We sincerely thank you for your valuable comments and constructive suggestions! We are encouraged that you find our idea insightful, our experiments extensive and our dataset valuable for the community. Below, we will address your proposed Weaknesses (W) and Questions (Q) in detail.

W1: The proposed framework seems lack of novelty. The main part of the proposed EDPN is the fusion model, which utlizes a cross attention block and a fusion block. However, both cross-attention fusion and conv fusion are all widely used methods in the field of AI. The rest of the proposed framework are all directly rely on existing methods.

We appreciate your feedback and understand this concern. We respectfully argue that the novelty of our work should be viewed from the holistic perspective of the proposed framework (please refer to the proposed second contribution in our paper L63-65). We are the first to introduce an EDFV-based prompting method for depth estimation to tackle the problem of scale ambiguities and visual illusions in image foundation models, and experimental results (Sec. 5.3 and 5.4 of our main paper and Sec. E of Supplementary Material) have shown our superiority against foundation models and conventional DFF methods. Each component in the pipeline is essential to the overall method (as discussed in Ablation studies in Supplementary Material Sec. D).

While a single component may seem straightforward, our combined pipeline introduces a cohesive and tailored framework that enables effective prompting in a challenging hybrid modality. As discussed in the Introduction (L48-50), the information given by event-based sensors is different from other sensors such as traiditional RGB sensors and LiDAR, as it mainly triggers at texture and edge regions. This poses unique challenges and opportunities for our design. In our pipeline, we first utilize proposed EDFV to process events into depth map. After that, we design a SAME strategy to extract effective information from events (L179-181 of our paper). Besides, considering the modality discrepencies between events and image, we employ different layers to capture their features respectively (L199-201 of our paper). We also introduce deformable convolutions to handle the sparsity and possible misalignments between two modalities (L207-209 of our paper). While we agree that the cross-attention fusion and conv fusion may be widely used methods in the field of AI, our adaptation of these modules within a prompting framework tailored for event-based DFF is novel in its use case and validated through extensive experiments. We will revise the paper to clarify this contribution and better differentiate it from related works.

W2: The proposed approach utlizes a pre-trained image foundation model as the image encoder, which seems unfair compared to the other methods. The authors should prove the performance gain is not caused by this.

Thank you for raising this point. Although other traditional DFF methods (including DefocusNet [3], DFF-FV [4], DFF-DFV [4], and DDFS [5]) do not utilize existing image foundation models, Depth Anything V2 [1] and Metric3D V2 [2] are foundation models themselves. The results of DA V2 in Table 1&2 of main paper are finetuned from its pretrained ViT-L checkpoint, while ours also utilizes the pretrained DA V2 ViT-L checkpoint as image encoder. Therefore, the performance gain of ours against the finetuned DA V2 should not be attributed to the pretrained foundation model, but rather the adaptation of the proposed EDFV and prompting framework.

Besides, in order to exclude the influence of pretrained checkpoints, we further conduct ablation studies on the Blender-Syn dataset. We train DA V2 from scratch on the Blender-Syn dataset without any pretrained checkpoint. Then we adopt this DA V2 as the image encoder to train our own framework. The results are shown in the following table, where DA V2 trained from scratch is denoted as "DA V2*", and our method trained with this scheme is denoted as "Ours*". It can be found that even without any pretrained checkpoint, our method still outperforms most of the existing DFF methods in most metrics. Therefore, we believe that the performance gain is not caused solely by pretrained image foundation models.

Table R1: Quantitative comparisons of in-domain metric depth estimation.

Blender-Syn

One of our contributions is to fuse image foundation models with event-based DFF through our EDFV-based prompting strategy. Therefore, from the comparison with the image foundation models (DA V2 [1] and Metric3D V2 [2]), we demonstrate that our strategy boosts performance, even outperforming these advanced pretrained models. This shows the benefit of our proposed EDFV and fusion framework, beyond the strength of the encoder itself.

W3: The authors use a pre-trained visual foundation model in zero-shot learning task which may causes information leak.

Thank you for this important concern. We will address this concern in the following two aspects:

• We have carefully verified the training datasets of our utilized pretrained foundation models, including DA V2 [1] and Metric3D V2 [2], and confirm that none of the image datasets used in our evaluations — specifically Sintel and 4D Light Field Dataset — were included in their training corpora.

