acceptodds
Under review as a conference paper at ICLR 2027

Geometry First, Scattering When Reliable: Sparse-View SAR Target Generation

Abstract

Sparse-view Synthetic Aperture Radar (SAR) undersamples the azimuth-depression space, making unseen-view generation severely under-constrained as view changes jointly alter coarse image-domain geometry and anisotropic electromagnetic scattering response. Yet coarse geometry admits a relatively stable cross-view approximation, whereas anisotropic scattering remains poorly determined across large angular gaps due to view-specific radar-target interactions. We therefore propose a geometry-first, scattering-when-reliable principle: coarse geometry is made explicit as a conditioning signal, while scattering information is introduced only when sufficient angular evidence supports its query-view variation. We instantiate this principle with -Flow. For geometry, -Flow represents coarse geometry as spatial support, constructing Canonical Spatial Support (CSS) from sparse observations and aligning it to the query view as Query-Aligned Spatial Support (QSS). For scattering, we first investigate which response information remains useful beyond QSS, and accordingly introduce Compatible Scattering Transfer (CST), which transfers relative, query-dependent scattering modulation from a compatible reference with denser angular coverage. QSS, optionally complemented by CST, conditions a flow-matching (FM) generator for unseen-view synthesis. Geometry ablations show that QSS provides query-aligned geometry that model scaling cannot recover. Scattering analyses reveal an inferability-utility trade-off: response information with greater complementary value beyond QSS is harder to infer from sparse observations. CST mitigates this limitation, consistently reducing azimuth error at unseen azimuths and even when the query depression is entirely unobserved for the target. Evaluated on MSTAR and ATRNet-STAR, -Flow achieves state-of-the-art generation quality over all sparsity levels and the best recognition performance under the largest sparsity setting, with only 1.63% of the FM baseline's parameters.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Synthetic Aperture Radar, Sparse-View Generation, Flow Matching, Novel-view synthesis

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Nfn4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies SAR target generation at unobserved views from sparsely sampled viewing angles. The proposed Φ-Flow aligns observed responses in a canonical frame, constructs geometric support, and obtains query-view conditions through depression-angle interpolation and azimuth rotation. When a compatible, densely observed reference target is available, it additionally incorporates scattering modulation. A conditional flow-matching model generates images for recognition data augmentation. The main contribution is to distinguish the roles of geometric support and fine-grained scattering under sparse observations and demonstrate a compact generator built around geometry conditioning.

Critical strengths

1. The most convincing part of this paper is the conditioning ablation in Table 1. With generator size fixed, recognition accuracy is 34.80% without QSS, 56.97% with the aligned mask, and 58.15% with QSS. The aligned mask already accounts for much of the improvement. This gives concrete support to the motivation of first placing the target in the right geometric configuration and shows that the benefit is not simply a consequence of increasing network capacity. The design is relatively straightforward, but it has clear value when observations are sparse.

2. The main experiments support this observation beyond a single ablation. Φ-Flow substantially improves recognition over FM in the 60° settings on both MSTAR and ATRNet-STAR, while using 3.87M rather than 237.26M parameters. Reporting classification and angular errors separately is also useful. For example, CST mainly improves azimuth error in Table 2 without improving classification accuracy, and the text interprets this accordingly. This makes the contribution of the conditioning information easier to understand than a blanket claim of improvement across all aspects of generation.

Critical weaknesses

1. The operational criterion for determining scattering-reference reliability remains unclear. Section 4.3.2 selects two category pairs with similar cross-view scattering behavior, but the main text does not explain how that similarity is established or how much target and reference data must be available beforehand. For a new target with sparse observations, can a suitable reference be selected from those observations, or does selection require prior knowledge of its scattering at unobserved angles? The latter would substantially restrict CST's applicability. This is not an allegation that test information was used; the missing piece is a selection criterion that can be applied before CST. The QSS-only results remain valid, but the scattering component is currently demonstrated primarily on preselected compatible categories.

2. The analysis supports the usefulness of geometry more strongly than it establishes why each part of the complete design is needed. The aligned mask and QSS already perform similarly, yet the main comparisons do not directly show the additional benefit of conditional generation over simple geometric augmentation using the same angle information. Figure 3(b) also needs a bounded interpretation: inferability depends on the tested angular rules, whereas utility comes from an oracle experiment using the measured query response. This is a useful diagnostic, but it does not establish that a learned scattering predictor must face the same trade-off to the same extent. These issues do not overturn the reported results, but they limit the scope of the conclusions.

Questions for authors

1. Please specify the criteria used to select the two reference-category pairs. In particular, which target observations were available during selection, and how would CST be enabled or disabled for a new category whose scattering behavior is unknown? If the intended setting assumes access to a known compatible reference, please make that assumption explicit.

