene, with the pose and phase errors of a real platform. GOTCHA therefore provides the SAR sampling a real radar actually records. On it RIFT leads on every column except IoU (three training seeds, three vehicles, means):

MethodTrain complex RelMSE (%)Validation complex RelMSE (%)Chamfer (m^2)Max. Hausdorff (m)HD95 (m)IoUF-score
RIFT40.6653.850.066160.92440.42540.081370.5281
SpINR-style73.9115.30.83313.1191.7010.051810.3343
BackprojectionN/AN/A1.1173.2712.6530.095100.4416

The two datasets differ in data origin, sampling geometry, stand-off, electrical size and bandwidth, and RIFT uses the same representation and forward model on both, with dataset-specific data terms and optimizer settings (Appendix D). We have not tested other sensors, such as automotive MIMO radars with narrow bandwidth and small apertures, or other capture geometries such as a linear trajectory; limited azimuth arcs are tested on the simulated B787 (see our answer to Weakness 5).

Our response continues in Part 3 of 7.

AuthorsOpenReview ↗

Official Comments to Reviewer omzY (Part 3 of 7, Round 1)

Weakness 5: We thank the reviewer for this list. Appendix H of the updated manuscript (Tables 10-15, summary in Section 5.2) covers most of it, and we will run the items it does not cover. Each Appendix H variant changes one choice of the main recipe on A320, B787 and the race car, with the same training command, and is scored on the 1,000 reserved-test views.

(a) Number of training views. The B787 sweep over 10% to 90% of the views, with NVS and geometry metrics, is the table in our answer to Weakness 3. We will run a three-point sweep (25, 50 and 75% of the training units) on the Camry.

(b) Angular distribution of views. On B787 we trained on 480 views concentrated in an azimuth window of 180 or 90 degrees centered on the nose and compared with 480 views spread over the full sphere, scoring all 1,000 test views inside and outside the window:

Training viewsComplex RelMSE inside / outside the window (%)Chamfer (1e-4 m^2)F-score
480 in a 90-degree window3.400 / 89.390.94190.6198
480 over the full sphere (90-degree region)9.970 / 14.382.2980.6394
480 in a 180-degree window7.953 / 76.170.78850.7125
480 over the full sphere (180-degree region)12.83 / 14.902.2980.6394

The Chamfer distance and F-score are for the whole object. We will add a planar 180-degree pass at the same number of views: the 480 training views of the 180-degree window with the smallest elevation (within about 25 degrees of the horizontal plane). All views are at the same distance, so only the angular footprint of a straight pass is represented.

(c)-(e) Spherical-harmonic order, adaptive allocation and loss terms. Appendix H (test complex RelMSE in percent, one run per object):

VariantA320B787Race car
RIFT (main run)0.31680.33610.4003
(a) No angular unlocks (maximum degree 0)1.2301.1841.318
(c) No spatial splits0.34970.36770.4143
(h) Uniform lattice, same point count, all bands up to degree 3 from the start0.45810.44220.5085
(i) Fixed isotropic lattice2.4102.0671.702
(d) No group-sparse prior11.9015.3418.93
(e) No SH-degree prior0.31660.33590.4002

The maximum degree on the RIFT dataset is 3; degree 0 only is 3.3-3.9 times worse. The objective is a data term and two penalties. We will run maximum degrees 1 and 2 on A320, B787 and the race car.

Runtime and memory. A benchmark of RIFT's forward model on one Intel Data Center GPU Max card (one view, 600 frequencies):

ScatterersForward, NUFFT / direct (s)Forward + backward, NUFFT / direct (s)Peak memory, forward + backward, NUFFT / directTraining update (s)Update peak memory (MB)
110,5920.041 / 0.136 (3.3x)0.080 / 0.194 (2.4x)57 MB / 1.61 GB0.085150
128,5820.045 / 0.144 (3.2x)0.089 / 0.202 (2.3x)62 MB / 1.90 GB0.099*163*
262,1440.080 / 0.149 (1.9x)0.160 / 0.220 (1.4x)65 MB / 3.82 GB0.173267
  • The two training-update entries of this row were measured with 125,000 active points in 128,582 slots; every other entry of the row is at 128,582 scatterers.