• We validate generalization performance on our captured EDFV-real dataset, as shown in Fig. 5 of main paper and Fig. 12 of Supplementary Material. These results corroborate that our framework performs well in real, unseen scenarios, which could not be attributed to information leakage of foundation models.

W4: This paper is more suitable for the Datasets & Benchmarks track rather than the Main track.

We respectfully argue that although the constructed datasets are one of the contributions of our paper, the proposed EDFV representation as well as the prompting framework are also our key contributions. Reviewer CTBx has highlighted the novelty of EDFV, and Reviewer kC23 has emphasized our innovative framework. As far as we know, we are the first work to solve event-based DFF with prompting approach. Please refer to our responses in W1 for our contributions to the pipeline. Thanks to our EDFV and prompting framework, we could achieve better results than the existing foundation models on both synthetic and real-world datasets (please refer to Sec. 5.3 and 5.4 of our main paper). Therefore, we believe this work aligns well with the Main track.

Q1: The authors should highlight their contributions on in the framework, as it currently appears that they have merely assembled a set of existing modules.

Please refer to our responses in W1 and W4, where we elaborate on the integration of EDFV, modality-aware fusion, and task-specific prompting — together forming a task-tailored framework.

Q2: The authors should discuss the implications of using a foundation model, such as the fairness of comparisons with other baselines and the potential risk of information leakage in the zero-shot setting.

Please refer to our detailed discussions in W2 and W3 for the comparison fairness and information leakage.

[1] Yang et al. Depth Anything V2. NeurIPS 2024.

[2] Hu et al. Metric3d V2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. TPAMI 2024.

[3] Maximov et al. Focus on defocus: Bridging the synthetic to real domain gap for depth estimation. CVPR 2020.

[4] Yang et al. Deep Depth from Focus with Differential Focus Volume. CVPR 2022.

[5] Fujimura et al. Deep depth from focal stack with defocus model for camera-setting invariance. IJCV 2023.

We sincerely thank you for taking the time to re-evaluate our submission and the positive feedback. We greatly appreciate your thoughtful comments throughout the review process, which helped us improve the quality and clarity of our work. We are glad to hear that our responses addressed your concerns and that you are willing to raise your score. Thank you again for your support and encouragement !

We sincerely thank you for your valuable feedback and constructive suggestions! We are encouraged that you find our framework novel, dataset construction comprehensive and performance superior. Below, we will address your proposed Weaknesses (W) in detail.

W1: The organization of the paper could be improved. Ablation studies and a more detailed description of the dataset creation process should be placed in the main paper.

Thank you for your suggestions! Due to the page limit of the submitted version for peer review, we have to put some details in the Supplementary Material. However, we understand the importance of these components in demonstrating the validity and interpretability of our proposed approach, and will include the important details in the final version of our main paper if accepted as follows:

• We will put Additional Observation 1&2 (Supplementary Material Sec. A) and their brief derivations in the main paper (Sec. 3.2), while retaining the extended math in the Supplement.

• Since the construction of the datasets is one of our core contributions, we will migrate more dataset construction details (Supplementary Material Sec. B) to the main paper (Sec. 5.1), including the camera parameters and scene setup details.

• We will move the ablation studies (Supplementary Material Sec. D and Table 5) to the main paper as Sec. 5.5, following Sec.5.4.

We will condense our texts to make sure the final version of our main paper within page limit, concise, and more inclusive.

W2: The comparative analysis lacks baseline methods from 2025.

Thank you for highlighting this. We have now incorporated a new method, namely Depth Any Camera (DAC) [1], published in 2025 to further extend the scope of compared methods. Here we finetune DAC using its pretrained outdoor ResNet101 checkpoint. Please refer to the following tables for our comparison, where we denote DFF methods as "DFF" and monocluar methods as "Mono". We will update the Table 1 and 2 in our final version accordingly. Note that DAC may not perform well because its main focus is wide-angle cameras.

Table R1: Quantitative comparisons of in-domain metric depth estimation.

Blender-Syn

Sintel-Dr. Bokeh

Table R2: Quantitative comparisons of zero-shot metric depth estimation.

Blender-Syn

Sintel-Dr. Bokeh

W3.1: There are no quantitative results reported for the 4DLFD-Semi-Real setting or the proposed dataset.