2. Which existing comparisons best distinguish the benefit of geometric alignment from that of conditional generation? Please also clarify the angle information available to the baselines in Tables 4 and 5. For Figure 3(b), please clarify whether the intended conclusion is specific to the current inference rules or establishes a broader limitation, since this distinction affects the scope of the contribution.

2 comments
AuthorsOpenReview ↗

Response to reviewer Nfn4 (W1&Q1) (1/2)

W1&Q1

Thank you for raising these questions. We clarify the original pair-selection procedure below and have added experiments selecting references using only sparse target training observations.

Q1.1: How were the original pairs selected, and which target data were available?

As described in Appendix A.2, the original pairs were screened using complete measured target and reference sequences, comparing their image appearance and view-dependent response changes after canonical alignment. The additional target views used for screening did not enter CSS construction, CST calibration, or FM training.

Q1.2: Can a reference be selected without knowing the target’s responses at unobserved angles?

Yes. The added procedure matches references using only the target’s sparse training views. Each target provides 5–12 training images per observed depression, while more densely observed reference sequences supply additional angular response evidence. The selected references reduce mean azimuth MAE relative to QSS-only in all four original CST protocols, demonstrating useful reference selection without access to the target’s missing views.

The table reports five-seed mean azimuth MAE (°), with changes from QSS-only in parentheses, calculated before rounding. Negative changes indicate improvement; column angles denote source-azimuth spacing.

ConditionAz. generation, 30°Az. generation, 60°Dep. interpolation, 30°Dep. interpolation, 60°
QSS-only25.51 (—)35.05 (—)39.66 (—)53.64 (—)
Sparse-selected CST22.18 (−3.33)33.09 (−1.97)37.43 (−2.23)46.01 (−7.63)

Complete results and uncertainties are provided in Appendix E (Tables [12-13], lines [1458-1499]).

Q1.3: What criterion determines whether CST is enabled for a new target?

CST is enabled when reference transfer improves response reconstruction on held-out observed target views. We select the reference-transfer strength, which controls how much the reference changes the target’s CSS, by minimizing this validation error, including zero as the no-transfer option. If zero is selected, generation uses QSS alone (Appendix B.2.1).

We also added the discussions of compatible reference selection in the main text (Section 4.3.3, Table 4, Line 426).

AuthorsOpenReview ↗

Response to reviewer Nfn4 (W2&Q2) (2/2)

W2&Q2

Thank you for these questions. We address the contributions of alignment and generation, the baselines’ angle information, and the scope of Figure 3(b) separately below.

Q2.1: Which comparisons distinguish geometric alignment from conditional generation?

We agree that query-aligned geometry accounts for much of the improvement in Table 1(b): recognition accuracy increases from 34.80% without QSS to 56.97% with the aligned mask, while the full QSS reaches 58.15%.

To test whether generation adds value beyond alignment, we additionally use the query-aligned QSS templates directly as augmentation images, without a generator. This comparison retains the same sparse observations, query angles, augmentation counts, downstream settings, and real test sets. Φ-Flow achieves higher mean recognition accuracy and lower mean azimuth-vector MSE in all six listed protocols:

DatasetSpacingQSS template Acc. (%) ↑Φ-Flow Acc. (%) ↑QSS template Az. MSE ↓Φ-Flow Az. MSE ↓
MSTAR30°61.17 ± 2.1180.23 ± 1.060.2614 ± 0.01110.1745 ± 0.0205
MSTAR60°39.58 ± 2.3061.06 ± 1.150.3790 ± 0.02220.3349 ± 0.0167
MSTAR120°35.26 ± 4.5045.25 ± 2.400.6313 ± 0.03050.4794 ± 0.0286
ATRNet-STAR60°41.02 ± 3.5357.68 ± 2.270.5987 ± 0.02320.5238 ± 0.0275
ATRNet-STAR90°47.93 ± 2.2649.86 ± 3.260.5689 ± 0.03820.5087 ± 0.0206
ATRNet-STAR120°31.72 ± 5.1750.16 ± 3.200.6111 ± 0.01840.6036 ± 0.0346

Values are mean ± sample standard deviation over five downstream seeds.

Figure 11 (Appendix C.3) shows that QSS provides coarse spatial support, while Φ-Flow better reproduces the measured query’s bright-response distribution through conditional generation.

Q2.2: What angle information is available to the baselines in Tables 4 and 5?

All baselines receive angle-label information for the requested view. ACGAN, Φ-GAN, Φ-StyleGAN, cDDPM, LDM, and FM receive class labels and encodings of the requested azimuth-angle labels. For depression interpolation, the requested depression-angle labels are additionally included. SAR-NV receives an observed SAR image and labels specifying the relative azimuth and depression from that observation to the query view. The architecture-specific encodings and conditioning mechanisms are detailed in Appendix B.4.

