acceptodds
Under review as a conference paper at ICLR 2027

Directional Positive-Definite Metric Consensus for Time-Modulated Hyperspectral Reconstruction

Abstract

Time-modulated hyperspectral imaging records multiple spatially aligned measurements of the same scene under different spectral responses for reconstruction. However, correlations among these responses lead to nonuniform sensitivity across spectral directions, while existing reconstruction algorithms generally coordinate learned priors using shared scalar penalties. Here, we propose the Time-Modulated Calibration-Conditioned Spatial–Spectral Reconstruction Network (TCSS), an ADMM-inspired unfolding network built on Directional Positive-Definite Metric Consensus. Using calibration-derived directions in spectral space, TCSS constructs branch-specific positive-definite metrics that jointly control prior strength and spectral–spatial allocation across directions. Spectral-Shape-Guided Dual-Branch Reconstruction (SDBR) and Cross-Stage Reconstruction Increment Propagation (CSIP) further improve spatial–spectral prior extraction and information propagation across reconstruction stages. Extensive experiments show that TCSS achieves superior reconstruction performance with a compact architecture and generalizes consistently across datasets and the tested calibrated response sets. The code is released at https://anonymous.4open.science/r/test24/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
time-modulated hyperspectral imaging, spectral compressive imaging, directional positive-definite metric consensus,ADMM unfolding

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_w9NXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates hyperspectral imaging, Rather then employing shared scalar penalties to coordinate learned priors, the authors introduce directional positive-definite metric consensus, which enables direction-specific spectral–spatial prior allocation and admits a closed-form image update.

Critical strengths
  1. This is an interesting task, different from traditional Cassi. The paper identifies specific challenges in hyperspectral imaging and highlights the limitations of existing methods.

  2. The proposed positive-definite-metric-based framework for time-modulated hyperspectral imaging is conceptually novel and offers good interpretability.

Critical weaknesses
  1. Insufficient Validation of Functional Complementarity. The functional complementarity among the components within the TCSS module has not been adequately validated. The ablation study in Table 2 only demonstrates the cumulative performance gains obtained by progressively introducing additional components, but lacks a thorough analysis of their individual contributions and potential functional redundancy.

  2. The effectiveness of the proposed directional metric design is not sufficiently supported by the experimental results. The paper emphasizes calibration-derived directional positive-definite metrics as its central contribution. However, Table A.8 shows that Generated-V performs worse than both Generated-Scalar and Shuffled-V in terms of mean SAM and MPSNR on unseen response sets. Particularly, Shuffled-V achieves better reconstruction performance despite disrupting the correspondence between singular-value features and spectral directions. These results raise questions about whether the performance improvements primarily originate from calibration-conditioned metric generation rather than the proposed direction-specific prior coordination. Further controlled experiments and direction-wise reconstruction error analysis would help establish the necessity and effectiveness of the directional metric design.

Questions for authors
  1. In Table 3(a) and Table A.8, Generated-V does not consistently outperform Generated-Scalar. Could the authors clarify what specific advantages the proposed direction-dependent metric provides over the simpler calibration-conditioned scalar metric?

  2. Validation of the proposed mechanism: The paper argues that directional metrics address nonuniform measurement sensitivity. Could the authors provide further evidence, such as reconstruction errors along different singular-vector directions, to demonstrate that the proposed method specifically improves reconstruction in low-sensitivity spectral directions?

  3. Why is the font so small in the figures?? I can't see what you want to convey at all. This is not a qualified paper. Also, why aren't the network and visualization results using 1\linewidth? The difference is not visible at all. Also, please don't adjust the paper spacing and layout format arbitrarily.