Direct is the summation over all scatterers; the training update is scene evaluation, NUFFT render, complex MSE, backward pass and an AdamW step, without the priors and the adaptive statistics. We will add a table of training time for all methods, with each method's update budget, to the paper.

Our response continues in Part 4 of 7.

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

RIFT considers explicitly complex-valued radar novel-view synthesis and 3D reconstruction. It combines learnable point scatterers, directional complex spherical-harmonic coefficients, and a radar forward model accelerated by NUFFT. The main contribution appears to be the gradient-guided adaptive spatial and angular refinement. The authors evaluate the method on six simulated objects and three vehicles from the measured GOTCHA dataset with the evaluation results on the validation data rather than independent test data.

Critical strengths
  1. Problem formulation: The paper tackles an interesting problem of predicting complex-valued radar measurements in novel views. This is critical as it preserves phase information which is valuable for coherent radar processing.

  2. Radar-only reconstruction: The method uses radar back-projection for initialization and learns the scene without external (e.g., RGB or LiDAR) supervision. The same learned field and selected checkpoint provide both novel-view predictions and a geometric readout. This is a useful capability, and reporting signal and geometry metrics separately helps assess both aspects.

  3. Representation and implementation: Combining explicit point scatterers, directional complex coefficients, adaptive spatial/angular capacity, and NUFFT evaluation is a reasonable design. The paper includes both simulated and measured radar experiments and provides substantial implementation details in the appendices.

Critical weaknesses
  1. Novelty: this is one of the biggest concerns. The radar/RF novel view synthesis literature is growing fast. The proposed paper appears to combine existing components: a) radar-only scene fitting in Radar Fields, b) directional scattering in DART, c) complex-valued RF synthesis in NeWRF, under various related settings. More closely, SH-SAS models complex directional scattering for synthetic aperture sonar, while GSRF (Yang et al., 2025), which is not cited, extends 3DGS to RF synthesis using explicit complex-valued primitives, directional coefficients, and adaptive spatial refinement. As a result, this paper appears to contribute more specifically in its radar-native point-scatterer formulation and its particular rules for allocating spatial resolution and angular capacity to individual scatterers, which appears to be more implementation details, rather than a fundamental breakthrough.

  2. Clarification and ablation: the paper lacks targeted ablations establishing the contributions of learned point locations, directional complex coefficients, spatial splitting, adaptive SH degree, and backprojection initialization. It is unclear how much of the improvement comes from the proposed adaptive refinement rules, versus the complex-valued objective, directional expressiveness, initialization, or increased representation capacity. Comparisons under the same radar forward model and loss, with comparable capacity and training budgets, would help establish the contribution of each component.

  3. Limitation: The forward model assumes scalar single scattering and uses frequency-independent directional coefficients. As stated in the paper, it lacks explicit treatment of occlusion, multiple scattering, diffraction, antenna patterns, polarization, and motion. It is unclear how reliably the learned field predicts unseen views when shadowing and multipath change with viewpoint, or when acquisition poses and phase calibration are imperfect.

  4. Evaluation: This is probably the weakest part of the paper. 4. a) In Appendix E, all six simulated objects are with a maximum dimension of 10 cm, observed at 10 m, and fitted using 2,400 training views from a full-sphere acquisition. At a carrier freq of10 GHz, this scaling substantially changes the electromagnetic size of the aircraft and vehicle models, i.e., the simulated objects are only 3.3 wavelengths wide. They provide limited evidence for realistic object dimensions, sparse acquisition trajectories, or changes in transmitter/receiver configuration (different sensor-to-object distance) 4. b) On GOTCHA, the reported signal results use the 30 validation units that also select the checkpoint; the 55 reserved azimuth sectors remain unevaluated in the reported results (Appendices D–E and Table 7). Validation-based checkpoint selection is standard, but the absence of a reported independent test evaluation limits the evidence for real-data novel-view performance.

