HQGS: High-Quality Novel View Synthesis with Gaussian Splatting in Degraded Scenes

We thank reviewer enTq for acknowledging the contribution of our paper and providing thoughtful comments.

Q1. Explanation of the COLMAP Section.

We follow the setup in 3DGS [1], which states, "the input to our method is a set of images of a static scene, together with the corresponding cameras calibrated by SfM （COLMAP）, which produces a sparse point cloud as a side effect. From these points, we create a set of 3D Gaussians." In our case, we replace the clean images with low-quality images to generate the corresponding point cloud by COLMAP. We make it clear in Sec. 3.1 of the revised manuscript.

[1] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. 2023: 1–14.

Q2, 6, 8. Explanation of M and ESFG.

We clarify in the revised manuscript that M represents the number of Gaussian primitives. As you mentioned, with the split-and-clone process, the value of M dynamically changes.

The ESFG output is a fused feature that contains global information of the scene. To handle the changes in M, we fix the scale of the fused feature and replicate it according to the varying number of Gaussian primitives (M). This approach ensures that the fused feature consistently retains global information while avoiding the time costs of adjusting dimensions through learnable parameters.

Q3. Explanation of the Downsampling Parameters in the ESFG Module.

As mentioned in line 263 in the original manuscript, we introduce downsampling in the ESFG module to reduce dimensionality and save computational resources. To further investigate its impact, we supplement additional experiments under two configurations: Without Downsampling and 4 $\times$ Downsampling. Last, we choose the 2 $\times$ to balance the training time and the performance.

Q4. Some Typos.

Thanks for your advice! The paper has been revised to enhance the writing quality.

Q5. Are there layers after the fusion features but before the sigmoid?

There are no additional learnable parameters between the fusion features and the sigmoid activation. All learnable parameters are placed prior to the fusion feature generation process.

Q7. Ablation Studies on the Proposed Loss.

As you suggested, we add the ablation studies on D-SSIM and our 𝓛₍SCS₎. In the following Table, our 𝓛₍SCS₎ have better results. For 3D reconstruction in degraded scenes, our framework first leverages edge maps to enhance the perception of high-frequency object edges. Simultaneously, the proposed 𝓛₍SCS₎ focuses more on the low-frequency global structure, forming a complementary relationship guided by edges for improved results. In contrast, D-SSIM is not specifically designed as a loss function for low-frequency components. And it does not effectively complement our high-frequency edge perception strategy, resulting in relatively limited performance.

Dear Reviewer enTq,

We deeply appreciate your valuable feedback during the first round of review and the thoughtful discussion that has significantly helped us refine our work. Since the discussion phase ends on Nov 26, we would like to know whether we have addressed all the issues, and we would greatly welcome any additional feedback or suggestions you may have.

Thank you again for your devotion to the review. If all the concerns have been successfully addressed, please consider raising the scores after this discussion phase.

Best regards,

Paper4233 Authors

Thank you for your thoughtful feedback and for taking the time to review our rebuttal. Both high-quality and low-quality images need to be fed into COLMAP to obtain estimated poses and initialized point clouds for 3DGS-based methods. As you suggested, we add a new visualisation of the poses for both high- and low-quality scenes in Figure 4 of the submitted Supplementary Material. These are two scenes from DeblurNeRF and LLFF, respectively, where the red cones represent the cameras. Since the scenes in DeblurNeRF are smaller, the cameras are also smaller, while those in LLFF are larger. We can observe that in low-quality scenarios, the camera's perspective shifts, and the decline in image quality affects the accuracy of pose estimation in COLMAP, resulting in poses that differ from those derived from high-quality images. This has been analyzed in other NeRF-based papers [1], which focus on posing issues to address blurry scenes.

However, our paper, along with some contemporary 3DGS-based studies [2,3,4], emphasises the sparsity of the point cloud, which is one of the key motivations behind our proposed method. Low-quality scenes significantly contribute to the sparsity of the initialized point cloud. Due to the limited information in low-quality images, the initialized point cloud contains significantly fewer points and is much sparser, as shown in Figure 2(a) of our previous and revised manuscript. Additionally, we have provided the number of points in the initialized point clouds in Figure 4 of the submitted Supplementary Material. The points in the point cloud decrease from 32,338 to 11,021 and from 6,174 to 3,348 in high-quality to low-quality scenes across the two datasets.

Nonetheless, all comparison models are retrained in the same setting, ensuring the fairness of the comparisons.

[1] Li Ma, Xiaoyu Li, Jing Liao, Jue Wang, Qi Zhang, Xuan Wang, Pedro V. Sander. Deblur-nerf: Neural radiance fields from blurry images. CVPR. 2022: 12861-12870.

[2] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Wen Li, Yifei Chen, Zhenzhong Kuang, Jiajun Ding, Jianping Fan, and Jun Yu. Srgs: Super-resolution 3D Gaussian Splatting. arXiv:2404.10318, 2024

[3] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, and Eunbyung Park. Deblurring 3d gaussian splatting. European Conference on Computer Vision. Springer, Cham, 2025: 127-143.

[4] Shiyun Xie, Zhiru Wang, Yinghao Zhu, and Chengwei Pan. SuperGS: Super-Resolution 3D Gaussian Splatting via Latent Feature Field and Gradient-guided Splitting. arXiv preprint arXiv:2410.02571, 2024.

We thank reviewer s4XL for acknowledging the contribution of our paper and providing thoughtful comments.

Q1, 3, 4, 5, 6, 7, 8. Improving and Rewriting Some Sentences.

Thanks for your valuable suggestions! We have made several improvements to the writing quality in the revised manuscript. The key changes include:

1. The introduction of ESFG in the abstract has been revised: "The fused features serve as prior guidance to capture detailed distribution across different regions, bringing more attention to areas with detailed edge information and allowing for a higher concentration of Gaussian primitives to be assigned to such areas."
2. Figure 4 has been updated, and its caption revised to: "The ESFG module separately learns semantic-aware feature and edge-aware feature, and then jointly guides the training of HQGS."
3. The explanation of the notation for matrix multiplication in Equation 2 has been improved for clarity.
4. Several sentences have been enhanced for better clarity and readability.
5. Figures 7 and 8 from the original manuscript have been modified: Figure 7 has been expanded with additional comparison algorithms, and it is now presented in table format as Table 6 in the revised manuscript. In Figure 8, we have redrawn the figure with more distinguishable colors and adjusted the fonts to eliminate overlapping issues.

Q2. Further Explanation on the Distortion in Line 48.

As we initially explore the common effects of various degradation scenarios on the 3DGS model, the experiments involve multiple scenarios. Thus, the term "distortion" here refers to the impact across all these blur, noise, low resolution, and compression scenarios. In the revised manuscript, we add further clarification on this point.

Q9. Regarding selecting JPEG Compression Quality Parameters and Additional Experiments under Different Settings.

We follow several image restoration methods [1, 2, 3], where the JPEG compression quality factors are typically chosen within the range of 10–40. For our experiments, we select a factor of 10, representing the most severe scenario. Additionally, as you suggested, we add experiments under other quality factors, as well as tests for 2 $\times$ and 8 $\times$ downsampling scenarios.

[1] Li Y, Fan Y, Xiang X, et al. Efficient and explicit modelling of image hierarchies for image restoration. CVPR. 2023: 18278-18289.

[2] Simon Welker, Henry N. Chapman, Timo Gerkmann. DriftRec: Adapting diffusion models to blind JPEG restoration. IEEE Transactions on Image Processing, 2024: 2795-2807.

[3] Li B, Li X, Lu Y, et al. PromptCIR: Blind Compressed Image Restoration with Prompt Learning[J]. arXiv:2404.17433, 2024.

