- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes a Trajectory-Aware Differentiable Blur Augmentation framework for improving image deblurring, particularly under limited training data. The central motivation is that existing blur augmentation methods mainly manipulate motion magnitude and orientation without explicitly modeling the temporal dynamics and geometric structure of motion trajectories. The proposed framework contains two components. First, Global Trajectory-aware Augmentation (GTA) models camera motion using velocity, acceleration, curve and S-curve primitives, and in-plane rotation. Second, Local Trajectory-aware Augmentation (LTA) represents spatially varying residual motion through learned local trajectory components and spatial mixing weights. A differentiable blur renderer synthesizes augmented images, while a learnable Augmentor is optimized using the deblurring objective to generate challenging trajectory perturbations.
Critical strengths
The paper presents a well-motivated and technically reasonable trajectory-aware blur augmentation framework that addresses the limitations of conventional magnitude- and orientation-based augmentation methods. By combining Global Trajectory-aware Augmentation (GTA) and Local Trajectory-aware Augmentation (LTA), the proposed approach effectively models diverse temporal motion dynamics and spatially varying blur patterns. The differentiable augmentation strategy further enables trajectory perturbations to adapt to the deblurring objective, providing a meaningful contribution beyond predefined random augmentation. Experimental results demonstrate consistent improvements across multiple datasets, deblurring architectures, and training-data regimes, with particularly encouraging gains under limited-data settings. The component ablations and qualitative visualizations further support the effectiveness of the proposed framework, while the clear presentation and detailed implementation settings enhance the overall quality of the paper.
Critical weaknesses
Although the proposed method achieves consistent performance improvements across multiple experiments, several aspects could be further strengthened. First, the trajectory diversity analysis relies on the authors' own trained trajectory estimation network, which may introduce evaluation bias and does not sufficiently validate the physical realism of the generated motion trajectories. Second, the additional performance gains brought by Local Trajectory-aware Augmentation (LTA) and the differentiable augmentation strategy are relatively modest, and their individual contributions require further investigation. Moreover, the framework requires pretraining a reblurring network and adopting an alternating optimization strategy, which increases training complexity, yet the associated computational overhead is not quantitatively evaluated. Finally, the comparison with GeoSyn mainly relies on previously published results rather than a fully unified experimental protocol, which somewhat limits the conclusiveness of the reported performance advantages.
Questions for authors
- Can the authors provide an independent evaluation to validate the diversity and physical realism of the generated motion trajectories?
- Can the authors further clarify the individual contributions of Local Trajectory-aware Augmentation (LTA) and the differentiable augmentation strategy to the overall performance improvements?
- Can the authors provide a quantitative analysis of the additional training time, computational cost, and GPU memory overhead introduced by the proposed framework?
- Can the authors further clarify the fairness and consistency of the experimental comparison with GeoSyn?
- Can the authors further evaluate the generalization ability of the proposed method under complex real-world degradations, such as rolling-shutter effects and sensor noise?
4 comments
[A3] Computational Overhead
We appreciate the reviewer's concern regarding the additional computational cost introduced by reblurring network pretraining and alternating optimization. To clarify this overhead, we compare estimated training time and measured peak GPU memory usage for the baseline, ID-Blau, and our method on GoPro using MIMO-UNet+ as the deblurring backbone. The results are reported in Table 8 of the revised manuscript (reproduced below).
For this cost comparison, we follow the original MIMO-UNet+ training protocol for all three methods. We separately report the estimated GPU-hours for reblurring pretraining, deblurring pretraining, and main training. Training costs are estimated from step times measured on a single NVIDIA RTX 3090 with BF16 mixed precision. ID-Blau additionally uses 5,000 epochs of diffusion training and 500 epochs of deblurring pretraining on a generated dataset five times the size of the original training set. The reported total includes both pretraining stages and excludes offline dataset generation for ID-Blau. Our method uses 1,000 epochs of reblurring pretraining, and its main-training time includes online Augmentor optimization. As shown in Table 8, our method achieves a comparable estimated total training time to ID-Blau, although it requires higher peak GPU memory due to batch concatenation.
| Aug. | Estimated training time (GPU-hours)↓ | Peak memory (GB)↓ | |||
|---|---|---|---|---|---|
| Reblur pretraining | Deblur pretraining | Main training | Total | ||
| – | – | – | 66.96 | 66.96 | 6.31 |
| ID-Blau | 13.91 | 55.90 | 66.96 | 136.77 | 6.31 |
| Ours | 9.59 | – | 149.70 | 159.29 | 13.03 |
[A4] Fairness of the Comparison with GeoSyn
We appreciate the reviewer's concern regarding the fairness of our comparison with GeoSyn. As clarified in Appendix A.2, GeoSyn's official code and pretrained blur synthesis model are not publicly available, preventing us from reproducing its results under consistent experimental settings or conducting additional comparisons across different backbones and datasets. We therefore report the RSBlur result from the original GeoSyn paper as a reference, rather than as a direct comparison under identical experimental settings.