Questions for authors
  1. As stated in the weakness, the authors fails to provide targeted ablations for learned versus fixed point locations, isotropic versus directional scattering, fixed versus adaptive SH degree, spatial refinement on/off, and backprojection initialization? These should use the same complex forward model and loss, report both signal and geometry metrics, and account for differences in final capacity and computation.

  2. Can the authors report native-complex prediction results on the 55 reserved GOTCHA azimuth sectors, using checkpoints selected only with the existing validation split? How does performance change when the held-out views form larger contiguous angular gaps?

  3. Why are all simulated objects scaled to 10 cm at most? Can the authors evaluate realistic object dimensions, reduced angular coverage, different sensor-to-object distances, and phase/pose perturbations?

  4. Can the authors also include a stronger complex-domain baseline such as SpINRv2 and provide convergence or compute comparisons?

6 comments
AuthorsOpenReview ↗

Official Comment to Reviewer mBBP (Part 1 of 3, Round 1)

Dear Reviewer mBBP,

We appreciate your constructive feedback on our work. Here is a point-by-point response to the weaknesses and the questions you brought up across 3 posts.

Weakness 1: We thank the reviewer for the careful comparison; we will cite and discuss GSRF and NeWRF in Related Work. We see the contribution as a combination that none of these methods provides: one explicit radar field, fitted to the complex frequency response with phase, and evaluated for both complex novel-view synthesis and 3D geometry.

  • Radar Fields fits received power on 2D automotive range-azimuth scans, so phase is not modelled, and its evaluated geometry is a bird's-eye-view map.
  • DART fits range-Doppler images from a moving radar in static scenes and requires radar motion and velocity.
  • NeWRF and GSRF synthesize wireless channel data (narrowband channels, spatial spectra, RSSI and CSI) and report no geometry metrics. They share complex-valued radiance, and GSRF shares adaptive densification.
  • SH-SAS is the closest: complex SH scattering fields with 3D geometry and complex novel-view synthesis, in sonar, stored in a hash grid and MLP with a fixed SH degree. RIFT stores complex SH coefficients on explicit point scatterers with adaptive spatial resolution and angular order, and uses a radar forward model.

Weakness 2: We thank the reviewer for this suggestion. In the updated manuscript, we included Appendix H (Tables 10-15) as the ablation study, with a summary in Section 5.2. We retrain RIFT on three RIFT-dataset objects (A320, B787 and the race car) with one design choice changed at a time. Every variant uses the training command of the main run (360,000 updates, minimum-validation checkpoint) and is scored on the reserved-test views for the signal and for the geometry.

In brief, relative to the main run, the reserved-test complex RelMSE rises 3.3-3.9 times when the angular unlocks are removed, by 27-45% for a uniform lattice with the same number of points and all SH bands active from the start, and 8.7 and 6.2 times on A320 and the race car (1.05 times on B787) when the backprojection start is replaced by zero, which supports these design choices. Spatial splitting changes the signal by 3-10% and learned offsets by 0.5% or less. Geometry does not always follow the signal, and we report both.

Weakness 3: We agree that RIFT does not represent occlusion explicitly, as stated in the paper. Occlusion and shadowing are represented implicitly, through the direction dependence of each point's response: a point hidden from some directions can learn a small response there, within the angular resolution of its spherical-harmonic degree, without explicit ray casting (Section 3). On the three ablation objects, the held-out views of the full sphere, where the shadowing pattern changes with viewpoint, are predicted at 0.32-0.40% reserved-test complex RelMSE (Appendix H, main run).

The imperfect acquisition poses and phase calibration the reviewer mentions are present in the GOTCHA dataset: it is measured airborne radar data, with clutter and car-ground multipath as well, and RIFT is evaluated on it. We load it with the dataset's published autofocus corrections (range and phase), applied once, and apply one estimated phase constant per pass and vehicle, from leave-one-pass-out fits at the vehicle, identically for every method and frozen before the reserved-sector scoring. RIFT leads on it: the validation complex RelMSE is 53.85% for RIFT against 115.3% for SpINR-style (Table 2, three-seed means). RIFT's global complex gain absorbs one scene-wide scale and phase offset but not viewpoint-dependent calibration errors, which can dominate the GOTCHA RelMSE (Limitations). A controlled sweep of pose and phase errors would quantify this further.

Weakness 4(a): We respectfully disagree that the evidence for realistic object dimensions and for different sensor-to-object distances is limited: the realistic dataset is GOTCHA, and we present the result. The RIFT dataset is a controlled simulation: six meshes scaled to at most 0.10 m (3.3 wavelengths) at a 10 m stand-off (Appendix E). GOTCHA is measured airborne data of real vehicles about 4.5-4.8 m long (roughly 150 wavelengths at 9.6 GHz) at a slant range of about 10 km, three orders of magnitude farther than the RIFT dataset; RIFT leads on all three vehicles (Table 2: validation complex RelMSE 53.85% against 115.3% for SpINR-style, three-seed means). The distance is fixed within each dataset and we do not vary it. For sparse trajectories, RIFT trained on 480 of the 2,400 B787 views (20%) spread over the sphere reaches 13.98% complex RelMSE on the 1,000 test views, better than SpINR-style on all 2,400 views (78.11%; Table 6a); concentrated in a 90- or 180-degree azimuth window, it does not extrapolate outside the window (our answer to Question 2).

Our response continues in Part 2 of 3.

AuthorsOpenReview ↗

Official Comment to Reviewer mBBP (Part 2 of 3, Round 1)

Weakness 4(b): We did not include the reserved-sector result in the paper because it poses a physically different and materially harder problem from the one the paper evaluates. For a monostatic sensor, the complex echo of a target of cross-range extent L decorrelates after an aspect change of θc=λ/2L\theta_c = \lambda/2L, the angle after which a fixed receiver sees a new speckle lobe (Shirley et al., 1992); for the vehicles at 9.6 GHz (λ\lambda = 3.1 cm), θc\theta_c is about 0.2 degrees. θc\theta_c Our validation units are close to training in this sense: the nearest training look is about a quarter of θc\theta_c away, because the same sector is in training in the other seven passes (within pass 4 alone, the gap would be about one θc\theta_c). The 55 reserved sectors are held-out azimuth gaps of 1 to 3 degrees, spread around the full circle and sealed in all 8 passes, so no pass observes them: their median gap is about 2.5 times that of validation, about one θc\theta_c at each unit's own azimuth. Predicting them means carrying the phase of the response across about one decorrelation angle, where interpolating from neighboring looks no longer works and only a model that tracks the scatterer positions to a small fraction of the wavelength (a quarter wavelength is 0.8 cm) can, an accuracy that the pose and phase calibration of airborne data limits.

Look-angle gap to the nearest training lookValidation units (pass 4; 30 per vehicle)Reserved-sector units (440 per vehicle)
Median gap (degrees)0.080.20
Maximum gap (degrees)0.190.92
Units within 0.1 degrees of a training look17 of 300 of 440
Median gap divided by θc\theta_c at the unit's own azimuth (Camry, Sentra, Santa Fe)0.26, 0.24, 0.250.98, 0.92, 0.93
Units with a gap above θc\theta_c (Camry, Sentra, Santa Fe)2, 0, 0 of 30211, 161, 166 of 440 (48%, 37%, 38%)

We nevertheless ran it: after selecting the checkpoints on validation, we froze the selection (the validation controls reproduce the paper's values) and scored the 55 reserved sectors of all 8 passes once, without further tuning. The complex RelMSE (%, native response) of the same frozen seed-42 checkpoints is:

CamrySentraSanta FeMean
RIFT, validation51.4157.9553.5454.30
RIFT, reserved sectors90.0891.4094.3991.96
SpINR-style, validation115.2111.4121.0115.9
SpINR-style, reserved sectors150.0150.2189.1163.1

Even on this harder problem RIFT is best on every vehicle and stays below the 100% level of a zero prediction, whereas SpINR-style exceeds it on every vehicle. RIFT's prediction remains correlated with the measurement (correlation 0.28 to 0.33, against 0.66 to 0.70 on validation), whereas SpINR-style's is essentially uncorrelated (0.05 to 0.07), which is why its error exceeds the zero-prediction level.

Our response continues in Part 3 of 3.

AuthorsOpenReview ↗

Official Comment to Reviewer mBBP (Part 3 of 3, Round 1)

Q1 (targeted ablations). Please refer to our response to Weakness 2. Appendix H (Tables 10-14) and Section 5.2 report these ablations (learned versus fixed point locations, isotropic versus directional scattering, fixed versus adaptive SH degree, spatial refinement on or off, and backprojection initialization) with the same forward model and loss and signal and geometry metrics on two lattices. We will add the final point count, the number of active spherical-harmonic coefficients and the training time of every variant to Appendix H.

Q2 (reserved sectors and larger gaps). The reserved-sector result, with checkpoints selected only on the validation split and frozen before scoring, is in our response to Weakness 4(b). The reserved sectors are held-out gaps of 1 to 3 degrees (the longest run is 3 degrees). Larger contiguous gaps are the subject of the azimuth-window test on the simulated B787 described under 4(a), in which views outside a 180-degree or 90-degree training window are held out. RIFT does not predict the unseen azimuths in that test (complex RelMSE of 76.2% and 89.4% outside the 180- and 90-degree windows). We do not claim viewpoint extrapolation, which is materially harder than interpolation: the held-out views of the paper are drawn by a random permutation among the training views on the sphere (RIFT dataset) or lie between neighboring passes (GOTCHA), as stated in Appendix E. Large contiguous gaps remain a limitation.

Q3 (scale, coverage, distance, perturbations). The RIFT dataset meshes are scaled to the same maximum dimension, 0.10 m, at a 10 m stand-off, so that the 3D reconstruction of the different objects can be compared fairly. Distance metrics such as the Hausdorff distance (maximum and 95th percentile) and the Chamfer distance are in absolute length units, so they grow with the size of an object and would not be comparable, or meaningful to average, across objects of different sizes. For realistic object dimensions, reduced angular coverage and the sensor-to-object distance, please refer to our response to Weakness 4(a), where the measured GOTCHA data provide real vehicle sizes and a very different range; for pose and phase imperfections, which the measured GOTCHA data contain, please refer to our response to Weakness 3.

Q4 (SpINRv2). We implemented a SpINRv2-style model (SpINRv2's warm-up and regularization on our SpINR-style implementation) and are training it on the B787 with the same 150-epoch protocol; we will report the result. As a first cost figure, in a benchmark of our forward model (110,592 to 262,144 point scatterers, one view), the NUFFT range operator is 1.9–3.3× faster than direct summation over the scatterers (1.4–2.4× including the backward pass), and its peak memory including the backward pass is 57–65 MB against 1.6–3.8 GB.

We thank the reviewer again for their thoughtful feedback, which has helped us improve the manuscript.

Reference:

L. G. Shirley, E. D. Ariel, G. R. Hallerman, H. C. Payson and J. R. Vivilecchia, "Advanced Techniques for Target Discrimination Using Laser Speckle", The Lincoln Laboratory Journal, vol. 5, no. 3, pp. 367-440, 1992.

Best, Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Updates in the Manuscript (Revision 1)

To begin with, we appreciate the reviewers for the constructive comments. Here, we present an updated manuscript of our work where some questions and weaknesses the reviewers suggested are already addressed in preparation for this update. We will update the manuscript again near the end of the author-reviewer discussion period. Here is a summary of the updates we made in the new upload.