Dear Reviewer s4XL,

We deeply appreciate your valuable feedback during the first round of review and the thoughtful discussion that has significantly helped us refine our work. Since the discussion phase ends on Nov 26, we would like to know whether we have addressed all the issues, and we would greatly welcome any additional feedback or suggestions you may have.

Thank you again for your devotion to the review. If all the concerns have been successfully addressed, please consider raising the scores after this discussion phase.

Best regards,

Paper4233 Authors

Thank you for your insightful comments and appreciation of our work and rebuttal. It is good to see that our comments could address your concerns. We will do our best to improve the final version of our paper based on your valuable suggestions.

We thank reviewer cE6C for acknowledging the contribution of our paper and providing thoughtful comments.

Q1. Lack of Robustness Experiments on Edge Maps

We leverage the well-known fact that edge maps have sufficient frequency information [1] and can be obtained by an edge detection operator even from degraded images. As suggested, we visualize edge detection results obtained from the Sobel operator under progressively challenging conditions in Figure 1 of the submitted Supplementary Material. Additionally, we compute the PSNR values based on images and edge maps under various conditions with respect to clean ones, then present the PSNR variance of the images and edge maps: 10.18/0.285. These results further demonstrate that the edge maps are more robust under progressively degraded conditions. Thus, edge detection performs well in extracting edge information to a considerable extent, even under relatively challenging conditions.

In our experimental setups, we follow the degradation settings commonly used in image restoration tasks [2, 3, 4, 5], ensuring that the chosen degradation ranges align with existing works (4$\times$ downsampling, JPEG with quality level 10, and so on). Under these settings, our method consistently performs favorably against existing algorithms. Furthermore, we discuss the progressively challenging scenarios in Figure 7 of the original manuscript, showing that our method performs better than other approaches under severe conditions (e.g., noise level of 50 or 8 $\times$ downsampling). We include comparisons with more methods in Table 6 of the revised manuscript.

[1] Lindeberg T. Scale space. Encyclopedia of Computer Science and Engineering. 2009: 2495–2504.

[2] Li Y, Fan Y, Xiang X, et al. Efficient and explicit modelling of image hierarchies for image restoration. CVPR. 2023: 18278-18289.

[3] Ren B, Li Y, Mehta N, et al. The ninth NTIRE 2024 efficient super-resolution challenge report. CVPR. 2024: 6595-6631.

[4] Lu Z, Li J, Liu H, et al. Transformer for single image super-resolution. CVPR. 2022: 457-466.

[5] El Helou M, Süsstrunk S. Blind universal Bayesian image denoising with Gaussian noise level learning. TIP. 2020, 29: 4885-4897.

Q2. Training and inference time.

In 3D reconstruction, rendering time corresponds to inference time. As shown in Table 2 of the original manuscript (Table 1 of the revised manuscript), we report the rendering times of various methods. Our HQGS achieves comparable rendering times to other 3DGS-based methods and is significantly faster than NeRF-based methods.

Regarding training time, Figure 8 in the original manuscript compares existing methods under the same training time for fairness. HQGS demonstrates a faster convergence rate and consistently performs well against other methods, highlighting its training efficiency. Whether compared under the same training time or at convergence with the same number of iterations, our method shows better performance. To enhance clarity, we have updated the subsection title from "ANALYSIS ON RECONSTRUCTION TIME OF SOME 3DGS-BASED METHODS." to "TRAINING TIME VS QUALITY." and improved the Figure's readability in the revised manuscript to eliminate overlapping issues.

Q3. Some Figures and Tables Need to be Improved.

Thank you for your valuable feedback! We have revised the manuscript to improve the quality of figures and tables, as detailed below:

1. The colors in Figure 1 have been adjusted to enhance distinguishability and readability.

2. For the order of the Figures, Figure 1 is referenced on line 42, and Figure 2 on line 48 in the original manuscript.

3. The formatting of Tables 1 and 2 in the original manuscript has been updated for consistency and improved readability.

4. Regarding the visualized input views, since rendering unseen views in 3D reconstruction only requires the given camera viewpoint information and not the input images. As suggested, the clean ground truth used for PSNR calculation is provided in Figure 5 of the revised manuscript.

Q4. The Settings of Experiments in Table 3 and Table 4.

Thank you for the reminder. Both Table 3 and Table 4 analyze the 'Wine' scene with blurry degradation from the DeblurNeRF dataset. We have updated their titles in the revised manuscript to reflect this context.

Q5. Performance and Explanation Under Challenging Conditions.

Thanks for your reminder and suggestions. In Figure 2(a) of the original manuscript, we illustrate the task settings by generating low-quality images with various degradations and the corresponding initialized point clouds. These examples are solely for demonstration purposes and are not used for network training.

In Figure 2(b), we visualize and render 2D images of Gaussian primitives after training on datasets constructed based on degradation settings from prior image restoration works [2, 3, 4, 5], as detailed in Section 3.1. All methods are trained under these settings for fair comparison. Our method significantly enhances rendering quality for unseen views within a certain range and maintains better performance and robustness compared to the baseline and other methods, even in challenging conditions (Figure 7, original manuscript).

For extremely adverse scenarios with minimal valid information—where existing image restoration algorithms fail to recover such cases. Addressing the reconstruction of unseen views under such extreme conditions remains an open challenge for future exploration.

Q6. Training Iterations/Time and Parameters Settings.

Thanks for your reminder. In Section 4.1 of the revised manuscript, we clarify that all 3DGS-based methods (3DGS, SRGS, and HQGS) are trained under the same settings for 50k iterations to ensure fairness. Performance comparisons based on the same number of iterations are provided in Tables 1 and 2 of the original manuscript, and comparisons based on the same training time are shown in Figure 8 of the original manuscript.

Specifically, in 9 minutes, HQGS completes 35k iterations, while 3DGS completes 50k iterations, with both methods approaching near convergence. The training speeds for 3DGS, SRGS, and HQGS are 0.535 dB/kiters, 0.550 dB/kiters, and 0.824 dB/kiters, respectively. In summary, although our model has more parameters, it converges more easily and learns faster.

The densification parameters are applied with a fixed number of iterations and are not learnable. For a fair comparison, we adopt the same parameters as the 3DGS pipeline.

Q7. The Weights of MLP.

The MLP is optimized individually for each scene, as per the design established by the 3DGS pipeline [6, 7]. This means that each scene requires separate training from scratch. All comparison methods, whether they involve MLPs (such as NeRF, which uses multiple MLPs) or not, follow the same optimization approach, training each scene individually.

[6] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Trans.Graph, 2023, 1–14.

[7] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Wen Li, Yifei Chen, Zhenzhong Kuang, Jiajun Ding, Jianping Fan, and Jun Yu. Srgs: Super-resolution 3D Gaussian Splatting. arXiv:2404.10318, 2024

Dear Reviewer cE6C,

We deeply appreciate your valuable feedback during the first round of review and the thoughtful discussion that has significantly helped us refine our work. Since the discussion phase ends on Nov 26, we would like to know whether we have addressed all the issues, and we would greatly welcome any additional feedback or suggestions you may have.

Thank you again for your devotion to the review. If all the concerns have been successfully addressed, please consider raising the scores after this discussion phase.

Best regards,

Paper4233 Authors

Dear Reviewer cE6C,

We sincerely thank you for your valuable feedback during the first round of review and for the thoughtful discussions that have greatly contributed to improving our work. Your insights and suggestions have been instrumental in refining our submission, and we are deeply grateful for your time and effort.

We kindly wish to confirm whether we have satisfactorily addressed all your concerns. Thank you again for your devotion to the review. If all the concerns have been successfully addressed, please consider raising the scores after this discussion phase.

Best regards,

Paper4233 Authors

Dear reviewer cE6C,

Thanks for the comments and review. We have provided more explanations and answers to your questions. Since the deadline for discussion is near the end, please let us know whether we have answered all the questions. Please also consider raising the score after all issues are addressed.

If you have more questions, please raise them and we will reply ASAP.

Thanks,

Paper4233 Authors

We thank reviewer rTVv for acknowledging the contribution of our paper and providing thoughtful comments.

Q1. Lack of Robustness Experiments on Edge Maps

We leverage the well-known fact that edge maps have sufficient frequency information [1] and can be obtained by an edge detection operator even from degraded images. As suggested, we visualize edge detection results obtained from the Sobel operator under progressively challenging conditions in Figure 1 of the submitted Supplementary Material. Additionally, we compute the PSNR values based on images and edge maps under various conditions with respect to clean ones, then present the PSNR variance of the images and edge maps: 10.18/0.285. These results further demonstrate that the edge maps are more robust under progressively degraded conditions. Thus, edge detection performs well in extracting edge information to a considerable extent, even under relatively challenging conditions.

In our experimental setups, we follow the degradation settings commonly used in image restoration tasks [2, 3, 4, 5], ensuring that the chosen degradation ranges align with existing works (4$\times$ downsampling, JPEG with quality level 10, and so on). Under these settings, our method consistently performs favorably against existing algorithms. Furthermore, we discuss the progressively challenging scenarios in Figure 7 of the original manuscript, showing that our method performs better than other approaches under severe conditions (e.g., noise level of 50 or 8 $\times$ downsampling). We include comparisons with more methods in Table 6 of the revised manuscript.

[1] Lindeberg T. Scale space. Encyclopedia of Computer Science and Engineering. 2009: 2495–2504.

[2] Li Y, Fan Y, Xiang X, et al. Efficient and explicit modelling of image hierarchies for image restoration. CVPR. 2023: 18278-18289.

[3] Ren B, Li Y, Mehta N, et al. The ninth NTIRE 2024 efficient super-resolution challenge report. CVPR. 2024: 6595-6631.

[4] Lu Z, Li J, Liu H, et al. Transformer for single image super-resolution. CVPR. 2022: 457-466.

[5] El Helou M, Süsstrunk S. Blind universal Bayesian image denoising with Gaussian noise level learning. TIP. 2020, 29: 4885-4897.

Q2. The Impact of ESFG on Gaussian Primitives Density.

Figure 2(b) in the original manuscript demonstrates the centers of Gaussian primitives from the trained model, which are also the points in the point cloud. Compared to 3DGS, our method produces a denser distribution of Gaussian primitives, enabling enhanced coverage of finer details and textures rather than only redistributing the primitives.

Q3. Visualizations of Gaussian Primitives in high- and low-frequency regions.

The proposed ESFG is designed to guide the model in focusing on detailed regions and small objects, as validated by Figure 2(b) in the original manuscript. Following your suggestion, we provide additional evidence in Figure 2 of the submitted Supplementary Material.

Specifically, we illustrate the Gaussian primitives in both low- and high-frequency regions, alongside their corresponding rendering results. To highlight these regions, we adjust the point cloud's angle, select optimal viewpoints, and include enlarged screenshots. Our method demonstrates better visual quality, particularly in areas like the floor and windows.

Additionally, we show the difference map between rendered and clean images in Figure 3 of the Supplementary Material (Figure 6 in the revised manuscript). These results show that 3DGS and other methods produce significant inaccuracies in low- and high-frequency regions (highlighted as bright areas in the difference maps). In contrast, our method achieves smaller differences, demonstrating well performance in these challenging regions.

Q4. Robustness Validation on all methods.

As suggested, we provide the results of other methods in the following Table and Table 5 of the revised manuscript. Our method performs favorably against others and exhibits stronger generalization ability under challenging conditions.