We now provide the quantitative results on the 4DLFD-Semi-Real dataset in the following table. Here we convert all disparity predictions into depth values with unified curves for fair comparisons. It can be found that our method outperforms others in all metrics.

Table R3: Quantitative comparisons of zero-shot metric depth estimation on 4DLFD-Semi-Real.

For EDFV-Real, due to the absence of a depth sensor in our current system setup, we were unable to collect ground-truth depth maps for quantitative evaluation. Please refer to Fig. 5 of our paper and Fig. 12 of Supplementary Material for qualitative comparisons. In the future work, we aim to enhance our hardware system by integrating a depth sensor for accurate ground-truth depth capture, enabling quantitative analysis.

W3.2: The set of comparison models does not include the event-based DFF methods mentioned in the Related Work section.

Thank you for raising this point. The event-based DFF method mentioned in the Related Work section [2] could only predict sparse depth maps. We have contacted its authors, but their code is not yet ready for public release. Besides, we have also found a recent relevent method [3] published in WACV 2025, which leverages existing grayscale video reconstruction approaches for event-based depth estimation. Their method lacks the input image modality, which could constrain their zero-shot prediction performance. Their code is also not released yet, and our email is not responded as of the rebuttal deadline. We will add the comparisons of these methods as soon as they become available before camera ready if accepted.

[1] Guo et al. Depth Any Camera: Zero-Shot Metric Depth Estimation from Any Camera. CVPR 2025.

[2] Jiang et al. Learning depth from focus with event focal stack. ISJ 2024.

[3] Horikawa et al. Dense Depth from Event Focal Stack. WACV 2025.

We sincerely thank you for your insightful comments and constructive suggestions! We are encouraged that you find our idea novel, our performance superior, and our dataset valuable to the community. Below, we will address your proposed Weaknesses (W) and Questions (Q) in detail.

W1: The frequent use of abbreviations throughout the paper may confuse readers.

Thank you for your feedback. We will revise the entire manuscript to minimize the use of unnecessary abbreviations, and ensure they are clearly defined at their first appearance in the main text to improve readability.

W2: The paper’s structure needs improvement. The Ablation Study section is recommended to put in the main text.

Thank you for your suggestions! Due to the page limit of the submitted version for peer review, we have to put some details in the Supplementary Material. However, we understand the importance of these components in demonstrating the validity and interpretability of our proposed approach, and will include the important details in the final version of our main paper if accepted as follows:

• We will put Additional Observation 1&2 (Supplementary Material Sec. A) and their brief derivations in the main paper (Sec. 3.2), while retaining the extended math in the Supplement.

• Since the construction of the datasets is one of our core contributions, we will migrate more dataset construction details (Supplementary Material Sec. B) to the main paper (Sec. 5.1), including the camera parameters and scene setup details.

• We will move the ablation studies (Supplementary Material Sec. D and Table 5) to the main paper as Sec. 5.5, following Sec.5.4.

We will condense our texts to make sure the final version of our main paper within page limit, concise, and more inclusive.

W3: The experimental evaluation lacks comparisons with state-of-the-art approaches from the past two years.

We respectfully clarify that Depth Anything V2 [1] and Metric3D V2 [2] are both state-of-the-art depth estimation methods published in the year 2024. In addition, we have now incorporated a new method, namely Depth Any Camera (DAC) [3], published in 2025 to further extend the scope of compared methods. Here we finetune DAC using its pretrained outdoor ResNet101 checkpoint. Please refer to the following tables for our comparison, where we denote DFF methods as "DFF" and monocluar methods as "Mono". We will update the Table 1 and 2 in our final version accordingly. Note that DAC may not perform well because its main focus is wide-angle cameras.

Table R1: Quantitative comparisons of in-domain metric depth estimation.

Blender-Syn

Sintel-Dr. Bokeh

Table R2: Quantitative comparisons of zero-shot metric depth estimation.

Blender-Syn

Sintel-Dr. Bokeh

W4: No quantitative experiments are conducted on 4DLFD-Semi-Real and EDFV-Real datasets.

We now provide the quantitative results on the 4DLFD-Semi-Real dataset in the following table. Here we convert all disparity predictions into depth values with unified curves for fair comparisons. It can be found that our method outperforms others in all metrics.

