acceptodds
Withdrawn from ICLR 2027

COHERENT SURFACE GAUSSIANS FOR A SPECULAR SYNTHETIC APERTURE

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Millimeter-wave radar sees through packaging, fabric and clutter, which makes it an attractive sense for robots, but its images are not shapes. We trace the reason to the optics of a synthetic aperture: each position is a co-located source and detector that receives glints only from surface points whose normals point back at it. Each sub-aperture therefore observes its own band of normals, back-projection images a convex surface behind its front, at a depth we predict in closed form, and the phase history determines a glint's center of curvature and its radius minus the radar's range offset, but not the two separately. We represent the scene by coherent surface Gaussians, curved primitives whose near-field echo has a closed form under a quadratic range approximation and whose angular response follows from their size and curvature. The closed form matches quadrature over the same primitives to within 1.4%, while flat Gaussians err by 24–98%. The same optics predicts that holding out a sub-aperture, a natural analogue of novel-view synthesis, is a weak learning signal, because the other sub-apertures see different normals; in a ten-seed physical-optics study it removes phantoms only by stopping early and losing surface, whereas generic cross-validation selects models reliably. Multipath ghosts seen by all sub-apertures are virtual glints and cannot be rejected by any hold-out. An encoder trained in simulation reduces the phantom rate of back-projection by an order of magnitude on unseen shapes, and a measured 77–81 GHz scan follows the predicted streak geometry.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Gaussian Splatting, Synthetic Aperture Radar, Near-Field Radar, Millimeter-Wave Sensing, Differentiable Rendering, Physics-Informed Learning, Inverse Problems, Sub-Aperture Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YXkGOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper addresses the issue of information loss and degraded imaging quality caused by specular reflections in near-field millimeter-wave (77–81 GHz) synthetic aperture radar (SAR) imaging. To tackle this, the authors propose a Gaussian-based representation framework. The core intuition is that since the sub-apertures capable of observing a surface patch depend on its size and curvature, the scene can be represented using "coherent surface Gaussians." Each primitive is a small surface patch with explicit curvature and a complex reflectivity; its backscattered signal yields a closed-form solution under a second-order range expansion (Theorem 2), with its angular response derived directly from geometry rather than learned. Reconstruction is performed in the raw echo domain by optimizing these surface patches so that the echoes predicted by the forward renderer match the measured signals.

The authors also introduce several additional modeling techniques and heuristics. However, despite reading these parts multiple times, I found the remainder of the paper virtually impossible to follow—it is exceedingly dense, over-complicated, and obscure. While I regret saying this, I believe it is remarkably difficult for a human reader to parse and digest the presented methodology.

Furthermore, the experimental validation is conducted almost on simulated data, and the manuscript lacks any qualitative imaging results on real-world objects.

Critical strengths
  • The paper features highly polished and visually appealing figures.

  • The analysis presented in the manuscript appears remarkably thorough.

  • I agree with the core motivation: specular reflection is indeed a critical challenge in mmWave SAR imaging. Unfortunately, despite reading the abstract and introduction carefully multiple times, I still struggle to grasp the underlying logic—including what exact problem is being solved, how the proposed method works step by step, and how effectively it performs on real-world data.

Critical weaknesses

This paper leaves me with many puzzling concerns:

  1. Lack of Real-World Object Reconstruction/Imaging Results: For a paper targeting shape reconstruction in near-field millimeter-wave (mmWave) SAR, what readers expect most is a clear result showing "measured real-world object →\rightarrow surface reconstructed by the proposed method." However, across all 45 pages, the only figures associated with measured data are Fig. 1(b, c) and Fig. 6, both of which merely display standard Backprojection (BP) images and their fringe geometry, rather than the reconstruction outputs of the proposed method. Meanwhile, the per-object fitting results in Fig. 12 stem from 2D simulations. For a SAR imaging paper, this manner of presenting results is very strange. I suggest that in future submissions, the authors include at least one real-world scan of an object with known geometry, demonstrating the reconstruction results of both the proposed method and baselines alongside comparisons with the ground truth. This would serve as the most direct evidence to support the claims of the paper.

  2. Visually Polished Figures with Elementary Flaws: While the figures appear high-end, they contain basic errors. For example, in Figure 1, labels like "five sub-apertures" and "height" are visibly occluded. Furthermore, the introduction states that "Figure 1(b) shows a target scanned by a 77–81 GHz rail radar," but based on my domain experience, Figure 1(b) does not look like the imaging result of the depicted target. There are many similar issues throughout the paper—the figures look sophisticated at first glance, but are fraught with problems upon closer inspection.

  3. Unusual Terminology and Conceptual Metaphors: The paper contains numerous strange expressions, such as "compare the optics with a measured scan." The authors continuously attempt to frame SAR imaging as "optics," which I find quite unprofessional as there are fundamental physical differences between the two.

  4. Extremely Obscure and Hard to Follow: The main text introduces a massive number of symbols, equations, and definitions, yet lacks a main thread to connect the underlying logic. As a result, even though I am very familiar with this field, I had to repeatedly flip back and forth between the main body and the appendix just to piece together the chain of reasoning, yet many details remain obscure. I strongly advise a complete rewrite of the paper to highlight a clear main storyline.

In summary, I am not personally opposed to AI auto research. However, the premise is that AI-written papers must remain meaningful, provide convincing experimental results, and be as understandable and readable to humans as human-authored papers. Unfortunately, this paper fails to satisfy these three criteria (or at best only fulfills the first one).