  4. Why is the unfolding comparison only presented in 5 stages, and what about the results of 7 and 9 stages? Such a comparison is not convincing, and it also lacks some SOTA method such as Mijun, Lade-DUN, etc

Reviewer_9PRCOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the non-uniform directional sensitivity across the spectral domain caused by correlated broadband responses in temporally modulated multi-frame hyperspectral reconstruction. The authors propose TCSS, an ADMM-inspired unfolding network centered on directional positive-definite metric consensus. Specifically, the method leverages the right singular vectors of the calibrated response matrix to construct positive-definite metrics for two branches, separately regulating the overall prior strength along each spectral direction and the spectral--spatial allocation, while integrating spectral-shape-guided dual-branch reconstruction and cross-stage incremental propagation. The paper reports reconstruction results on the KAIST, Harvard, and ICVL datasets, alongside experiments on cross-response generalization, varying numbers of measurements, and calibration error perturbations. While the reconstruction performance and parameter compactness of the full network are well-supported, the primary concern is that the current controlled ablation studies have not yet established the independent gains attributable to the directionality itself and the correspondence to the correct calibration.

Critical strengths
  1. The problem motivation is concrete and well-founded. Equations (2)–(3) on page 4 and Table A.2 on page 16 explicitly link response correlations to weakly observable spectral directions. With 40 measurements for 28 spectral bands, although the forward operator has full column rank, its condition number remains approximately 1.13×1071.13 \times 10^7, demonstrating that a sufficient number of measurements does not guarantee reliable observability across all spectral directions.

  2. The core computational mechanism is clear and mathematically sound. Equations (4)–(9) on pages 4–5 and Appendix B.2 formulate the positive-definite consensus update under a shared singular-vector basis. Because the metrics of both branches are positive-definite, the quadratic image subproblem admits a unique solution that can be computed direction-by-direction.

  3. The reconstruction performance and parameter compactness of the full network are well-supported. In Table 1, TCSS-5stg achieves an MPSNR of 47.305 dB and a SAM of 3.186∘3.186^\circ with only approximately 0.06 M parameters. Across three training seeds in Table A.9, TCSS attains 47.416±0.11247.416 \pm 0.112 dB and 3.046±0.148∘3.046 \pm 0.148^\circ, outperforming DPU-SCI (45.934±0.19945.934 \pm 0.199 dB and 7.067±0.202∘7.067 \pm 0.202^\circ). Cross-dataset evaluations further corroborate its effectiveness under the tested protocols.

  4. The experimental reporting includes valuable controlled and unfavorable results. The authors provide ablations on generative scalars and shuffled directional correspondence, comparisons under recommended learning rates, and evaluations under fewer measurements, registration errors, and calibration mismatches. Moreover, the per-scene averages in Table A.5 are fully consistent with Table 1, with no substantive arithmetic errors found.

Critical weaknesses
  1. Table 2 introduces generative parameterization and directionality simultaneously, making it impossible to disentangle their respective contributions. In Table 3(a) and Table A.8, Generated-V does not exhibit a clear advantage over Generated-Scalar or Shuffled-V. While Table 3(b) supports the role of jointly modulating τ\tau and π\pi within the directional configuration, it does not resolve the aforementioned attribution ambiguity. We recommend conducting multi-seed controlled comparisons within the full SGA+CSIP architecture and aligning the claims regarding core contributions with the empirical findings. This is the primary issue affecting the assessment of novelty.

  2. Table A.4 shows that a physically equivalent row permutation alone causes substantial degradation in the current adapters for MST++, SPECAT, and S2S^2-Transformer. While the appendix acknowledges this limitation, the main text should reflect this as well and prioritize permutation-stable baselines such as DPU-SCI and PADUT when interpreting transfer advantages. Furthermore, the baseline improvements obtained under the recommended learning rates in Table A.18 should also be presented in the main text. This affects the interpretation of both the magnitude of the performance gains and the scope of the comparisons.

  3. The manuscript does not specify a traceable source for the response curve library, its amplitude normalization/processing, exact wavelengths, or the indices of the selected response subsets; the complete list of test scenes and preprocessing protocols for Harvard and ICVL also need to be provided. In addition, the conditioning encoder and query/key dimensions in CSA, as well as the metric merging and image update rules in the spatially modulated extension, require concrete definitions. These omissions hinder independent reproducibility, and whether the accompanying code resolves them needs verification.

  4. Figure A.4 does not fully define the directional projection error, the magnitude calculation for each dynamic quantity, or the normalization denominators, which limits the quantitative interpretation of the plotted curves. We suggest supplementing the exact mathematical formulas, numerical scales, and grouping criteria for strong versus weak directions. This issue pertains to under-specified diagnostic definitions rather than confirmed numerical errors based on current evidence.

Questions for authors
  1. Within the complete architecture retaining SGA, CSIP, the loss function, and the training budget, what are the seed-paired results for Generated-Scalar, Generated-V, and Shuffled-V? Does preserving the exact correspondence between singular-value features and spectral directions further reduce reconstruction errors along weakly observable directions? Given that the appendix has verified the numerical equivalence between Free-V and Generated-V under matched weights, and Table A.10 shows that different parameterizations learn distinct stage-wise penalty schedules, how do the authors disentangle and interpret the respective roles of reparameterization, optimization dynamics, and calibration conditioning in the gains achieved by the generator?

  2. For the order-sensitive adapters in Table A.4, could an explicit and consistent train/test ordering convention—or other order-control mechanisms—be applied to evaluate how much of the cross-response performance gap remains (with measurements and response matrices synchronously permuted during control)? Furthermore, how do DPU-SCI and DAUHST trained with their recommended learning rates in Table A.18 perform under the exact same frozen-weight cross-response evaluation protocol?

  3. Could the authors provide a traceable source or data file for the measured response curves, details of the amplitude processing, the ordered indices for each response subset, and the exact wavelength vectors, along with the complete list of test scenes, resampling procedures, and intensity preprocessing configurations for Harvard and ICVL? In the spatially modulated extension, what specific window-metric merging formula and actual image update rule are employed?

  4. Could the authors provide the explicit mathematical expressions for the directional projection error, spectral/spatial consensus residuals, and dual-variable corrections in Figure A.4, including their within-scene aggregation schemes and numerical scales? Are the vertical axes scaled, what normalization denominator is used for the reconstruction increments, and which specific directions are included in the strongest and weakest 20% subsets?

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

This paper proposes a directional positive-definite metric consensus framework for time-modulated multi-frame hyperspectral reconstruction. The method exploits the SVD of the spectral response matrix to construct adaptive direction-dependent metrics within an unrolled optimization framework. Experiments demonstrate promising reconstruction performance compared with existing methods.

Critical strengths
  1. The paper addresses a meaningful problem in hyperspectral reconstruction and proposes a physically motivated, mathematically interpretable optimization framework.
  2. The proposed method achieves competitive reconstruction performance. The extensive supplementary experiments further support its empirical effectiveness.
Critical weaknesses
  1. Insufficient justification of the core directional design. In Table 3(a), Shuffled-V achieves better SAM than Generated-V, while Generated-Scalar performs comparably. Similar observations appear in Table A.8. These results question whether accurate alignment with singular directions is essential to the proposed method.
  2. Concerns about baseline fairness. Tables A.17–A.18 show substantial improvements in baseline performance after adjusting training schedules, suggesting that the performance advantages reported in the main comparison may be overstated.
  3. Limited real-world validation. The experiments primarily use simulated measurements with measured spectral responses. The method also shows considerable sensitivity to calibration errors (Table A.16), while real hardware acquisition and inter-frame misalignment are insufficiently evaluated.
Questions for authors
  1. Why does Shuffled-V perform comparably to or better than Generated-V in Tables 3(a) and A.8? Can the authors clarify the specific contribution of accurate singular-direction alignment?
  2. How much of the reported performance advantage remains when using the improved baseline training configurations in Tables A.17–A.18?
  3. Given the sensitivity to calibration errors reported in Table A.16, how robust is the method under realistic calibration uncertainty and inter-frame misalignment?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.