- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes I²EM, a dual-exposure night image restoration framework that explicitly separates inter-exposure displacement and intra-exposure motion blur using optical flow and spatially varying kernels. The method achieves strong quantitative results on the D2 benchmark. However, the methodological novelty appears relatively limited compared with existing multi-exposure restoration approaches, and the effectiveness of the proposed framework on real captured dual-exposure images requires more convincing validation.
Critical strengths
-
Well-motivated design. Explicitly distinguishing inter-exposure displacement from intra-exposure blur provides a reasonable approach to improving cross-exposure alignment and information fusion.
-
Strong quantitative performance. The method achieves 40.78 dB PSNR on the D2 benchmark, outperforming the directly comparable D2HNet baseline by 2.83 dB.
-
Extensive ablation studies. The paper evaluates several key components, including bidirectional flow estimation, kernel modeling, reconstruction constraints, and confidence-aware fusion.
Critical weaknesses
-
Insufficient dual-exposure comparisons. Although Table 1 includes numerous baselines, D2HNet is the only directly comparable method under the same dual-exposure setting. Additional comparisons with relevant methods, such as LSD₂, LSFNet, and DeLECA, would strengthen the evaluation.
-
Unclear baseline training protocols. The paper does not clearly specify whether the compared methods were retrained on the D2 dataset or evaluated using pretrained weights. The authors should clarify the training settings, checkpoint sources, and input adaptations to ensure fair comparisons.
Questions for authors
Please refer to the weakness.
3 comments
Q2:Unclear baseline training protocols. The paper does not clearly specify whether the compared methods were retrained on the D2 dataset or evaluated using pretrained weights. The authors should clarify the training settings, checkpoint sources, and input adaptations to ensure fair comparisons.
In the manuscript, we have specified the common D2-2880p scenes, full-resolution inference, and metric implementation. In the revision, Section 4.1 clarifies that we use released weights where an applicable checkpoint is available and otherwise adapt the implementation and train it on the official D2 split (manuscript lines 331–340).
The “Weights/training” columns in Tables 1 and 2 identify the protocol for each baseline (manuscript lines 345–377). DarkIR, NoiseDiff, CDVD-TSP, ESTRNN, HINet, LSFNet-RGB, and DeLECA use D2-adapted training. The other listed baselines use released checkpoints. The tables also distinguish long-only and short-only input from the common long–short pair used for the ordinary dual-exposure methods. Appendix E.5 specifies the crop setting, number of epochs, optimizers, and loss for the two capacity controls (manuscript lines 1410–1419).
Table 1. Single-exposure night-image restoration comparison on D2-2880p
| Method | Input | Weights/training | PSNR | SSIM | LPIPS | Params (M) | FLOPs (G) |
|---|---|---|---|---|---|---|---|
| DeepDeblur | L-only | Released | 36.55 | 0.9636 | 0.095 | 42.3 | 156.8 |
| DeblurGAN-v2 | L-only | Released | 36.82 | 0.9655 | 0.090 | 60.9 | 182.3 |
| MPRNet | L-only | Released | 37.08 | 0.9684 | 0.086 | 23.5 | 98.2 |
| SwinIR | L-only | Released | 37.70 | 0.9712 | 0.079 | 12.1 | 62.3 |
| MIMO-UNet+ | L-only | Released | 37.31 | 0.9690 | 0.083 | 16.1 | 42.1 |
| MPRNet-DN | S-only | Released | 37.35 | 0.9580 | 0.085 | 15.74 | 92.3 |
| NAFNet | L-only | Released | 37.70 | 0.9710 | 0.079 | 23.4 | 92.4 |
| Restormer | L-only | Released | 37.87 | 0.9717 | 0.077 | 26.1 | 105.7 |
| Uformer | L-only | Released | 37.55 | 0.9702 | 0.081 | 28.3 | 118.4 |
| HAT | L-only | Released | 37.85 | 0.9715 | 0.078 | 12.4 | 98.3 |
| PromptIR | L-only | Released | 37.45 | 0.9695 | 0.082 | 87.2 | 138.6 |
| UFPNet | L-only | Released | 37.72 | 0.9712 | 0.079 | 80.3 | 290.4 |
| EVSSM | L-only | Released | 38.09 | 0.9727 | 0.072 | 17.1 | 126.0 |
| DeblurDiff | L-only | Released | 37.55 | 0.9703 | 0.050 | 386.2 | 2890.0 |
| DarkIR | S-only | D2-adapted training | 37.65 | 0.9705 | 0.080 | 35.2 | 128.5 |
| NSP | S-only | Released | 37.25 | 0.9685 | 0.085 | 28.5 | 115.2 |
| NoiseDiff | S-only | D2-adapted training | 37.95 | 0.9718 | 0.075 | 120.5 | 285.3 |
| I²EM | Long–short pair | D2 training | 40.78 | 0.9783 | 0.054 | 133.79 | 462.8 |
Table 2. Dual-exposure night-image restoration comparison on D2-2880p with a shared long–short input pair
| Method | Weights/training | PSNR | SSIM | LPIPS | Params (M) | FLOPs (G) | Time (ms) |
|---|---|---|---|---|---|---|---|
| CDVD-TSP | D2-adapted training | 37.00 | 0.9668 | 0.087 | 16.6 | 88.5 | 23 |
| ESTRNN | D2-adapted training | 37.10 | 0.9675 | 0.088 | 20.5 | 78.5 | 25 |
| LSD₂ | Released | 38.42 | 0.9741 | 0.071 | 28.5 | 142.8 | 18 |
| HINet | D2-adapted training | 37.22 | 0.9685 | 0.083 | 88.7 | 112.3 | 6 |
| D2HNet | Released | 37.95 | 0.9723 | 0.076 | 35.2 | 165.3 | 12 |
| LSFNet-RGB | D2-adapted training | 38.15 | 0.9732 | 0.073 | 32.5 | 158.4 | 22 |
| SelfHDR | Released | 38.85 | 0.9752 | 0.065 | 118.2 | 425.8 | 38 |
| DeLECA | D2-adapted training | 39.15 | 0.9758 | 0.063 | 45.8 | 198.5 | 25 |
| BracketIRE | Released | 39.25 | 0.9761 | 0.061 | 128.5 | 498.2 | 45 |
| STGDNet | Released | 38.12 | 0.9729 | 0.072 | 28.4 | 142.6 | 15 |
| I²EM | D2 training | 40.78 | 0.9783 | 0.054 | 133.79 | 462.8 | 78 |
I maintain my weak reject decision, as I believe the experimental comparisons are unfair.
Q1:Insufficient dual-exposure comparisons. Although Table 1 includes numerous baselines, D2HNet is the only directly comparable method under the same dual-exposure setting. Additional comparisons with relevant methods, such as LSD₂, LSFNet, and DeLECA, would strengthen the evaluation.
In section 4.2(manuscript lines 342–378), we have compared with D2HNet on the same pair and have also evaluated methods that use multiple observations, such as CDVD-TSP and ESTRNN, on D2. To extend the direct comparison, the revised Table 2 evaluates LSD₂, LSFNet-RGB, and DeLECA using the same long–short pair as I²EM (manuscript lines 363–377). We use the released checkpoint for LSD₂ and adapt and train the other two implementations on the official D2 training split (Section 4.1, manuscript lines 331–336).
Under the shared protocol, these three methods reach 38.42, 38.15, and 39.15 dB, respectively, while I²EM reaches 40.78 dB. The comparison therefore shows that the PSNR advantage persists across the additional methods with the same exposure inputs.
Table 2. Dual-exposure night-image restoration comparison on D2-2880p with a shared long–short input pair