Therefore, I recommend Rejecting this paper.

Questions for authors
  • Reconstructing real-world objects is crucial; the authors should supplement the paper with qualitative imaging/reconstruction results on real-world experimental data.

  • What is the core narrative and main storyline of this paper? How effective is the proposed method on actual real-world SAR data? I am genuinely curious to see how the proposed framework performs when applied to real SAR measurements.

  • Has the encoder undergone ablation experiments regarding set attention, the number of sub-apertures, and different input representations (e.g., Backprojection volumes vs. raw echo signals)?

  • I strongly advise the authors to significantly streamline and condense the appendix.

Reviewer_z8shOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes SPEC GS, a method for reconstructing radar-visible surfaces from near-field millimeter-wave synthetic-aperture measurements. It analyzes how specular reflection, aperture geometry, and range offsets affect the apparent locations of surfaces in radar images. The method represents surfaces using curved Gaussian patches with complex reflectivity and derives an efficient approximate closed-form renderer whose directional response follows from patch geometry. A learned encoder initializes the reconstruction, followed by measurement-domain refinement. The paper also examines when held-out measurements can support model selection and why they cannot reliably reject certain multipath artifacts. Experiments assess rendering accuracy, reconstruction quality, computational efficiency, and failure cases.

Critical strengths
  1. The paper addresses an important problem in radar reconstruction. Bright responses and streaks in radar images do not necessarily correspond to physical surfaces. Understanding these effects and recovering reliable surface geometry is valuable for radar sensing.

  2. The physical analysis provides useful insights. The paper explains why different sub-apertures place the same curved reflector at different apparent locations, and why full-aperture imaging can place a peak behind the actual surface. It relates these effects to aperture diversity, bandwidth, and surface curvature, with numerical validation.

  3. The surface representation is well motivated. Each Gaussian models local geometry, curvature, and complex reflectivity. Its directional response follows from its geometry, creating a clear connection between surface shape and predicted radar measurements. The comparisons with flat Gaussians support the importance of modeling curvature.

  4. The proposed renderer offers a useful balance between accuracy and speed. The closed-form approximation closely matches dense numerical integration over the same Gaussian primitives. The paper also reduces computation by skipping negligible contributions, reporting substantial acceleration with little additional error.

  5. The analysis of geometric ambiguity is valuable. The paper explains how a fixed system range offset can be confused with changes in surface geometry. This clarifies why accurate signal fitting alone does not guarantee accurate reconstruction and provides practical guidance for calibration.

  6. The experiments are extensive and closely connected to the main claims. They cover rendering accuracy, runtime, reconstruction quality, unseen shapes, training objectives, validation strategies, and multipath failure cases. The ablations examine several important design choices, while independently generated physical-optics simulations strengthen the evaluation. The results also show that network initialization followed by brief refinement can approach the quality of much slower per-scene optimization.

Critical weaknesses
  1. Despite extensive experiments, real-world coverage. The measured targets described in the paper are a pipe-like object and a trihedral corner reflector, both captured using the same radar scanning setup. These experiments help validate the specular reflection analysis, but provide limited evidence of reliability for complex objects, multiple occluding objects, or different acquisition conditions.
  2. The integration formula is accurate, but the overall surface representation still introduces substantial error. The 0.1%–0.4% errors in Table 1 are measured against dense numerical integration over the same Gaussian primitives, primarily validating the integration formula itself. Against physical-optics returns from the actual surfaces, the total errors reach 9.1% for the pipe and 15.8% for the plate, indicating remaining errors from approximating surfaces and hard edges with Gaussians.
Questions for authors
  1. Choice of Gaussian representation. Beyond enabling closed-form integration, what advantages do curved Gaussian primitives offer over alternative radar surface representations, such as triangular meshes or oriented points? Given the surface approximation and hard-edge errors reported in Table 1, could the authors clarify the trade-offs that motivated this choice?
  2. Calibration procedure. How was the fixed range offset calibrated in the real experiments? Appendix D states that all baselines use the same calibrated offset, whereas Appendix H describes the corner reflector as providing only a qualitative check of its sign. Could the authors reconcile these descriptions and specify the reference measurement, the estimated offset, and its uncertainty?
Reviewer_VGWEOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies surface reconstruction from specular near-field SAR. It proposes coherent surface Gaussians, analyzes the synthetic aperture as a co-located specular sensor and its held-out prediction.

Critical strengths

a. The paper analyzes the physics of specular SAR, providing useful insights into why accurate prediction may still fail in geometry correctness.

Critical weaknesses

a. The machine learning contribution in this paper is not clearly demonstrated. Most parts of the paper are analysis of radar imaging and its underlying physics. While the authors mentioned that held-out and retained measurements should observe the same underlying structure (contribution 3), it is not developed into a more general ML principle. b. The experimental setup is not described transparently. E.g. for table 4, the paper does not specify the measured objects and how ground-truth surfaces are obtained. Besides, the paper is difficult to follow as I do not see a clear central thread. c. The paper uses the wrong template “Under review as a conference paper at ICLR 2026” and exceeds the page limit (table 4 of the main text appears on page 10).

Questions for authors

a. Most experiments, including table 2 and 3, are conducted in 2D simulation. Will the conclusions of this paper hold in 3D? b. How would the method change when roughness of the surface is comparable to the wavelength where diffuse scattering is significant?

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

Related papers

Discussion (0)

Sign in to comment.