Fair comparison under training stochasticity (Reviewer XE1v, W3 and the question on the adapted baselines). We re-ran RIFT and the four learned baselines under the same number of gradient updates (360,000 updates) and three training seeds, scoring the minimum-validation checkpoint of each run; the protocol is in Appendix E and the baseline adaptations are in Appendix F. We also ran GeRaF with its model step advancing (Appendix F; see our response to XE1v). Tables 1 and 2 now report three-seed means, the per-seed tables are in Appendix G, and the re-run baselines replace the corresponding rows of Table 1 and their view panels. RIFT still performs best after this update despite slight number changes: in every column of Table 1 on the RIFT dataset and, on GOTCHA, in every column of Table 2 but IoU.

Ablations and further evidence (Reviewers mBBP W2 and Q1, omzY W5, XE1v W1). We added an ablation study on a subset of the RIFT dataset, three objects (A320, B787 and the race car), where each variant changes one design choice of the main recipe and is scored on the reserved-test views (Appendix H and Section 5.2). Removing the angular unlocks makes the held-out complex error 3.3-3.9 times larger and removing the group-sparse prior 38-47 times larger; the geometry does not always follow the signal. The other evidence the reviewers asked for is in our individual responses.

AuthorsOpenReview ↗

Summary of Updates in the Manuscript (Revision 2)

Dear Reviewers,

Here is a summary of the updates in the second upload of the manuscript, made in response to your comments and to our individual responses. The first upload is described in our comment "Summary of Updates in the Manuscript (Revision 1)".

Related work (Reviewers mBBP W1 and XE1v W1). We added GSRF, NeWRF, RFDT and mmIR to Related Work: NeWRF and GSRF address wireless channel estimation, and RFDT and mmIR fit differentiable mesh-based simulators with non-radar scene priors, a LiDAR-derived mesh in mmIR and optional shape priors in RFDT.

Number of views (Reviewer omzY W3). Section 4 now states the number of training views (2,400 per RIFT-dataset object, 2,410 azimuth units per GOTCHA vehicle), the Limitations name too few training views as a failure mode, and Appendix H gives the B787 results with 240 to 2,400 views (Table 17).

Reconstruction comparison (Reviewer XE1v W2). A compact comparison of the side-view and plan-view reconstructions of B787 with the baselines is now in the main text (Figure 2).

Ablations and capacity (Reviewers mBBP W2 and Q1, omzY W5). Appendix H now has Tables 10-17. Table 16 gives the final number of active points, the number of complex spherical-harmonic coefficients and the training time of every variant. The maximum spherical-harmonic degrees 1 and 2 are added to the ablation tables and figures as variants (L1) and (L2).

Again, we thank Reviewers mBBP, omzY and XE1v for their comments, which led to each update above.

Best, Authors

AuthorsOpenReview ↗

Summary of Updates in the Manuscript (Revision 3)

Dear Reviewers,

Here is a summary of the updates in the third upload of the manuscript, made in response to the follow-up comment of Reviewer mBBP. The earlier uploads are described in our comments "Summary of Updates in the Manuscript (Revision 1)" and "Summary of Updates in the Manuscript (Revision 2)".

Comparison with SH-SAS (Reviewer mBBP). Appendix H now states that the neural-field arms (G) and (G-bp) are a radar adaptation of the SH-SAS representation: the complex spherical-harmonic coefficients of RIFT's point scatterers are stored in a hash grid and an MLP of the kind used in SH-SAS (16 levels, finest resolution 4,096, two hidden layers of width 32, spherical-harmonic degree 3) and rendered by RIFT's radar forward model, and the sonar renderer, pulse deconvolution, transmission and surface-normal terms of SH-SAS are not used because of incompatibility in data modality (Radar/Sonar). Appendix H also describes the view panels of the two arms.

Related work (Reviewer mBBP). Related Work describes NeWRF and GSRF as applying radiance fields and Gaussian splatting to wireless channel prediction and data synthesis (channel responses, signal strength and spatial spectra).

We also corrected two typographical errors in Sections 1 and 3.

We thank Reviewer mBBP for the follow-up comment that led to these updates.

Best, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.