3D-GP-LMVIC: Learning-based Multi-View Image Compression with 3D Gaussian Geometric Priors

hanks to you for the valuable comments. We are appreciated for recognizing the strengths of our paper in terms of performance. We will alleviate your remaining concerns as follows:

R1[The novelty of the depth estimation process]

---

Our depth map estimation based on the 3D Gaussian representation differs from the original 3D Gaussian Splatting method [1] and is not mentioned in [2]. This difference improves disparity estimation accuracy compared to the original method, which calculates pixel depth using a weighted averaging approach:

$$d = \sum_{i=1}^{M} T_i\alpha_i z_i,$$

where $T_i, \alpha_i, z_i$ represent the transmittance, opacity, and depth of the i-th point, respectively. In our work, we adopt a median-based estimation approach:

$$d = z_{i^*}, \hspace{0.5em} \text{where} \hspace{0.5em} i^* = \min ( i \mid T_i < 0.5 ).$$

Since the transmittance decreases as light passes through the points, starting from an initial value of 1, we take the depth of the point at which the transmittance drops below 0.5 as the estimated depth value.

We tested the alignment performance of the original 3D Gaussian Splatting depth estimation method under the same experimental settings as described in Appendix E of this paper. Since the code for [2] does not appear to be publicly available, we were unable to test the method from [2]. We combined the results with Table 3, as shown in the table below:

3D-GP-A outperforms 3D-GP-A (original) in PSNR and MS-SSIM, indicating that the proposed depth estimation method improves the alignment accuracy compared to the original depth estimation method used in 3D Gaussian Splatting.

We have explained the differences between our depth map estimation method and the original 3D-GS depth estimation method in lines 204–208 of the main text. Additionally, we have included a comparison of alignment performance with the original 3D-GS depth estimation method in Appendix E.

[1] Kerbl B, Kopanas G, Leimkühler T, et al. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 2023.

[2] Safadoust S, Tosi F, Güney F, et al. Self-Evolving Depth-Supervised 3D Gaussian Splatting from Rendered Stereo Pairs. BMVC, 2024.

R2[Section 3.2 needs revision]

---

Thank you for your suggestion. We have made significant revisions to Section 3.2, which you can find in the updated paper. Alternatively, you may refer to the following response.

R2.1[Explanation of differences compared to existing MVIC methods in (5) and (6)]

---

Thank you for your suggestion. We have revised the descriptions in lines 268–289 (5) and lines 320–328 (7) (previously (6)) to highlight the unique aspects of our compression model compared to other methods.

In (5), we primarily designed the disparity extractor and reference feature extractor. These two modules help the compression model integrate aligned reference features into the main network to improve compression efficiency.

In (6), we introduced the depth prediction extractor. This module predicts the current view's depth map based on the reference view's depth map, thereby reducing geometric redundancy between views.

R2.2[Purpose of the mask map in encoding]

---

The mask map is designed to identify non-overlapping regions between viewpoints and mask out irrelevant information in the reference features, thereby avoiding the introduction of noise. When capturing a scene from different viewpoints, an object may appear in one view but not in another. The mask map's purpose is to determine whether a pixel in the current view corresponds to the same object as the aligned pixel in the reference view. We have added the motivation for incorporating the mask map in lines 315–316.

The ablation study in Section 4.3 compares the proposed method with W/O Mask, demonstrating that incorporating the mask map contributes to improved compression efficiency.

R2.3[Figure 3 is complicate]

---

Thank you for your suggestion. We have simplified Figure 3 by dividing the flowchart into a high-level overview of the overall process and detailed components for each module. You can find the corresponding changes in the revised paper.

R2.4[$y_n$ is quantized in (4) and (5), but $z_n$ is quantized in Figure 3]

---

Thank you for your suggestion. We have added descriptions in lines 293–296 and 331–333 of the paper, explaining the quantization of $z_n$ and $z_{d_n}$ and the modeling of the latent variable probability distribution based on the quantized hyperprior representation. Since the entropy model primarily uses existing methods, we have kept this part concise.

Thank you once again for dedicating your valuable time to reviewing our paper and providing constructive comments!

As the end of the discussion period approaches, we kindly ask if our responses, which address all the questions and concerns you raised, have satisfactorily resolved your queries. Your feedback would be greatly appreciated, and we would be delighted to engage in further discussions if needed.

Sincerely,
The Authors

R2.5[Why is 'MaskConv' used in Figure 3?]

---

In the autoregressive entropy model, 'MaskConv' is used to divide the latent variable $y_n$ into disjoint segments, which are then encoded and decoded sequentially. Following your suggestion in 'R2.3', we have simplified Figure 3 by representing the context and hyperprior entropy model, which are not part of our proposed method, with a single block. As a result, 'MaskConv' has also been omitted. This revision makes the flowchart more concise and highlights our proposed approach.

R3[Suggestion to move (3) into appendix]

---

Thank you for your suggestion. Since the notation $d'_{n-1}$ in (3) is used in (4), and Section 3.3 defines the inter-view distance based on (3), we have decided to retain (3) in the main text for now. This ensures continuity in the article and improves readability for the audience.

Thank you again for reviewing our paper. We have addressed all the issues you raised and believe our responses sufficiently resolve your concerns. We have revised the description and flowchart in Section 3.2 to highlight the proposed components of the image compression model, including the disparity extractor, reference feature extractor, and image context transfer module. Additionally, we clarified the differences between our depth estimation method and the original approach used in 3D-GS, as well as the improvements in alignment performance resulting from these differences.