Q2.3: Does Figure 3(b) establish a general limitation of scattering prediction?

Thank you for asking us to clarify the scope of this conclusion. We have clarified in Section 4.3.1 that the finding concerns the tested angular inference rules, rather than a general limitation on learned scattering predictors. The revised text distinguishes relative inference improvement over CSS (SrS_r) from complementary generation utility when the measured query representation is supplied (UrU_r). The key observation is that response information with substantial complementary value remains poorly recovered by these rules, motivating CST to obtain additional angular evidence.

Reviewer_zQ43OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an SAR image-generation framework based on the key idea of "Geometry First, Scattering When Reliable". It addresses novel-view SAR image generation from sparsely sampled azimuth and depression angles. Its design is motivated by the insight that coarse image-domain geometry can be approximately aligned across views, whereas anisotropic electromagnetic scattering is difficult to infer over large angular gaps. The proposed Phi-Flow explicitly represents geometry and introduces scattering information only when a densely measured, compatible reference target is available.

Critical strengths
  • The geometry-first insight is well motivated and technically sound. It is well matched to the SAR problem.

  • The experiments were conducted on two SAR datasets and covered multiple levels of angular sparsity, cross-scene testing, downstream classification, and viewpoint estimation.

  • The paper presents useful ablation studies on model capacity, alignment, amplitude, incorrect spatial conditions, and response representations.

Critical weaknesses
  • The visual results (Figures 4 & 5) for the generated SAR images are not impressive. It is not clear how the generated images (given its low resolution) will be useful for downstream applications in practice.

  • All experiments involve stationary ground vehicles. No other objects (e.g., aircraft, bus, drones, etc.) were evaluated. Thus, the scope is narrow.

  • The generalizability of the proposed framework is unclear. It does not provide sufficient evidence that the model can synthesize views of a previously unseen target instance.

Questions for authors

The paper presents a novel-view SAR image generation framework with a clear, problem-specific insight, strong ablations for QSS, and impressive computational efficiency. However, it appears that the main cross-model gains stem from handcrafted query-aligned geometry, while CST (i.e., the more distinctive scattering contribution) requires a densely-measured reference selected using information that would be unavailable in a realistic sparse-view deployment. In addition, I have the following concerns.

  • The visual results (Figures 4 & 5) for the generated SAR images are not impressive. The images are obscure and difficult to interpret visually. It is unclear whether these results will be useful for downstream applications in practice.

  • As acknowledged in the paper, all experiments involve stationary ground vehicles. There is no evaluation on aircraft, ships, buildings, articulated objects, or moving targets. Also, there is no evaluation for a cross-sensor, cross-frequency, or cross-polarization study.

  • CSS is aggregated by target class and depression angle, and the generator receives the class label. It is unclear whether the method can synthesize views of a previously unseen target instance or category from its sparse observations.

  • CST appears to rely on the strong assumption that the reference target is measured on a dense 5-degree azimuth grid and that the dimensions of both the target and reference are available for registration. Such a reference library may not be available in practice.

4 comments
AuthorsOpenReview ↗

Response to reviewer zQ43 (W123&Q123) (1/2)

W1&Q1

SAR intensity images represent electromagnetic backscattering which is distinct from object appearance in vision domain. Their concentrated bright responses and speckled backgrounds can therefore be less intuitive to interpret visually. Related SAR-generation studies also illustrate these characteristics in real measurements, including SAR-DisentDM (AAAI 2026, Fig. 1(c)) and Φ-GAN (ICCV 2025, Fig. 5(h)). The MSTAR target chips used in these studies also have a size of 128×128 pixels. In Figures 4–5, the generated samples should therefore be assessed against the corresponding measured SAR images in terms of target support, orientation, and dominant scattering responses.

To assess whether the generated samples are useful despite their limited visual interpretability, we directly evaluate downstream utility through target recognition and azimuth estimation on held-out real images (Tables 5–6). For example, at 60° azimuth spacing, Φ-Flow augmentation improves recognition accuracy over Sparse-only from 37.35% to 61.06% on MSTAR and from 45.16% to 57.68% on ATRNet-STAR, while reducing azimuth-vector MSE in both cases. These results demonstrate useful information for SAR recognition and viewpoint estimation at the evaluated chip resolution.

W2&Q2

Thank you for raising the question of evaluation scope. We would first clarify that the ATRNet-STAR evaluation includes two bus categories. The ten-category ATRNet-STAR benchmark contains Dongfeng EQ6608LTV and Yutong ZK6120HY1, representing medium and large buses, alongside cars/SUVs, heavy trucks, and construction vehicles. Together with the ten MSTAR categories, these experiments cover diverse target sizes and structures (Appendix A.1).