Q5. Training Time vs Performance.

Figure 8 in the original manuscript compares existing methods under the same training time for fairness. HQGS demonstrates a faster convergence rate and consistently performs well against other methods, highlighting its training efficiency. Whether compared under the same training time or at convergence with the same number of iterations, our method shows better performance.

To improve clarity, we have updated the subsection title from "ANALYSIS ON RECONSTRUCTION TIME OF SOME 3DGS-BASED METHODS." to "TRAINING TIME VS QUALITY." We have also improved the readability of Figure 8 in the revised manuscript to eliminate overlapping issues.

Q6. The Decline in Edge Detection Performance and Its Impact After Applying Image Restoration Processing.

Figure 7 of the original manuscript validates this claim: under challenging conditions (e.g., noise level 50 and 8 $\times$ downsampling), edge detection performance declines compared to the clean case, leading to a drop in the quality of rendered unseen views. Despite this, our method still performs better, achieving a 3.27 dB improvement over 3DGS at a noise level 50.

Additionally, as shown in the Table in Q4, our HQGS remains superior even in clean scenes (noise = 0, 1 $\times$ resolution), which can be considered the upper limit after image restoration. These results demonstrate that our method is effective not only under challenging conditions but also in clean settings, excelling in capturing small objects and enhancing global learning in low-frequency regions.

Q7. Can the ESFG module improve the density of the point cloud while maintaining the total number of Gaussian elements? Is there a densification strategy or explanation of how the ESFG module affects 3DGS densification to better handle the sparse point clouds generated by low-quality images?

In 3DGS, once training concludes, the points in the point cloud correspond directly to the centers of Gaussian primitives [6, 7], meaning their quantities are fixed and equal, cannot be independently adjusted.

Regarding density, the original densification strategy in 3DGS duplicates points based on factors like positional importance or the presence of details. However, 3DGS exhibits limited sensitivity to fine details, leading to missing object features (e.g., the "wires" in Figure 2(b) of the original manuscript). In contrast, our ESFG emphasizes finer details and provides this information to the model, resulting in a higher density of Gaussian primitives, specifically in detailed regions, and effectively capturing these intricate features.

[6] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Trans.Graph, 2023: 1–14.

[7] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Wen Li, Yifei Chen, Zhenzhong Kuang, Jiajun Ding, Jianping Fan, and Jun Yu. SRGS: Super-resolution 3D Gaussian Splatting. arXiv:2404.10318, 2024

Dear Reviewer rTVv,

We deeply appreciate your valuable feedback during the first round of review and the thoughtful discussion that has significantly helped us refine our work. Since the discussion phase ends on Nov 26, we would like to know whether we have addressed all the issues, and we would greatly welcome any additional feedback or suggestions you may have.

Thank you again for your devotion to the review. If all the concerns have been successfully addressed, please consider raising the scores after this discussion phase.

Best regards,

Paper4233 Authors

Dear Reviewer rTVv,

We sincerely thank you for your valuable feedback during the first round of review and for the thoughtful discussions that have greatly contributed to improving our work. Your insights and suggestions have been instrumental in refining our submission, and we are deeply grateful for your time and effort.

We kindly wish to confirm whether we have satisfactorily addressed all your concerns. Thank you again for your devotion to the review. If all the concerns have been successfully addressed, please consider raising the scores after this discussion phase.

Best regards,

Paper4233 Authors

Dear Authors,

Thank you for your detailed and thorough response. I appreciate the effort you put into addressing the concerns raised during the review process, as well as for providing additional data and visualizations.

