We express our gratitude to Reviewer YXjG for acknowledging our extensive experiments and state-of-the-art performance. Your assistance in refining metric definitions is highly appreciated.

Q1: Contact Error depends on absolute scale.

We agree that Contact Error should not depend on the absolute scale. Therefore, before calculating Contact Error, we normalize the DMI field by the limb radius of the character. Please refer to line 511 in MeshRetCode/model/retnet.py from our supplemental zip file for this detail.

The precise definition of Contact Error should be:

where $R_A$ and $R_B$ represent the radius of each character's arms.

We apologize for not including this implementation detail in our paper and will correct this definition in our final manuscript.

Q2: How diverse are the skeletons in your dataset? Need for results on characters with different relative bone lengths.

We plot the skeleton length distribution in Figure 15 of our rebuttal PDF. In the ScanRet dataset, the arm lengths of characters range from 51.9 cm to 67.0 cm, body heights range from 153.4 cm to 185.1 cm, and arm-height ratios range from 0.32 to 0.39. In the ScanRet dataset, the arm lengths of characters range from 35.8 cm to 134.2 cm, body heights range from 120.4 cm to 368.3 cm, and arm-height ratios range from 0.28 to 0.45.

To see results on characters with different relative bone lengths, please refer to Figures 10-13 in our paper and the last two clips in our demo video from the supplemental zip file. In these results, our method performs well on characters with different relative bone lengths.

Q3: What if I would like to use the method on a character system with a different number of bones and connectivities?

To process characters with different bone numbers, such as varying spine bone numbers, we can split or merge neighboring bones and then attach sensors to the modified virtual bone system. This approach requires no training or modification of the neural network. However, to process characters with different bone connectivities, we need to collect a new dataset and retrain the neural network. Additionally, some hyper-parameters, such as the dimensions of certain embedding layers, must be modified. Notably, previous state-of-the-art methods like R2ET and SAN also require characters to share a common skeleton topology.

Dear Reviewer,

Thank you once again for your feedback. As the rebuttal period is nearing its conclusion, could you kindly review our response to ensure it addresses your concerns?

We appreciate your time and input.

Sincerely,

The Authors

Thank you for your valuable feedback and support. We are pleased to confirm that we have thoroughly addressed your concerns. Given the limited time remaining in the rebuttal phase and the necessity of refactoring some of our code, we will strive to provide experimental results on characters with varying bone numbers. If we are unable to complete this before the rebuttal deadline, we commit to including experiments with varying bone counts in our final manuscript, accompanied by a comprehensive discussion of the results. All the discussion about our method's limitation will also be included.

We extend our appreciation to Reviewer utNP for recognizing the novelty of our method and the effectiveness of the proposed DMI field. We sincerely value your recommendation to include additional experimental results.

Q1: More ablation studies on different sensor arrangements.

We provide additional ablation experiments here. We conducted further studies on different sensor arrangements. Specifically, we evaluated the performance of a model trained with half the sample points in the $\phi$ space in SCS, denoted as ${Ours}\_\phi$, and another model trained with half the sample points in the $l$ space in SCS, referred to as ${Ours}\_l$. As shown in the following table, ${Ours}_\phi$ compromises the interpenetration ratio, indicating that sufficient sample points in the $\phi$ space are crucial for avoiding interpenetration. We also found that both models introduce artifacts; please refer to Figure 14 in our rebuttal PDF.

We will add the results to the final manuscript.

Q2: More results with ScanRet characters as target characters.

We provide additional qualitative results using ScanRet characters as target characters. Please refer to Figure 17 in our rebuttal PDF. For quantitative results, the metrics in Table 1 are computed using both ScanRet characters and Mixamo characters as targets.

Q3: Comparison with "Contact-aware Retargeting of Skinned Motion."

We value your advice on including more baseline methods. We intended to compare our method to "Contact-aware Retargeting of Skinned Motion," but we were unable to do so because the authors did not release their code, and some implementation details, such as the vertex correspondence feature, were not described in detail. Notably, the authors of R2ET did not compare their method to this one in their paper. In general, we have made our best effort to provide a comprehensive comparison with existing methods, including the currently open-sourced state-of-the-art method, R2ET.

Q4: The results only trained on the Mixamo dataset.

We value your advice on including more results. The following table presents a comparison between our methods and state-of-the-art methods trained only on the Mixamo dataset.

Because the Mixamo dataset does not always provide clean contact input, the performance of our method is affected. However, our method still achieves much lower Contact Error and Penetration Ratio compared to baseline methods. Please refer to Figure 16 in our rebuttal PDF for qualitative results.

L1: Failure cases under noisy conditions.

We provide some failure cases in Figures 16 and 17 in our rebuttal PDF. We will include these in our final manuscript.

L2: How to attach sensors for disabled people with missing limbs and characters with different bone numbers?

To process characters with different bone counts, such as varying spine bone numbers, we can first split or merge neighboring bones and then attach sensors to the modified virtual bone system. However, our method cannot process characters with missing limbs. We will address this limitation in our final manuscript and leave it to future work.

We are grateful to Reviewer WToy for acknowledging our novel technical contributions and the effectiveness of our experimental evidence. We also appreciate your guidance in clarifying the notation in the DMI field description.

Q1: What are the number $S$ and $K$?

We apologize for any confusion. The number $S$ is determined by the size of the SCS semantic coordinates set. Please refer to Line 257 for the specific semantic coordinates set used in our experiment, where $S = 288$. Your understanding of $K$ is correct. After masking the full DMI field $\overline{**D**} \in \mathbb{R}^{S \times S \times P}$ with $\mathbf{M}_{\text{src}} \in \mathbb{R}^{S \times S}$, we obtain the final DMI field $\mathbf{D} \in \mathbb{R}^{K \times L \times P}$, where $K$ represents the number of observation sensors in $\mathbf{D}$. We will include a more detailed explanation in the final version of our paper.

We extend our gratitude to Reviewer yJ9L for recognizing the novelty and state-of-the-art performance of our method. Your guidance on balancing proximal and distal sensor pairs has been particularly insightful.

Q1: How to choose the “cls”, “far”, or “dm” version?

We appreciate your suggestion to discuss the balance problem between proximal and distal sensor pairs. In the ablation study presented in Section 4.3, the "dm" version ($Ours_{dm}$), the "cls" version ($\mathrm{Ours}_{cls}$), and the "far" version ($\mathrm{Ours}\_{far}$) all exhibit worse performance compared to the full approach (Ours). The full approach can be considered a mixed version of $\mathrm{Ours}\_{cls}$ and $\mathrm{Ours}\_{far}$, utilizing an equal distribution of proximal and distal sensor pairs. To better illustrate this balance, we provide additional experimental results by testing different ratios of proximal to distal sensor pairs. The following table compares our method's performance with varying percentages of proximal sensor pairs under the Mixamo+ setting. As the percentage of proximal sensor pairs decreases, the interpenetration ratio fluctuates mildly, while the contact error initially decreases and then increases. Finally, with no proximal pairs (equivalent to the "far" version), the performance drops significantly.

In Figure 18 of our rebuttal PDF, we present a qualitative comparison of our methods using different proximal sensor pair ratios. Except for the 100% Proximal version (equivalent to $\mathrm{Ours}\_{cls}$) and the 0% Proximal version (equivalent to $\mathrm{Ours}\_{far}$), our method demonstrates fair robustness to the proximal sensor ratio in the 25%-75% interval.

Based on these results, we conclude that choosing 50% proximal sensor pairs strikes a reasonable balance for achieving good performance. We will include the additional results in the final manuscript.

As for the balance between enforcing contact and avoiding interpenetration, our method directly captures the interaction between mesh surfaces, as illustrated in Figure 3. Therefore, there is no explicit contradiction between preserving contact and avoiding interpenetration.

L1: The SCS extraction may be affected by noisy mesh input.

Thank you for your valuable comments. Noisy mesh surfaces can indeed pose challenges. One potential solution is to adopt a Laplacian-smoothed proxy mesh of the character for the SCS extraction. We have not implemented this due to the limited time available during the rebuttal phase but plan to discuss this further in our final manuscript.

Dear Reviewer,

Thank you once again for your feedback. As the rebuttal period is nearing its conclusion, could you kindly review our response to ensure it addresses your concerns?

We appreciate your time and input.

Sincerely,

The Authors