Nevertheless, GeoSyn primarily modifies the magnitude and direction of estimated motion, whereas our method explicitly augments trajectory geometry and temporal motion dynamics. We will further clarify the scope and limitations of this comparison in the revised manuscript.
[A5] Generalization to Real-World Degradations
We appreciate the reviewer's concern regarding generalization to complex real-world degradations, such as rolling-shutter effects and sensor noise. These effects depend on camera-specific imaging processes and are difficult to explicitly model within our trajectory-based augmentation framework. Nevertheless, we can evaluate the generalization ability of our method on real-world datasets that reflect practical camera acquisition conditions. To this end, we consider both our existing RSBlur evaluation and additional cross-domain experiments on RSBlur and RealBlur-J/R.
As shown in Table 2 of the revised manuscript (reproduced below), our method achieves higher PSNR on RSBlur than the unaugmented baseline and ID-Blau under our training protocol, and also exceeds the result reported in the original GeoSyn paper. Furthermore, Table 5 of the revised manuscript (reproduced below) shows that, when trained on GoPro, our method generally achieves better performance than the unaugmented baseline and ID-Blau on RSBlur, RealBlur-J, and RealBlur-R.
| Methods | GMACs↓ | PSNR↑ |
|---|---|---|
| NAFNet | 63.64 | 33.97 |
| + ID-Blau | 63.64 | 34.01 |
| + GeoSyn | 63.64 | 34.23 |
| + Ours | 63.64 | 34.34 |
| Test Set | Aug. | PSNR↑ | LPIPS↓ |
|---|---|---|---|
| GoPro (Train) → Real-world Blur (Test) | |||
| RSBlur | – | 27.36 | 0.334 |
| ID-Blau | 26.34 | 0.386 | |
| Ours | 27.84 | 0.310 | |
| RealBlur-J | – | 26.54 | 0.174 |
| ID-Blau | 26.35 | 0.190 | |
| Ours | 26.55 | 0.165 | |
| RealBlur-R | – | 33.65 | 0.097 |
| ID-Blau | 33.76 | 0.095 | |
| Ours | 33.77 | 0.080 |
[1] Exposure Trajectory Recovery from Motion Blur, TPAMI 2021.
Thank you for your detailed response. My concerns have been addressed, and I will therefore maintain my positive rating.
We sincerely appreciate your encouraging follow-up and continued support. Your valuable feedback has helped us strengthen our work by more thoroughly examining the benefits and computational costs of our framework and presenting experimental comparisons more transparently. We are pleased that our responses have addressed your concerns, and we have incorporated the additional results and clarifications into the revised manuscript. We are truly grateful for your careful review and constructive suggestions, which have been instrumental in improving our work.
We thank the reviewers for their constructive feedback and address their concerns below. Newly added text and figure/table captions in the revised manuscript are temporarily highlighted in blue, while table contents remain black for readability. We have also condensed existing text to accommodate additional experiments and welcome further discussion.
[A1] Independent Evaluation of Trajectory Diversity and Blur Realism
We appreciate the reviewer's concern regarding the independent validation of trajectory diversity and generated blur realism. To address this, we conduct an additional trajectory diversity analysis using Motion-ETR [1], an independent estimator that models motion trajectories over the exposure interval.
As shown in Figure 7 and discussed in Section 4.4 of the revised manuscript, the independent analysis reveals diverse trajectory clusters in the original GoPro dataset, consistent with our original findings. Existing augmentation methods remain concentrated in a limited number of clusters, whereas our method achieves broader coverage across different trajectory patterns.
Our primary objective is to improve downstream deblurring performance by adaptively generating diverse blur patterns tailored to the training dataset and target network. To further evaluate the generated blur, we measure recall and coverage for diversity, along with FID and CMMD for distributional similarity. As shown in Table 4 of the revised manuscript (reproduced below), our method outperforms ID-Blau across all four metrics. Although these results do not directly establish physical realism, they suggest that our method generates diverse blur patterns without deviating substantially from the original GoPro blur distribution.
[A2] Contributions of LTA and Differentiable Augmentation
We appreciate the reviewer's concern regarding the individual contributions of LTA and differentiable augmentation. To further clarify their effects, we expanded our ablation study by including an additional GTA-only setting with random trajectory perturbations.
As shown in Table 3 of the revised manuscript (reproduced below), the expanded ablation enables direct comparisons between random and differentiable augmentation, both with and without LTA, allowing us to assess the individual contributions of local trajectory modeling and feedback-driven optimization. The results show consistent improvements from adding LTA and applying differentiable augmentation across GoPro and HIDE, with the full configuration achieving the best performance across all metrics. Together, these results support the complementary roles of LTA and differentiable augmentation in improving downstream deblurring performance.