To extend the evaluation beyond ground vehicles, we have added experiments on two aircraft categories, Kodiak 100 and AT-504. In these experiments, QSS is constructed and the generator is trained using sparse aircraft observations: six or three measured views per category at 60° or 120° azimuth spacing, respectively. As shown below, Φ-Flow augmentation improves both mean accuracy and mean azimuth MAE over Sparse-only in the two tested settings.

Azimuth spacingTraining conditionAccuracy (%) ↑Azimuth MAE (°) ↓
60°Sparse-only64.17 ± 8.8949.51 ± 4.49
60°Φ-Flow (QSS-only)77.58 ± 1.8729.08 ± 2.40
120°Sparse-only53.75 ± 6.8175.54 ± 1.94
120°Φ-Flow (QSS-only)65.75 ± 3.8340.15 ± 4.17

These results extend the evidence beyond ground vehicles under the evaluated sparse-view setting. Experimental settings, seed aggregation, and qualitative examples are provided in the new aircraft Appendix H (Table [20], lines [1890-1943]).

We have also updated the existing “Limitation and future work” paragraph to clarify this. Other target families, articulated or moving targets, and controlled cross-sensor, cross-frequency, or cross-polarization transfer remain future research directions.

W3&Q3

Our study focuses on novel-view synthesis from sparse observations of the evaluated target categories. The generalization studied here is from observed to unobserved viewing directions, rather than applying a trained generator directly to previously unseen instances or categories. We have not explicitly evaluated whether a trained generator can synthesize instance-specific views of previously unseen target instances or unseen categories from their sparse observations.

The key benefit is to obtain useful missing-view samples from very limited measurements. QSS makes the available coarse geometric information explicit, allowing a compact 3.87M-parameter generator to model the remaining view-dependent appearance. Our evaluation covers unobserved azimuths and entirely withheld target-depression planes, with recognition and viewpoint-estimation gains on held-out real images (Tables 5–6). These results demonstrate useful novel-view augmentation beyond reconstructing the observed training views. The additional aircraft experiments further extend the proposed principle beyond ground vehicles.

AuthorsOpenReview ↗

Response to reviewer zQ43 (Q4) (2/2)

Q4

Thank you for raising this practical concern. We conducted two additional experiments showing that physical vehicle dimensions and native 5° reference sampling are not essential for useful CST transfer. First, image-based registration replaces physical vehicle dimensions with pixel extents extracted from the target and reference CSS images. Second, we reduce the reference azimuth sampling from 5° to 15°.

The table reports five-seed mean azimuth MAE (°), with the change from QSS-only in parentheses; negative changes indicate improvement. Both tasks use 60° source-azimuth spacing.

ConditionAzimuth generationDepression interpolation
QSS-only35.05 (—)53.64 (—)
CST29.05 (−6.01)34.15 (−19.49)
CST (image registration)32.25 (−2.80)41.51 (−12.14)
CST (15° reference grid)31.92 (−3.13)39.67 (−13.97)

These results show that only using the image data for registration (CST w/ image registration) or relatively less reference measured targets (CST w/ 15° reference grid) can also improve the performance significantly compared with QSS-only, although their errors are higher than those of the original CST setting. It demonstrates that the registration metadata and reference-coverage requirements can each be relaxed while retaining viewpoint-estimation benefits. CST still assumes access to a suitable reference sequence. When no suitable reference is available, Φ-Flow can operate with QSS alone. Complete settings and results, including uncertainties, are provided in Appendix E.5 (Table [14], lines [1512-1532]).

We also added the discussions of compatible reference selection in the main text (Section 4.3.3, Table 4, Line 426).

I thank the authors for responding to my comments. I understood this took a lot of effort. I appreciate it. However, I remain concerned about the image quality of the results. The authors state that “[their method] improves recognition accuracy over Sparse-only from 37.35% to 61.06% on MSTAR and from 45.16% to 57.68% on ATRNet-STAR.” While these improvements are considerable, I do not consider a recognition accuracy of 57.68% satisfactory. Therefore, I am not comfortable raising my score.

Thank you for your continued consideration. The reported 57.68% corresponds to ten-class, cross-scene recognition with only six measured training views per category. Under this fixed observation budget, generated-view augmentation improves accuracy from 45.16% to 57.68%, demonstrating useful information beyond the sparse observations.

Prior SAR synthesis studies also report meaningful augmentation gains despite modest absolute accuracy. SAR-DisentDM (AAAI 2026, Table 4) improves average accuracy across six classifiers from 37.81% to 49.80% at 40° sampling intervals [1]. Φ-GAN (ICCV 2025, supplementary Table 6) reports 49.11% to 81.75% on 5% MSTAR [2]. Its 5% subset contains 121 measured images across ten categories—approximately 12 views per category, comparable in observation count to our 30° protocol, rather than the six-view 60° setting discussed here.

These studies use different evaluation protocols, but illustrate why absolute accuracy should be interpreted in the context of observation availability and test difficulty. Our downstream gains, together with improved target-support IoU and orientation accuracy, provide complementary evidence that the generated images preserve useful target information and improve missing-view coverage.

[1] SAR-DisentDM: A Semantic-Disentangled Diffusion Model for Limited-Data SAR Image Synthesis. AAAI 2026. DOI: 10.1609/aaai.v40i14.38165

[2] Φ-GAN: Physics-Inspired GAN for Generating SAR Images Under Limited Data. ICCV 2025. DOI: 10.1109/ICCV51701.2025.02700

Reviewer_ywofOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In sparse-view SAR target generation, a model generates vehicle images at unseen viewing angles from a few real views to help target recognition (ATR). Authors argue that coarse target shape changes predictably with viewing angle, whereas fine scattering does not ("geometry first, scattering when reliable"). They propose Φ\Phi-Flow, a 3.87M flow-matching U-Net with two modules: Query-Aligned Spatial Support (QSS, which interpolates class-mean maps across depression angles and rotates them to the target azimuth) and Compatible Scattering Transfer (CST, which transfers log-ratio scattering from a dense reference class). On MSTAR-10 (30-121 images) and 10-class ATRNet-STAR, Φ\Phi-Flow mostly beats 7 generative baselines and sparse controls in downstream accuracy (e.g., 61.06% vs 46.12% at MSTAR 60∘60^\circ; 74.33% vs 57.50% on ATRNet-STAR depression interpolation).

Critical strengths
  1. Clean QSS ablations (Tab. 1b, Tab. 9). Step-by-step ablations from unaligned CSS to aligned soft mask to full QSS cut orientation error (from 39.74∘39.74^\circ to 19.08∘19.08^\circ to 8.53∘8.53^\circ). Also, QSS beats wrong-class and wrong-azimuth controls across all 6 view spacings (+7.8 to +10.5 pts over wrong-class at 45∘45^\circ-120∘120^\circ).
  2. Large gains over no-generation controls (Tab. 5, Tab. 12). With only 3.87M parameters, Φ\Phi-Flow beats data-size-matched Repeated-sparse on MSTAR (+5.36 to +16.84 pts) and Sparse-only on ATRNet-STAR (+12.52 pts at 60∘60^\circ; +18.83 pts on withheld 30∘30^\circ depression).
  3. Strict data-split hygiene and honest negative results. Held-out 30∘30^\circ depression views are excluded from QSS, CST, and flow training. Authors also candidly report where CST hurts accuracy (30∘,60∘30^\circ, 60^\circ in the additional set).
Critical weaknesses

Φ\Phi-Flow clearly beats non-generative controls (Sparse-only and Repeated-sparse, this paper's own data controls). However, experiments do not support three claims (paraphrased): "model scaling cannot recover QSS geometry", "gains over generative baselines come from the method rather than baseline under-tuning", or "CST works without oracle target views".

W1. Generative baselines lack training details, perform worse than no generation, and miss simple rotation controls (Tab. 4, 5, 12).

  1. Baseline training details are missing. Sec. 4.4 points to Appendix A for baseline setups, but Appendices A-B only describe Φ\Phi-Flow. For the 7 generative baselines (12.45M-1,286M parameters trained on just 30-121 images), paper gives no training steps, learning rates, angle-conditioning details, or code sources.
  2. Baselines hurt accuracy below using no synthetic data. On ATRNet-STAR, all 7 generative baselines score lower than Sparse-only (no synthetic data) at 60∘60^\circ, 90∘90^\circ, and 120∘120^\circ (Tab. 5). At 120∘120^\circ, ACGAN gets 9.84% accuracy and its azimuth predictions are no better than constant zeros (MSE 0.500±0.0000.500 \pm 0.000). On MSTAR, 15 of 28 baseline entries fall below Repeated-sparse, including all 7 baselines at 120∘120^\circ (Tab. 12).
  3. Simple rotation and FM+QSS controls are missing. QSS is simply a rotated class-mean template (Eq. 2-4). Yet paper never tests two simple controls: (a) augmenting the classifier directly with rotated real images (or using the rotated QSS template itself as the synthetic image), and (b) feeding the QSS channel into the 237M FM baseline (FM+QSS), which already matches Φ\Phi-Flow at 30∘30^\circ (80.75% vs 80.23%, Tab. 4).

W2. Claim that model scaling cannot learn geometry relies on a broken no-QSS baseline (Tab. 1a, 1b, 4, Fig. 7). In the width sweep (0.52M-13.9M parameters, Tab. 1a), the no-QSS U-Nets fail to train properly (support IoU 0.011, orientation error 39.11∘39.11^\circ). Adding their synthetic images drops 30∘30^\circ accuracy to ~55% (Fig. 7e), which is much worse than training on real Sparse-only data alone (67.84%, Tab. 12). By contrast, paper's own 237M conditional FM baseline (also w/o QSS) gets much better FID at 30∘30^\circ (105.87 vs ~380-420 for no-QSS U-Nets, Fig. 7a) and reaches 80.75% accuracy, matching Φ\Phi-Flow. This suggests a properly trained model without QSS can learn target geometry at 30∘30^\circ, so the no-QSS U-Net in Tab. 1 is an overly weak baseline.

W3. CST "when reliable" screening uses dense oracle target views and lacks a negative control (Tab. 2, 10). To pick "compatible" reference classes for CST, authors compute similarity using the full measured target sequence (including unseen test azimuths). In a real sparse-view setting, these dense target views do not exist, so CST relies on an oracle screening step. Also, paper never tests CST with an incompatible, random, or 90∘90^\circ-rotated reference class to prove that compatibility screening actually matters. On the additional pairs at 30∘30^\circ (Tab. 10), CST's azimuth gain is within noise (38.64±3.9638.64 \pm 3.96 vs 37.41±2.1937.41 \pm 2.19) while classification accuracy drops (from 83.63% to 80.66%).

W4. Inferability-utility trade-off is based on 6 noisy points and is partly built in (Fig. 3b). Fig. 3b plots only 6 single-run points, and the rank correlation between inferability (SrS_r) and utility (UrU_r) is not statistically significant (ρ=−0.37\rho = -0.37). Also, the trend only appears inside the nested Top-32/64/128 group, and UrU_r is defined as gain over QSS-only generation, so it excludes everything QSS already captures by definition.

Questions for authors
  1. Baseline setups & simple controls (W1). How were the 7 baselines angle-conditioned and tuned on 30-121 images? On MSTAR 60∘/120∘60^\circ/120^\circ and ATRNet-STAR 60∘60^\circ, how does Φ\Phi-Flow compare against (a) FM+QSS (237M FM with the QSS channel) and (b) augmenting the classifier with rotated real images (or rotated QSS maps directly)?
  2. No-QSS failure vs FM-237M (W2). Why does the no-QSS U-Net drop 30∘30^\circ accuracy below Sparse-only (~55% vs 67.84%), whereas 237M FM (also w/o QSS) reaches 80.75%? What are the support IoU and orientation error of 237M FM?
  3. CST without oracle views (W3). Can compatible reference classes be selected using only the sparse training views? Also, what happens if you run CST with an incompatible or random reference class?
6 comments
AuthorsOpenReview ↗

Response to reviewer ywof (W1.1&W1.2&Q1) (1/4)

W1.1&Q1

Thank you for identifying this omission. We have added Appendix B.4, “Generative-baseline implementations and training,” documenting the seven baselines’ optimization settings, angle-conditioning mechanisms, implementation sources, and pretrained components. The main settings are summarized below; complete details are provided in Appendix B.4 (Table [10],lines [1026-1041]).

Hyperparameter choices were based primarily on the corresponding papers and implementations. Method-specific adjustments were guided by training-loss convergence and generated-image quality at observed training views, following the same tuning criteria as Φ-Flow. Checkpoint selection follows the same fixed-budget principle: the diffusion baselines use their final checkpoints, and Φ-StyleGAN uses its final EMA generator. The red-framed training-view examples in Figs. 14–16 also show that several baselines reproduce the observed target appearances well. The distinction between observed-view fitting and unseen-view generation is discussed further in our response to W1.2.

We also corrected a naming error: “StyleGAN” should read “Φ-StyleGAN,” the StyleGAN variant reported in the Φ-GAN paper (Zhang et al. (ICCV 2025)).

W1.2&Q1

Correction of the 120° GAN results: We thank the reviewer for drawing attention to the anomalous ACGAN result. Three 120° comparisons in Tables [5, 6, 17, 18] were affected by incomplete-batch handling: under the original loader configuration, 30 training images with batch size 32 yielded no training batches. We corrected the loader to retain incomplete batches and reran the affected comparisons.