However, I still have some unresolved questions regarding the ESFG module and its impact on the densification of Gaussian primitives in HQGS. Specifically, I have observed that HQGS consistently produces more Gaussian primitives compared to 3DGS, both in high-frequency regions and in low-frequency areas. This observation raises the following points:

1. Are you using the same densification strategy as the original 3DGS method? If so, what mechanism within the ESFG module leads to the observed increase in Gaussian primitives across both high- and low-frequency regions?
2. If the densification frequency or strategy has been intentionally adjusted in HQGS, could you provide more details on how it differs from the original 3DGS approach? Specifically, has the frequency of densification been increased, or has a new mechanism been introduced to achieve this result?

Clarifying these points would provide valuable insights into how HQGS operates and the role of ESFG in improving reconstruction quality. I look forward to your response.

Reviewer rTVv

Thank you for your thoughtful feedback and for taking the time to review our rebuttal. We will address your points regarding the proposed modules below and include these in the revised version to improve its readability.

1. Are you using the same densification strategy as the original 3DGS method? If so, what mechanism within the ESFG module leads to the observed increase in Gaussian primitives across both high- and low-frequency regions?

We adopt the same densification strategy as the other 3DGS-based methods [1,2], performing a densification operation every 100 iterations to expand the point cloud density in important regions. The observed increase in Gaussian primitives is due to the introduction of additional information that guides the model to areas where point cloud expansion is necessary.

For the ESFG module, we aim to improve attention to these areas by incorporating a high-frequency map. The high-frequency branch contains only high-frequency information, which has superior perceptual ability for small target regions. Additionally, we introduce the original low-quality images with strong semantic information, which include both high- and low-frequency area information and serve as supplementary input. Furthermore, the proposed SCS loss function can also be used to supplement low-frequency information.

Our ablation studies in Tables 3 and 4 of the revised paper validate these previous points. Moreover, as seen in Figure 2 of the supplementary materials, the distribution of Gaussian primitives indicates that our method generates more Gaussian primitives in both high-frequency regions (e.g., the dinosaur) and low-frequency regions (e.g., the floor), compared to 3DGS. Due to the ESFG module, the number of Gaussian primitives in high-frequency regions increases significantly.

2. If the densification frequency or strategy has been intentionally adjusted in HQGS, could you provide more details on how it differs from the original 3DGS approach? Specifically, has the frequency of densification been increased, or has a new mechanism been introduced to achieve this result?

We do not increase the densification frequency; all of our parameter settings are consistent with other 3DGS-based methods [1,2] to ensure fairness. Densification occurs once every 100 iterations. We believe the increase in the number of Gaussian primitives is primarily due to the enhanced attention information. Since the densification operation is performed based on the importance of the information, it only occurs in regions where the model determines that higher distinction is needed, thereby improving the accuracy of the model's rendering results.

The ESFG module provides sufficient guidance (as mentioned in Question 1), directing the model to focus on specific regions. This helps ensure that more Gaussian primitives are allocated to fit the ground truth, thereby reducing the loss.

[1] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Trans.Graph, 2023: 1–14.

[2] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Wen Li, Yifei Chen, Zhenzhong Kuang, Jiajun Ding, Jianping Fan, and Jun Yu. SRGS: Super-resolution 3D Gaussian Splatting. arXiv:2404.10318, 2024

Dear reviewer rTVv,

Thanks for the comments and review. We have provided more explanations and answers to your questions. Since the deadline for discussion is near the end, please let us know whether we have answered all the questions. Please also consider raising the score after all issues are addressed.

If you have more questions, please raise them and we will reply ASAP.

Thanks,

Paper4233 Authors

Dear Reviewer rTVv,

Thank you once again for your insightful feedback. With the deadline approaching on December 2, we would greatly appreciate the opportunity to clarify any remaining concerns or answer any questions you may have.

If all issues have been addressed to your satisfaction, we kindly ask you to consider revising the scores accordingly after this discussion phase. We look forward to your continued feedback and hope to resolve any lingering doubts as efficiently as possible.

Thank you again for your time and dedication to this review!

Best,

Paper4233 Authors

Thank you for your clear and detailed responses. Your explanations have addressed my concerns thoroughly, and I will increase the scores.