- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes CoDA, a test-time adaptation framework that jointly adapts classifiers to corrupted inputs and their restored counterparts. It uses cross-view gating, selective prediction consistency, and entropy-aware fusion to control the interaction between the two branches.
Critical strengths
- The idea of jointly using corrupted and restored views while controlling unreliable cross-view interaction is intuitive.
- The experiments cover multiple TTA methods, backbone architectures, and evaluation protocols, with additional tests on alternative restoration models and distribution shifts.
Critical weaknesses
-
The reliability of cross-view agreement and entropy is not sufficiently analyzed. Both branches are initialized from the same classifier, and the restored image is generated from the corrupted input, so they may share incorrect predictions. Lower entropy also does not necessarily indicate better restoration or more accurate classification.
-
The comparison with closely related diffusion-based adaptation methods is limited. Diffusion-Calibrated Continual Test-Time Adaptation and Collaborative Model and Data Adaptation are discussed in Related Work, but are not directly evaluated.
-
The corruption experiments only consider severity level 5. It remains unclear whether the same strategy is beneficial under milder corruptions, where restoration may alter useful details in otherwise reliable inputs.
-
The computational cost remains substantial despite the small additional interaction overhead. Table 3 mainly compares runtime with restoration-based DDA and SDA, rather than model-only TTA.
Questions for authors
See Weaknesses.
4 comments
Part 2: Response to Weakness 2
W2: We agree that comparisons with CMDA and diffusion-calibrated continual test-time adaptation (DCTA) are important for a more complete evaluation of CoDA. Following your suggestion, we have added the following comparisons to the revised manuscript.
Comparison with CMDA: We have integrated CoDA with CMDA and added a comparison between CMDA and CMDA+CoDA in Section 5.1, Table 4. The results show that incorporating CoDA improves all three evaluated backbones, demonstrating that the proposed cross-view interaction can further enhance an existing joint data–model adaptation framework. These results support the applicability of CoDA beyond the default DDA-based restoration setting. These additions are marked in blue in the revised manuscript (L336–L345, L357–L362). The results are shown below:
| Method | Gauss | Shot | Imput | Defoc | Glass | Motion | Zoom | Frost | Snow | Fog | Brit | Contr | Elastic | Pixel | JPEG | Avg |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ResNet50 | ||||||||||||||||
| source | 6.1 | 7.5 | 6.7 | 14.3 | 7.6 | 11.8 | 21.5 | 16.2 | 16.1 | 21.4 | 55.1 | 3.6 | 14.5 | 33.3 | 42.1 | 18.5 |
| CMDA | 51.5 | 46.5 | 46.5 | 18.1 | 19.3 | 22.2 | 29.7 | 28.2 | 34.2 | 16.2 | 58.1 | 1.2 | 43.8 | 52.8 | 54.6 | 34.9 |
| +CoDA | 53.1 | 50.4 | 49.8 | 23.4 | 28.0 | 37.6 | 46.0 | 45.9 | 44.5 | 54.6 | 65.1 | 23.2 | 56.4 | 59.0 | 56.9 | 46.3 |
| SwinT | ||||||||||||||||
| source | 28.8 | 27.1 | 27.2 | 22.1 | 9.5 | 23.6 | 27.4 | 47.1 | 36.0 | 46.8 | 68.0 | 34.0 | 21.4 | 26.9 | 50.2 | 33.1 |
| CMDA | 54.9 | 50.1 | 50.3 | 34.9 | 37.1 | 40.3 | 39.7 | 43.2 | 38.6 | 43.6 | 64.5 | 35.2 | 49.5 | 58.0 | 56.4 | 46.4 |