For MSTAR ACGAN, accuracy changed from 18.62 ± 10.13% to 29.58 ± 1.38%, and azimuth MSE from 0.5229 ± 0.0309 to 0.7432 ± 0.0341. For ATRNet-STAR ACGAN, accuracy changed from 9.84 ± 2.60% to 31.81 ± 3.18%, and MSE from 0.5000 ± 0.0001 to 0.7400 ± 0.0188. For ATRNet-STAR Φ-GAN, accuracy changed from 15.47 ± 13.73% to 34.11 ± 0.92%, and MSE from 0.5144 ± 0.0321 to 0.7140 ± 0.0188. The corresponding entries have been updated. These corrections leave unchanged the conclusion that synthetic augmentation would underperform the corresponding no-generation controls under severe view sparsity.

Explanation of lower downstream accuracy than Sparse-only:

The downstream benefit of generated-data augmentation depends on whether it adds useful, sufficiently accurate novel-view examples beyond the sparse observations. The red-framed examples in Figs. 14–16 show that most compared methods can achieve good generation results at observed views. With denser sampling, such as MSTAR at δθ=30∘\delta \theta =30^\circ, most compared methods are capable to generate target with relatively decent orientations and scattering details. As angular gaps widen, however, some compared methods exhibit incorrect orientations, distorted or blurred structures, or reproduce an observed appearance instead of the requested view. These examples are consistent with the hypothesis that view-inconsistent synthetic samples can add misleading supervision and reduce the benefit of augmentation. By contrast, Φ-Flow provides query-aligned spatial support and yields lower image-orientation error and better downstream recognition in the reported wide-gap settings.

Effect of query-aligned geometry: The matched FM+QSS versus no-QSS comparison discussed in our response to W2 provides further support: on the same FM backbone, QSS improves both query-view geometry and downstream recognition under larger angular gaps, reinforcing the importance of useful novel-view samples for augmentation (Appendix C.1).

AuthorsOpenReview ↗

Response to reviewer ywof (W1.3&Q1) (2/4)

W1.3&Q1

Thank you for this suggestion. We added "augmenting the classifier with rotated real images or QSS maps" experiments, as well as "FM+QSS" experiments. Within each protocol, these conditions use the same sparse real observations, augmentation counts, downstream settings, and real test sets as Φ-Flow.

Augmenting the classifier with Rotated real or QSS maps:

Recognition accuracy (%) ↑

DatasetSpacingRotated realQSS templateΦ-Flow
MSTAR30°70.11 ± 2.0961.17 ± 2.1180.23 ± 1.06
MSTAR60°46.37 ± 1.2739.58 ± 2.3061.06 ± 1.15
MSTAR120°35.96 ± 3.3335.26 ± 4.5045.25 ± 2.40
ATRNet-STAR60°44.20 ± 4.9741.02 ± 3.5357.68 ± 2.27
ATRNet-STAR90°41.71 ± 1.3547.93 ± 2.2649.86 ± 3.26
ATRNet-STAR120°47.63 ± 3.4631.72 ± 5.1750.16 ± 3.20

Azimuth-vector MSE ↓

DatasetSpacingRotated realQSS templateΦ-Flow
MSTAR30°0.2994 ± 0.01710.2614 ± 0.01110.1745 ± 0.0205
MSTAR60°0.5301 ± 0.02290.3790 ± 0.02220.3349 ± 0.0167
MSTAR120°0.7704 ± 0.02800.6313 ± 0.03050.4794 ± 0.0286
ATRNet-STAR60°0.7795 ± 0.03740.5987 ± 0.02320.5238 ± 0.0275
ATRNet-STAR90°0.7492 ± 0.05950.5689 ± 0.03820.5087 ± 0.0206
ATRNet-STAR120°0.8714 ± 0.03480.6111 ± 0.01840.6036 ± 0.0346

Values are mean ± sample standard deviation over five downstream seeds.

Φ-Flow achieves higher mean recognition accuracy than both controls in all six listed protocols. Its lower azimuth-vector MSE than the QSS-template control further supports the value of conditional generation beyond using the query-aligned template itself as an image. The comparison in Fig. 11 (Appendix C.3, lines [1287-1307]) illustrates the difference: the simple rotation of a SAR image (Rotated real) or QSS template cannot simulate the real imaging geometry change (Measured query) due to the special electromegnatic behaviours of SAR. It can explain the inferior classifier performance of simple rotation augmentation compared with the proposed Φ-Flow generation.

FM+QSS versus Φ-Flow:

For FM+QSS, we added the requested paired comparison using the same ADM-style FM backbone, matching initialization, optimization, training budget, and generation noise. The larger backbone already performs well at 30° spacing but fails at larger sparsity; adding QSS yields clearer gains at the wider 60° and 120° gaps. Complete settings and results are reported in Appendix C.1 (lines [1104-1226]).

