- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies post-reconstruction errors in INR-based reconstruction and proposes SDR, a two-stage self-diagnostic refinement framework. The key observation is that residual errors are spatially concentrated, can be partially localized using K-fold O-O-F uncertainty , and exhibit different behaviors under continued optimization. SDR first uses O-O-F held-out errors to gate further optimization, then transfers structurally matched signed O-O-F residuals to high-uncertainty regions through KNN-based residual correction with strength .
Critical strengths
-
Post-reconstruction error analysis is a meaningful problem for INR-based reconstruction. I particularly appreciate the post-hoc and plug-and-play formulation: SDR does not require modifying the architecture or retraining the base reconstruction method with additional supervision, and it does not require ground truth at the genuinely missing entries.
-
The experimental evaluation is relatively thorough within the considered reconstruction setting. The paper evaluates multiple reconstruction methods, signal modalities, sampling ratios, and five random seeds, and reports standard deviations. The improvements are generally consistent, and the residual-transfer control experiments provide useful evidence that both the sign and structural correspondence of the transferred residuals matter.
Critical weaknesses
-
Although I believe post-reconstruction error refinement for INR is a meaningful research topic, I am not fully convinced by the current experimental setting. For natural RGB images, there already exist strong reconstruction methods based on pretrained generative priors, such as DPS, DDNM, and DiffPIR, where INR-based reconstruction may not be the most competitive choice. I think SDR would be more convincing if evaluated more extensively on settings where INR or continuous representations are naturally useful, such as MSI/HSI, scientific data, or NeRF. This is especially relevant because the key ingredients of SDR, including O-O-F residuals and cross-validation uncertainty, are not inherently specific to INR.
-
My impression is that the most fundamental contribution of this work may actually be the use of O-O-F residuals as observable proxies for unknown residuals, followed by residual/bias correction. In comparison, the additional diagnosis, optimization gating, masking, and the “high uncertainty + still moving” versus “high uncertainty + hardly moving” categorization feel relatively heuristic. In particular, a prediction that is still moving may simply indicate incomplete convergence or oscillation, while a stable prediction does not necessarily define a fundamentally different type of reconstruction error. The relatively large number of operations in the full pipeline also makes me wonder whether these additional components are necessary, or whether a simpler formulation centered on O-O-F residual transfer could capture most of the improvement.
Questions for authors
-
How essential are the optimization gate and the high-CV mask once reliable O-O-F residuals are available? I would be interested in an ablation that removes the gate and/or the mask and studies residual transfer more directly. Relatedly, the theoretical bound identifies the fold-to-full discrepancy as an important source of estimation error. Could the authors empirically measure this discrepancy and study when O-O-F residuals cease to be representative of the residuals of the full reconstruction? I think such an analysis would clarify the central mechanism of the method more directly than the current optimization-dynamics categorization.
-
Are the base reconstruction methods fully optimized before SDR is applied? The paper refers to the starting point as a converged reconstruction, while the experimental description states that the base methods are run for their original pre-set number of iterations. Since continued optimization alone sometimes yields substantial PSNR improvements, it would be useful to compare against a stronger “train-to-convergence” baseline. I would also encourage comparisons against simple alternatives such as per-channel/band residual-bias correction and an ensemble of independently trained reconstructions. Finally, evaluating the same O-O-F refinement principle on a qualitatively different continuous-representation problem, such as NeRF or a scientific/medical inverse problem such as CT reconstruction, would provide stronger evidence that the proposed mechanism is genuinely general rather than specific to the current tensor-completion setting.
Updated Paper and Additional Experiments
Thank you for your constructive feedback. We have uploaded a revised paper with additional experiments and clarifications addressing the reviewers’ concerns: