- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes SPIRIT, a physics-informed inverse-rendering framework for 3D reconstruction from severely undersampled FMCW SAR measurements. The method combines differentiable backprojection, a structured surface representation, coherent RF rendering, and a two-stage optimization strategy to improve geometric stability during phase-sensitive reconstruction.
Experiments on synthetic data with 1, 3, and 6 scan lines demonstrate improved geometric and image reconstruction compared with the evaluated baselines. On the 1-scan benchmark, SPIRIT achieves 23.1 mm Chamfer Distance (CD), compared with 48.0 mm for the strongest geometric baseline. Additional evaluation on real MITO measurements reports 3.15 mm CD after residual adaptation.
Critical strengths
-
The paper addresses the challenging problem of recovering 3D geometry from extremely sparse SAR measurements. Rather than simply enhancing reconstructed images, SPIRIT explicitly models geometric structure and coherent wave propagation. The reported improvements are substantial: under the 1-scan setting, CD decreases from 48.0 mm for the strongest evaluated baseline to 23.1 mm. The real MITO evaluation provides additional evidence of practical potential.
-
The two-stage optimization is well motivated by the sensitivity of radar phase to geometric errors. Table 3 shows that full Stage II unfreezing without the refiner results in substantially worse geometric accuracy (53.3 mm CD), whereas the final configuration with the refiner achieves 23.1 mm CD and higher PSNR (25.3 vs. 22.1 dB). This highlights an important limitation of optimizing signal fidelity alone.
Critical weaknesses
The paper does not clearly establish how the differentiable renderer is connected to the FMCW signal preprocessing pipeline. Appendix E.2 states that the measurements undergo a fast-time FFT, whereas Equations (8) and (35) formulate the renderer using per-wavenumber propagation. The exact mapping between these representations should be clarified to establish signal-domain consistency.
Furthermore, the complete Stage I/II loss functions, their weights, the geometric regularizers, and the parameterization of the complex scattering coefficients are not sufficiently specified. In particular, it remains unclear whether Stage I uses only sparse observations or also full-aperture or ground-truth geometric supervision.
These details are important for verifying the physical formulation and reproducing the proposed optimization strategy.
The baseline comparisons do not fully isolate the contribution of the structured surface representation. SPIRIT employs a dataset-trained backbone with approximately 122.2M parameters, whereas the implicit and volumetric neural-field baselines are optimized independently for each test instance with a budget of only 100 iterations. The height-map baseline is also restricted to a fixed plane without geometric deformation.
Consequently, the improvements cannot be attributed specifically to the proposed representation or training strategy. This is particularly relevant given related work on differentiable SAR rendering (Wilmanski and Tamir, 2023) and RF geometry reconstruction (GeRaF, 2025).
The real-data evaluation is also limited to 10 independent test objects. The paper does not establish whether these objects overlap with the synthetic mesh training corpus, nor does it clearly specify the sparsity ratio used in Table 4. These omissions limit the extent to which the real-data results establish performance under extreme aperture sparsity and generalization beyond the evaluated object split.
Questions for authors
-
How are the FFT-domain FMCW measurements mapped to the per-wavenumber signals used by the differentiable renderer? Could the authors provide the discrete formulation or a simple point-scatterer consistency test?
-
What are the exact Stage I/II loss functions and coefficients? Does Stage I use full-aperture measurements or ground-truth 3D geometry, or only sparse observations and regularization?
-
Could the authors provide a stronger comparison with a deformable height-field or amortized implicit-surface baseline under matched training data, supervision, and computational budgets?
-
What exact sparsity ratio is used for the MITO experiments? Are the real test objects excluded from the synthetic mesh training corpus, and are signal errors evaluated on observed or withheld aperture positions?