We also compare Φ-Flow with the requested QSS-conditioned ADM-style FM backbone. On MSTAR at 30° spacing, FM+QSS achieves higher mean recognition accuracy than Φ-Flow (82.27% vs. 80.23%) and lower mean azimuth-vector MSE (0.1506 vs. 0.1745). At wider 60° and 120° spacing, Φ-Flow obtains higher mean recognition accuracy (61.06% vs. 57.43%, and 45.25% vs. 41.91%, respectively). For azimuth-vector MSE, FM+QSS is lower at 60° (0.3184 vs. 0.3349), while Φ-Flow is lower at 120° (0.4794 vs. 0.4943). Thus, the two models show different trade-offs across sparsity levels and metrics; Φ-Flow retains higher mean recognition accuracy at the two wider MSTAR spacings.

MSTAR spacingFM Acc. (%) ↑FM+QSS Acc. (%) ↑Φ-Flow Acc. (%) ↑FM azimuth MSE ↓FM+QSS azimuth MSE ↓Φ-Flow azimuth MSE ↓
30°82.39 ± 3.1282.27 ± 1.2680.23 ± 1.060.1645 ± 0.01400.1506 ± 0.01680.1745 ± 0.0205
60°42.70 ± 0.9757.43 ± 1.7361.06 ± 1.150.5254 ± 0.02860.3184 ± 0.02200.3349 ± 0.0167
120°31.06 ± 1.6541.91 ± 1.7245.25 ± 2.400.6885 ± 0.04160.4943 ± 0.01790.4794 ± 0.0286
AuthorsOpenReview ↗

Response to reviewer ywof (W2&Q2) (3/4)

W2&Q2

Thank you for raising this concern. We have added matched experiments with and without QSS on the 237.26M-parameter ADM-style FM backbone across all six MSTAR protocols (Appendix C.1, Figs. 8–9, lines [1104-1226]).

Q2.1: Why does the compact no-QSS model perform worse than the benchmark FM?

The approximately 55% and 80.75% accuracies come from different architectures and training configurations, not two widths of the same network, so their performance gap cannot be attributed to parameter count alone. The compact ablation uses a three-resolution convolutional FiLM U-Net, whereas the benchmark FM uses a five-resolution ADM-style U-Net with base width 256, GroupNorm residual blocks, and self-attention (Appendix B.3). Geometrically inaccurate synthetic samples can introduce misleading supervision, helping explain why augmentation with the compact no-QSS model can perform worse than Sparse-only.

Q2.2: What are the support IoU and orientation error of the original benchmark FM?

The results for the original benchmark generations are:

MSTAR spacingFM support IoU ↑Φ-Flow support IoU ↑FM orientation error ↓Φ-Flow orientation error ↓
30°0.4680.4907.22°6.11°
45°0.4330.4678.56°7.06°
60°0.3400.42521.22°9.34°
120°0.2480.36831.53°9.35°

These metrics use all 2,747 generation queries, including observed views, consistently with the capacity analysis. They confirm good geometric recovery at 30°, but substantially larger errors at 60° and 120°.

Other concern: Does the capacity-related conclusion rely only on the compact no-QSS model?

No. We additionally tested QSS on the 237.26M-parameter ADM-style FM backbone. Across six MSTAR protocols, QSS improves spatial support and orientation metrics, with larger recognition gains at wider angular gaps (Fig. 8). Thus, QSS remains useful on the tested large backbone, although these results do not rule out further gains from model scaling.

MSTAR spacingFM accuracy (%) ↑FM+QSS accuracy (%) ↑FM azimuth MSE ↓FM+QSS azimuth MSE ↓
30°82.39 ± 3.1282.27 ± 1.260.1645 ± 0.01400.1506 ± 0.0168
45°66.56 ± 4.0272.22 ± 1.170.2959 ± 0.02200.2434 ± 0.0099
60°42.70 ± 0.9757.43 ± 1.730.5254 ± 0.02860.3184 ± 0.0220
72°43.97 ± 2.4459.07 ± 2.220.7125 ± 0.02250.3510 ± 0.0276
90°21.53 ± 0.9627.84 ± 4.420.5928 ± 0.01070.4108 ± 0.0370
120°31.06 ± 1.6541.91 ± 1.720.6885 ± 0.04160.4943 ± 0.0179

Values are mean ± standard deviation over five downstream seeds; azimuth MSE uses the sine/cosine representation. These matched reruns are reported separately from the original benchmark; the available no-QSS reruns differ from the original accuracies by at most 2.09 percentage points.(The detailed explanation can be found in Appendix B.3, lines [992-1003])

Comments on the paper

1 comment
AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewers and Area Chair,

Thank you for taking the time to review our submission. As the author discussion period draws to a close, we wanted to draw your attention to our rebuttal posted in this forum. We would appreciate your considering them in the evaluation of our submission. If any concerns remain or any part of our responses would benefit from clarification, we would be happy to follow up.

Thank you for your time and consideration.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.