We thank the reviewer for the additional comment on our response.
> [Q1] As for A1, please refer to my response to the rebuttal in Reviewer vns8.
- Please refer to *our response* to your concern [here](https://openreview.net/forum?id=JXvszuOqY3¬eId=KJgF8zZ1ZE).
> [Q2] As for A2, the original method evidently lags behind AugSeg and UniMatch in most splits. Besides, these comparisons should be included in the initial submission, rather than rebuttal. During rebuttal, the newly added component Adaptive CutMix (from AugSeg) would make the whole framework much more sophisticated (more hyper-parameters) and unstable.
- We respectfully disagree with the reviewer’s concern that we lagged behind AugSeg (*CVPR2023*) and UniMatch (*CVPR2023*) in most splits. We followed your suggestion and reported extra results to compare them fairly. We are confident that our original method (Ours) outperformed AugSeg (e.g., in 1/8 and 1/4 in ResNet-50, and 1/4 in ResNet-101) and UniMatch (e.g., 1/8 in ResNet-50, and 1/4 partitions in ResNet-101).
- Our point was not about the complexity of the method, but we just employed different augmentation, such as adaptive CutMix (dubbed Ours+). We found that ours were improved (e.g., 1/16, 1/8 in ResNet-50, and 1/4 in ResNet-101), and other results are comparable to the prior works. This experiment was *auxiliary* to this question but would partially address Q4 and Q5.
- Please remember that this paper's main contribution is NOT introducing a new augmentation method. Our main contribution to this paper is *switching temporary teachers* to secure teacher diversity, a simple yet efficient method for semi-supervised learning tasks.
> [Q3] As for A3, why not include UniMatch, especially since the authors have included UniMatch in A2. The results of UniMatch under the five splits are 71.9, 72.5, 76.0, 77.4, and 78.7, respectively. In comparison, the results of this submission are 69.3(-2.6), 71.02(-1.4), 76.09(+0.1), 77.86(+0.5), and 78.28(-0.4) respectively.
- We did not; because the input resolution of UniMatch differs from other methods as well as ours. Most of the methods (including ours) employed *512*; UniMatch used *321* for training resolution.
> [Q4] As for A4, under the same split, the improvement is very marginal, which are -0.4, +0.2, +0.2 under the three splits respectively. More importantly, I believe such fair comparisons under the same split should be included in the very initial submission, instead of reporting very mislearning results in Table 1 of the submission.
- As for the marginal improvement, we gently remind you UniMatch and AugSeg are *CVPR-2023* papers. Compared with PS-MT (*CVPR-2022*), we consistently achieved 0.5-1.5% gains. Moreover, on Cityscapes, our original method achieved the best performances in all but one partition (three were the best, and one was the second-best).
- We here provide a summarized table for **Cityscapes**.
|Method|1/16 (186)| 1/8 (372)| 1/4 (744)| 1/2 (1,488)|
|---|---|---|---|---|
|PS-MT|-|76.9|77.6|79.09|
|AugSeg|75.22|77.82|**79.56**|80.43|
|UniMatch|76.6|77.9|79.2|79.5|
|Ours|**76.81**|**78.4**|79.46 |**80.52**|
- Our main contribution is introducing switching temporary teachers (i.e., teacher diversity) can improve the performance of semi-supervised learning, not simply aiming for the absolute best performance across all datasets.
- UniMatch and AugSeg have significant contributions, and we recognize their importance. We believe the contributions are positively aligned with our method; as an example, we employed the adaptive CutMix utilized in AugSeg to provide an upbeat hybrid following your valuable comments. We believe Ours+ revealed room for ours' further improvements, but It is worth noting that the above table now does not include Ours+ (using adaptive CutMix).
- Please note that in response to your comment to Reviewer vns8, we corrected the performance comparison by assessing it on another split and datasets (we will revise the numbers in Table 1).
> [Q5] As for A5, I did not mean trying more advanced augmentations. I hoped to see the results without using ClassMix.
- We kindly remind your comment, “Does the improvement come from more diverse spatial augmentations, i.e., CutMix and ClassMix?” We understood the question was to see the applicability of further diverse augmentations, so we opted for a diverse augmentation.
- The main improvement of our method may come from teacher diversity and decoupling, as addressed in our [response](https://openreview.net/forum?id=JXvszuOqY3¬eId=MncAE7jA4h) to Reviewer vns8. In this respect, we could yield improvements by ensuring *more diversity* using adaptive CutMix with *switching temporary teaching*, as we reported in our previous response.
- We kindly remind you our method with *dual teachers employed two different augmentations*, such as CutMix and ClassMix, for sufficient diversity, so solely removing ClassMix would lose the performance; therefore used Adaptive Cutmix.