| +CoDA | 53.8 | 51.5 | 50.9 | 33.7 | 37.6 | 41.6 | 43.7 | 48.2 | 47.4 | 51.4 | 70.4 | 41.0 | 50.7 | 56.2 | 57.6 | 49.0 |
| ConvT | ||||||||||||||||
| source | 40.1 | 39.1 | 38.7 | 25.6 | 11.4 | 33.0 | 31.2 | 49.3 | 43.8 | 41.9 | 70.3 | 45.0 | 22.5 | 41.0 | 57.2 | 39.3 |
| CMDA | 59.7 | 55.4 | 54.9 | 44.3 | 44.8 | 49.5 | 46.4 | 50.6 | 43.3 | 42.4 | 70.4 | 43.8 | 53.3 | 62.5 | 63.2 | 53.2 |
| +CoDA | 60.0 | 57.8 | 57.9 | 45.3 | 46.3 | 51.5 | 53.1 | 57.4 | 57.4 | 61.4 | 72.4 | 56.8 | 59.9 | 64.8 | 65.0 | 57.8 |
Comparison with DCTA: We have added a comparison with DCTA in Appendix D.2, Table 19. As DCTA's code is not yet publicly available, we integrate CoDA with REM, a baseline used in DCTA, to provide a fair comparison using the same source checkpoint. Meanwhile, we are actively contacting the authors for further implementation details. The additions are marked in blue in the revised manuscript (L1260–L1269). The results are shown below:
| Method | Gauss | Shot | Imput | Defoc | Glass | Motion | Zoom | Frost | Snow | Fog | Brit | Contr | Elastic | Pixel | JPEG | Avg |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| source | 47.0 | 48.2 | 47.9 | 31.5 | 21.2 | 41.5 | 36.7 | 45.8 | 50.1 | 42.3 | 73.6 | 8.6 | 42.5 | 62.0 | 63.8 | 44.2 |
| Rem | 56.5 | 61.9 | 60.8 | 46.8 | 51.0 | 56.5 | 57.2 | 64.8 | 62.5 | 64.6 | 76.8 | 53.2 | 58.4 | 71.1 | 69.8 | 60.8 |
| CTTA | 57.8 | 59.3 | 61.8 | 51.7 | 48.8 | 59.0 | 56.7 | 66.0 | 65.1 | 63.7 | 79.8 | 55.1 | 58.9 | 69.8 | 71.2 | 61.7 |
| CODA(Rem) | 66.3 | 64.5 | 64.2 | 55.0 | 58.1 | 61.4 | 61.9 | 65.1 | 64.8 | 67.4 | 77.0 | 55.9 | 68.8 | 74.2 | 71.7 | 65.1 |
Part 3: Response to Weakness 3, 4
W3: Thank you for your constructive suggestion. We have extended our experiments to ImageNet-C at severity 3 using DPM-Solver++, which accelerates restoration with fewer denoising steps but may compromise restoration quality and introduce less reliable prior information. Despite this, CoDA still achieves performance gains, demonstrating its effectiveness under milder corruptions. The additional experiments are marked in blue in Appendix C.4, Table 17 (L1198–L1221). The results are shown below:
| Method | ResNet50 | ViT-B | ViT-S | ConvT | SwinT | Avg. |
|---|---|---|---|---|---|---|
| EATA | 57.1 | 68.6 | 57.6 | 58.9 | 53.5 | 59.1 |
| +DDA/CoDA | 58.0 / 59.5 | 68.4 / 69.6 | 59.6 / 61.2 | 59.8 / 60.7 | 56.0 / 57.1 | 60.4 / 61.6 |
| SAR | 54.3 | 62.2 | 51.1 | 58.9 | 51.8 | 55.7 |
| +DDA/CoDA | 56.6 / 57.9 | 62.6 / 64.5 | 54.7 / 56.6 | 58.6 / 60.2 | 54.5 / 55.8 | 57.4 / 59.0 |
| DeYO | 57.2 | 69.8 | 58.6 | 59.1 | 54.4 | 59.8 |
| +DDA/CoDA | 58.3 / 59.7 | 67.4 / 70.5 | 60.0 / 62.0 | 59.7 / 61.1 | 56.0 / 57.5 | 60.3 / 62.2 |
| COME | 53.5 | 69.8 | 58.5 | 60.4 | 57.9 | 60.0 |
| +DDA/CoDA | 55.2 / 55.9 | 66.5 / 70.6 | 59.6 / 61.9 | 61.1 / 62.1 | 58.8 / 59.9 | 60.2 / 62.1 |
| ZeroSiam | 59.3 | 68.5 | 59.8 | 59.7 | 54.8 | 60.4 |
| +DDA/CoDA | 60.2 / 61.9 | 67.9 / 69.4 | 61.3 / 62.8 | 60.4 / 61.6 | 56.6 / 57.9 | 61.3 / 62.7 |
W4: We agree that the computational cost deserves further clarification. We have added a direct comparison with model-only TTA, confirming that image restoration dominates the total runtime (9,854 ms out of 9,941 ms). Compared with DDA+ZeroSiam, CoDA introduces only 17 ms additional latency and 12 MB additional GPU memory usage. To reduce restoration costs, we further evaluated AdaIR, a non-diffusion restoration model, which substantially reduces computational cost while maintaining performance improvements. We have also explicitly discussed this limitation in the Limitations section. The additional results are presented in Tables 22 and 23 and marked in blue in the revised manuscript (L1374–L1403). The results are shown below:
| Method | Total (ms) | Restore (ms) | Adapt (ms) | Memory (MB) |
|---|---|---|---|---|
| ZeroSiam | 32 | – | – | 966 |
| DDA+ZeroSiam | 9924 | 9854 | 70 | 16163 |
| CoDA+ZeroSiam | 9941 | 9854 | 87 | 16175 |
| Method | Batch | Total (ms) | Restore (ms) | Adapt (ms) | Memory (MB) |
|---|---|---|---|---|---|
| AdaIR | 16 | 738 | 668 | 70 | 6892 |
| AdaIR | 32 | 1386 | 1299 | 87 | 13665 |
We hope that our responses and the additional experimental results in the revised manuscript have provided sufficient clarification and addressed your concerns. We sincerely appreciate your valuable feedback and welcome any further questions or suggestions!
Supplementary Experimental Results for W1
To provide further empirical evidence regarding W1, we have conducted an additional stress test on CoDA+SAR to evaluate the robustness of Selective Coupling (SC) to shared incorrect predictions.
Specifically, we randomly discard 50% of the correctly agreed samples while retaining all incorrectly agreed samples when computing the SC loss. Ground-truth labels are used only for this diagnostic experiment. The addition results are marked in blue in the revised manuscript (Appendix E.6, Table 25, L1418–L1440). The results are shown below:
| Backbone | Original CoDA | Stress Test |
|---|---|---|
| ResNet50 | 41.40 | 40.92 |
| ViT-B | 59.40 | 58.87 |
| ViT-S | 47.30 | 46.62 |
| ConvNeXt-T | 44.60 | 44.12 |
| Swin-T | 40.40 | 40.10 |
Despite the increased proportion of shared incorrect predictions in SC, accuracy decreases by only 0.30–0.68 percentage points across all five backbones. Together with the improved sample selection accuracy reported in Table 5, these results provide further empirical evidence of the robustness of CoDA's cross-view interaction mechanisms.
We sincerely hope that these additional results help address your concerns, and we greatly appreciate your careful consideration of our work.
Part 1: Response to Weakness 1
We sincerely appreciate your thoughtful comments and your recognition of CoDA’s intuitive design and our extensive experiments across different methods and settings. We have carefully considered all the points you raised and have addressed each of your concerns in detail below.
W1: We agree that neither cross-view agreement nor lower entropy necessarily guarantees correct predictions. CoDA does not rely on these signals alone. Instead, cross-view gating applies secondary filtering to samples already selected for updates by the underlying TTA method, while consistency regularization is restricted to samples passing both gates. Entropy serves as a relative uncertainty indicator rather than a guarantee of correctness. To further validate this design, we added Table 5, showing that the accuracy of samples selected for model updates increases from 45.89% to 56.53% for ZeroSiam and from 55.43% to 62.61% for SAR. These results demonstrate that our mechanism favors more reliable updates, reducing the risk of reinforcing shared errors. The additions are marked in blue in the revised manuscript (L396–L411). The results are shown below: