Authors' Response to Reviewer qCdJ
We thank Reviewer qCdJ for devoting time to the Author-Reviewer discussion session and providing follow-up comments. Our responses are as follows.
---
> ***Q-i:** "Do existing works that you listed all use I3D backbone only? Or are there other backbone being used?"*
**A:** Thanks for the question.
- Existing works that we are comparing with, including CO2A [13], CoMix [14], MM-SADA [18], STCDA [19], CMCD [20], and CIA [21], all adopt I3D as their backbones.
- Early works [11,12,16] used ResNet-101 as the backbone.
- It has been widely verified in empirical ways that using the I3D backbone can achieve satisfactory performance across existing video-based UDA benchmarks, including those large-scale ones like UCF-HMDB and Epic-Kitchen.
- We believe more powerful backbones are promising; as larger and higher-quality datasets come out, there will require an upgrade on the backbones for video-based UDA tasks.
---
> ***Q-ii:** "What notions are going to be added, could you please show me?"*
**A:** We omit the notation table in the previous rebuttal window due to the lack of space (there is a 6000-character limit). Specifically, we supplemented the following table in the revised manuscript to improve the readability:
| Notation | Description
|-|-
|$\mathcal{D}$|Domain
|$\mathcal{S}$ / $\mathcal{T}$|Source domain / Target domain
|$\mathbf{V}^\mathcal{D}$|A video sequence from domain $\mathcal{D}$
|$\{\mathbf{V}_i^\mathcal{D},y_i^\mathcal{D} \}$|The $i$-th video sequence and the corresponding action label of domain $\mathcal{D}$
|$\{\mathbf{x}_i^\mathcal{D},…,\mathbf{x}_T^\mathcal{D} \}$|$T$ frames of images in the video sequence $\mathbf{V}^\mathcal{D}$
|$\{\mathbf{z}_i^\mathcal{D},…,\mathbf{z}_T^\mathcal{D} \}$|The dynamic latent factors of the video sequence from domain $\mathcal{D}$
|$\mathbf{z}_d^\mathcal{D}$|The static latent factors of the video sequence from domain $\mathcal{D}$
|$\mathbf{x}_{<t}^\mathcal{D}$|The input sequence before timestamp $t$
|$\mathbf{x}_{1:T}^\mathcal{D}$|The full input sequence from $t=1$ to $t= T$
Please let us know if there is any other notation that is unclear and could be clarified. Thanks a lot.
---
> ***Q-iii:** "'Are there extra discussions/comparisons in Sec. 4.3 that are interesting and worth to be added?"*
**A:** Thanks for your question.
- Section 4.3 *'Compared to Multi-Modal Methods'* is to compare and analyze TranSVAE with existing multi-modal UDA methods, and Table 4 is to show the corresponding comparison results. In the submitted version, we accidentally forgot to refer the table to the Section, while the analyses are already there.
- In this revision, we have added comparisons and analyses on more multi-modal methods, including A3R, CydDA, MixDANN, and CleanAdapt, as suggested by *Reviewer cyvj* and *Reviewer AobY*. Please kindly check our responses to these reviewers for more details, thanks.
---
> ***Q-iv:** "What are the updates going to be made based on the above rebuttal in addressing the raised concerns?"*
**A:** Thanks for your question. We summarized the changes and improvements that we made during this rebuttal in the General Response section (copy below). We appreciate your valuable comments and suggestions. Please let us know if there is anything that we could further modify or improve. Thanks.
- We have supplemented more ablation experiments, including TranSVAE w/ and w/o disentanglement, to verify the effectiveness of each objective term in our framework.
- We have compared with more SOTA baselines that use multi-modality data sources for video-based UDA, including A3R, CleanAdapt, CycDA, and MixDANN. All the comparison results show that our TransVAE is still a competitive method even using only a single RGB modality.
- We have added complexity analyses on the number of trainable parameters, multiply-accumulate operations, floating-point operations, and inference frame-per-second for our TranSVAE as well as for prior arts.
- For better readability, we have supplemented a notation section for the maths symbols used in the paper.
- We have added missing references suggested by the reviewers.
- We have polished and improved the elaboration of this work.
- We have carefully addressed other comments from the reviewers point by point.
---
> ***Q-v:** "Any extra information in Fig. 5 & 6 worth discussing or analysis, or any interesting insights or comparisons to show readers?"*
**A:** Thanks for your question.
- The two figures which show the ablation study results on U/H datasets are indicating the effectiveness of each loss term in the TranSVAE framework. The reason why we showed two figures is for the completeness of U/H datasets, as some reviewers might be interested in the ablation results on both tasks.
- As more contents and improvements have been supplemented in the revision, we have moved Fig. 6 to the Appendix due to limited space.
---
Last but not least, we thank Reviewer qCdJ again for the time and effort devoted to this review.