Table R3: Quantitative comparisons of zero-shot metric depth estimation on 4DLFD-Semi-Real.

For EDFV-Real, due to the absence of a depth sensor in our current system setup, we were unable to collect ground-truth depth maps for quantitative evaluation. Please refer to Fig. 5 of our paper and Fig. 12 of Supplementary Material for qualitative comparisons. In the future work, we aim to enhance our hardware system by integrating a depth sensor for accurate ground-truth depth capture, enabling quantitative analysis.

W5: Some typo: L197, "we two".

Thank you for spotting this. Here we have left an "employ", which should be "we employ two U-Net-based networks". We will thoroughly revise the manuscript and eliminate all such typographical errors in our final version.

Q1&Q2: The authors should provide further analysis to demonstrate the superiority of using event cameras over traditional depth-from-focus methods. Based on the above question, if we replace the sparse depth generation step with normal depth-from-focus approaches, and keep the subsequent steps (MCPM part) unchanged, how would the performance be affected?

We appreciate this insightful question. We respectfully demonstrate the necessity of introducing events in the following three aspects:

1. Our current setup is just a camera prototype, and the rotation speed is limited by mechanical or manual speed. However, as far as we know, some industrial camera modules could reduce the focus ring rotation time to about 10ms such as the liquid lens, which would be hard for standard cameras to capture focal stacks. However, event cameras can easily handle the scenario with a microsecond-level temporal resolution.

2. As discussed in L171-174 of our paper, the performance of traditional DFF methods could be influenced by the discrete frame number in focal stacks. However, with events, we can capture the focal stack with a microsecond-level temporal resolution, which can provide nearly continuous transitions between in-focus and out-of-focus regions, and help to improve the accuracy of depth estimation. In addition, although events may be triggered less in smooth or textureless regions, traditional DFF methods also struggle in such areas [4] as they depend on textures to estimate the in-focus degree. Therefore, even if replacing the sparse depth generation step with normal DFF approaches, the performance may not increase.

3. As events only record intensity changes, the total raw data size with resolution $720\times 1280$ during one capture is averagely 4-5MB. However, a single image in the focal stack with the same resolution could take up to 2.7M, with 30 images up to 80MB. This means that the storage space required for event-based depth estimation is much smaller than that of traditional DFF methods, which make it more practical to deploy on resource-limited devices.

Q3: It is suggested to indicate the type of compared method (such as DFF and Mono) in Table 1 and 2 for better clarity.

Thank you for your suggestions. We have provided the updated version of Table 1 and 2 in the tables above, and will also include them in our final version.

[1] Yang et al. Depth Anything V2. NeurIPS 2024.

[2] Hu et al. Metric3d V2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. TPAMI 2024.

[3] Guo et al. Depth Any Camera: Zero-Shot Metric Depth Estimation from Any Camera. CVPR 2025.

[4] Yang et al. Deep Depth from Focus with Differential Focus Volume. CVPR 2022.

Thank you for your responses! We have quickly reviewed related literatures in the past two years and now added a new DFF method for comparison, namely HybridDepth [1], which is a recent method published in WACV 2025. This method incorporates both a pretrained traditional DFF method (DFF-DFV in their implementation) and a pretrained image foundation model (DA in their implementation) as a relative depth estimator. Although taking information from both sources, it may still face the challenge of discrete frame number and acquisition time discussed in our paper as they need image focal stacks as input. We have now fintuned it on our two synthetic datasets from its pretrained NYU checkpoint and provide the in-domain quantitative results in Table R1 below. HybridDepth generates its output by multiplying the estimation results of DA with a predicted scale map, which may be the reason why it doesn't preform well on Sintel-Dr.bokeh dataset with large scale variations across different scenes. In addition, we have also noticed a CVPR 2025 paper [2] that utilizes defocus clues for sparse and dense depth estimation. However, the code for dense depth estimation is not available yet, and we will try to include it for comparison once it is available.

Besides, in order to demonstrate our second point in Q1&Q2, we further conduct an ablation study, where we have replaced our sparse depth generation network with an existing DFF method (here we use DFF-DFV) and keep the rest of our framework unchanged. For fair comparison, we employ 32 frames as the input focal stack, and use the same random initialization and the same training scheme on this network (denoted as Ours#). Here we also report the results of this network in the Table R1 below, and provide the computational efficiency comparison in Table R2. It can be found that although Ours# brings minor improvements to some metrics compared with Ours (such as 0.001 in RMSE and 0.001 in $\delta_1$), it brings much more computational complexity (3.42 times parameters and 8.74 times MACs) to the whole pipeline, which is not desirable for resource-limited devices. Instead, our event-based method can achieve a balance between accuracy and efficiency, and is more practical to deploy on resource-limited devices.

Thank you again for your thoughtful comments and the continued discussion, which helps us further improve our work.

Table R1: Quantitative comparisons of in-domain metric depth estimation.

Blender-Syn

Sintel-Dr. Bokeh

Table R2: Computational efficiency comparison.

[1] Ganj et al. HybridDepth: Robust Depth Fusion by Leveraging Depth from Focus and Single-Image Priors. WACV 2025.

[2] Xu et al. Blurry-Edges: Photon-Limited Depth Estimation from Defocused Boundaries. CVPR 2025.

We sincerely thank you again for your suggestions and further discussions! Now we have completed our training and evaluation of HybridDepth on Sintel-Dr. Bokeh dataset, and the results can be found in the updated Table R1 of of our last Offical Comment for you. It can be found that our method still outperforms HybridDepth in all metrics. HybridDepth generates its output by multiplying the estimation results of DA with a predicted scale map, which may be the reason why it doesn't preform well on Sintel-Dr. Bokeh dataset with large scale variations across different scenes. We will add these comparisons to our final version if accepted. Please let us know if there are more comments and we will be more than happy to follow up.

We sincerely thank you for your insightful comments and constructive suggestions! We are encouraged that you find our method well-motivated and carefully designed, our paper well-organized, and our concept novel. Below, we provide detailed responses to your proposed Weaknesses (W), Questions (Q), and Limitations (L).

W1.1: Some important details are left to the Supplementary Material.

Thank you for your suggestions! Due to the page limit of the submitted version for peer review, we have to put some details in the Supplementary Material. However, we understand the importance of these components in demonstrating the validity and interpretability of our proposed approach, and will include the important details in the final version of our main paper if accepted as follows:

• We will put Additional Observation 1&2 (Supplementary Material Sec. A) and their brief derivations in the main paper (Sec. 3.2), while retaining the extended math in the Supplement.

• Since the construction of the datasets is one of our core contributions, we will migrate more dataset construction details (Supplementary Material Sec. B) to the main paper (Sec. 5.1), including the camera parameters and scene setup details.

• We will move the ablation studies (Supplementary Material Sec. D and Table 5) to the main paper as Sec. 5.5, following Sec.5.4.

We will condense our texts to make sure the final version of our main paper within page limit, concise, and more inclusive.

W1.2: The limitations section is quite short. A more systematic reflection on assumptions (e.g., sensitivity to event noise, synchronization of sensors) would improve transparency.

We appreciate this suggestion. While our method has shown robustness in various scenarios, we acknowledge its performance may degrade under extreme conditions. Specifically:

• In low light scenarios, high sensor noise can reduce the quality of EDFV and predicted sparse depth, thus causing the performance of our method to degrade.
• When applied to high-speed scenarios, the synchronization of sensors could be an important issue as we need the frame and events to capture the scene with similar timing. But there are some approaches such as the chip synchronization in DAVIS346 [1] and external clock triggering mechanism [2] to limit the time error within 10ms. Besides, a specifically designed electronic system [6] could reach the time precision of 0.5ms. These hardware solutions could enable our system to achieve high-precision synchronization.

We will add these discussions in the final version of our paper, and try to improve the robustness of our method in these challenging scenarios in our future work.

W2: While promising, the real-world applicability is still partially constrained by the hardware complexity of event-RGB hybrid systems and the assumption of focus sweeping capability. These aspects are not discussed in depth.

Thank you for highlighting these practical considerations.

• For the hardware complexity, our current setup is just a camera prototype, and the hybrid system could be further integrated by engineering approches to reduce the complexity. In addition, as far as we know, some industrial camera modules could reduce the focus ring rotation time to about 10ms such as the liquid lens, which could further improve the real-time efficiency of our system.
• For the assumption of focus sweeping capability, as indicated in the mathematical derivations in Sec. 3.2 and Supplementary Material Sec. A, the effectiveness of events mainly locates at edges and textures. Therefore, for textureless regions or areas with low event density, our estimation relies more heavily on the initial dense depth and sparse depth around textured regions. This may cause our method to perform suboptimally in highly smooth scenarios. Notably, such scenarios also remain open challenges for conventional DFF approaches [3], as they depend on textures to estimate the in-focus degree.

We will add these discussions in the final version of our paper, and work on improving the robustness of our method in these scenarios in future work.

W3: The idea of prompting IFMs with sparse depth is increasingly common (e.g., in LiDAR prompting literature). The prompting framework (EDPN) may not be as novel as the EDFV component.

Thank you for this thoughtful comment. We agree that the general idea of prompting image foundation models (IFMs) with sparse depth has gained increasing attention, such as in the context of LiDAR [4,5]. However, although the idea is common, the strategy may differ. Therefore, we respectfully argue that our approach differs from previous works in motivation and implementation. It should be noted that the novelty of our work should be viewed from the holistic perspective of the proposed framework, compromising the EDFV construction, sparse depth & initial dense depth estimation, and prompting network (EDPN) (please refer to the proposed second contribution in our paper L63-65). Each component in the pipeline is essential to the overall method (as discussed in Ablation studies in Supplementary Material Sec. D). While a single component may seem straightforward, our combined pipeline introduces a cohesive and tailored framework that enables effective prompting in a challenging hybrid modality.

As discussed in the Introduction (L48-50), the information given by event-based sensors is different from other sensors such as LiDAR, as it mainly triggers at texture and edge regions. This poses unique challenges and opportunities for our design.

In our pipeline, we first utilize proposed EDFV to process events into depth map. After that, we design a SAME strategy to extract effective information from events (L179-181 of our paper). Besides, considering the modality discrepancies between events and image, we employ different layers to capture their features respectively (L199-201 of our paper). We also introduce deformable convolutions to handle the sparsity and possible misalignments between two modalities (L207-209 of our paper). We will revise the paper to clarify this contribution and better differentiate it from related works.

Q1&L1: The method relies on event triggers from intensity changes, which may be sparse or absent in textureless or poorly lit regions. Could the authors show or describe failure cases where EDFV performs poorly? Do you apply any fallback mechanism or fusion weighting in such cases?

Please refer to our rebuttal in W1.2 and W2 for some challenging scenarios where our method may perform suboptimally. As we have considered the feature of sparse events in our design of prompting network, we did not incorporate a fallback mechanism in our method. The deformable convolutions and cross attention mechanism in our pipeline could help to handle the textureless regions. However, we agree that an additional fallback mechanism or fusion weighting may further improve the performance in extremely challenging conditions. These improvements are beyond our current scope, and will be considered as our future work.

Q2&L2: The proposed system requires precise synchronization between RGB and event cameras and a controllable focus mechanism. These requirements limit real-world usability. How to ensure temporal synchronization between the event camera and RGB sensor, especially during fast focus sweeping?

If applied to static or slow motion scenarios, the synchronization between the event camera and RGB sensor may be less crucial, as they record nearly the same scene over short durations. But if applied to high-speed scenarios, tighter synchronization could be required. As mentioned in W1.2, hardware-based solutions such as externally triggered clocking mechanisms [2] are for maintaining reliable timing within temporal tolerances.

[1] Zhu et al. The Multi Vehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. RAL 2018.

[2] Duan et al. EventAid: Benchmarking Event-aided Image/Video Enhancement Algorithms with Real-captured Hybrid Dataset. TPAMI 2025.

[3] Yang et al. Deep Depth from Focus with Differential Focus Volume. CVPR 2022.

[4] Park et al. Depth Prompting for Sensor-Agnostic Depth Estimation. CVPR 2024.

[5] Lin et al. Prompting Depth Anything for 4K Resolution Accurate Metric Depth Estimation. CVPR 2025.

[6] Zou et al. Learning to Reconstruct High Speed and High Dynamic Range Videos from Events. CVPR 2021.

We sincerely thank you for your positive feedback and for confirming that your concerns have been addressed! We appreciate your constructive suggestions, which have helped us enhance the quality and clarity of our manuscript. We will ensure all promised revisions are incorporated into the final version if accepted.

The authors have addressed my concerns.