We hope you will recognize the revisions of our paper and consider revising your score. Should you have any further questions, we would be happy to provide additional clarifications.

We deeply value your initial feedback and understand that your schedule may be demanding. However, we kindly request that you take a moment to review our responses to your concerns.

We have made substantial revisions to Section 3.2, including simplifying the parts of the flowchart that describe prior work and emphasizing the details of the proposed method. Equations (5) and (7) now highlight the unique contributions of our disparity extractor and depth prediction extractor. Additionally, we have focused the writing more explicitly on our contributions. All your suggestions regarding Section 3.2 have been considered, and the necessary modifications have been made.

For Section 3.1, we have added clarifications to distinguish our depth estimation method from the one used in the original 3D-GS framework. Furthermore, we demonstrated through experiments that our proposed method achieves superior alignment performance.

Lastly, regarding your suggestion to move Equation (3) to the appendix, we considered its critical role in defining the notation used in Equation (4) and in Section 3.3, where we propose the view-to-view distance metric based on insights derived from Equation (3). To maintain coherence and readability, we have retained Equation (3) in the main text for now.

In conclusion, we kindly ask for your response once more, as we are eager to engage in further discussion with you.

Thank you again for reviewing our paper! In addition to addressing your comments on the writing, we would like to further elaborate on the motivation, contributions, and significance of our work for your consideration.

---

1. Motivation and Practical Applications

Our goal is to design a multi-view image compression framework tailored for wide-baseline setups to achieve accurate disparity estimation and effectively eliminate inter-view redundancy. Unlike prior works that focus on compressing stereo images—captured by closely positioned cameras and characterized by small, primarily horizontal disparities—we address wide-baseline setups, where data features irregular view relationships and less consistent disparities. These challenges render existing disparity estimation methods, such as homography transformation and patch matching, less effective, as demonstrated in Table 3 on the TnT dataset.

Wide-baseline setups are crucial for practical applications, where scenes often consist of dozens to hundreds of images, posing significant challenges for storage and transmission.

---

2. Contributions

• We propose leveraging 3D-GS to estimate depth maps, enabling accurate disparity estimation under wide-baseline setups.
• In the image compression model, we introduce a context transfer module for aligning features across viewpoints while masking non-overlapping regions between views.
• In the depth compression model, we design a depth prediction extractor to predict the depth map of the current view based on the reference view’s depth map, effectively reducing geometric redundancy between views.
• We develop a multi-view sequence sorting method to maximize similarity between adjacent views, improving compression efficiency.

We validated these contributions through comprehensive alignment and ablation experiments.

---

3. Significance

We believe our work offers a novel perspective in the multi-view image domain by shifting the focus from sparse-view image compression (e.g., stereo images) to wide-baseline setups. The latter captures more comprehensive scene information from diverse viewpoints, which is becoming increasingly relevant as 3D applications continue to grow. This type of data is inherently large-scale, making efficient compression crucial.

Additionally, we implemented a complete encoding and decoding pipeline and will release the code to enable further research and practical applications in this area.

---

We respect your evaluation and sincerely hope you recognize the significance and value of our work. We look forward to hearing your response!

Dear Reviewer c6CX,

We greatly appreciate the time and effort you have already dedicated to reviewing our work. As the discussion phase is nearing its conclusion, we kindly request your feedback regarding our responses to your comments.

To summarize, here are the key updates and clarifications we have made in response to your review:

1. Section 3.2 Improvements:

   • We revised the writing and reorganized the flowchart to simplify the parts borrowed from existing works while emphasizing the details of our proposed methods, including the disparity extractor, reference feature extractor, and image context transfer module.
   • Equations (5) and (7) now clearly highlight the unique aspects of our approach.

2. Clarification of Depth Estimation:

   • We explicitly explained the differences between our depth estimation method and the original 3D-GS framework, highlighting the resulting improvements in alignment performance.

3. Response to Your Suggestions:

   • Regarding your suggestion to move Equation (3) to the appendix, we provided a rationale for keeping it in the main text due to its relevance to Equation (4) and its role in motivating the view-sorting mechanism described in Section 3.3.

We believe these revisions address your concerns and enhance the clarity and strength of the paper. Your feedback on these updates is invaluable to us, and we hope for your response to ensure a productive discussion.

Thank you again for your time and support.

Best regards,
The Authors of "3D-GP-LMVIC: Learning-based Multi-View Image Compression with 3D Gaussian Geometric Priors"

Dear Reviewer c6CX,

As the discussion phase is coming to a close, we would like to kindly remind you of our responses to your valuable comments and suggestions. We have thoroughly addressed the concerns you raised, including:

1. Revising Section 3.2 to improve the description and flowchart, with a focus on highlighting the unique contributions of our method.
2. Clarifying the differences between our depth estimation method and the original approach used in 3D-GS, demonstrating the resulting improvements in alignment performance through experiments.

We deeply value your feedback and hope you will have the opportunity to share any further thoughts before the discussion period concludes. Your insights are greatly appreciated and will help us further enhance the quality of our work.

Thank you once again for your time and effort in reviewing our paper.

Best regards,
The Authors of "3D-GP-LMVIC: Learning-based Multi-View Image Compression with 3D Gaussian Geometric Priors"

R3[Depth map bpp calculation and effectiveness]

---

We calculated the bpp of the depth map in the proposed 3D-GP-LMVIC framework. The effectiveness of the depth map is evaluated in the ablation study in Section 4.3 of the paper. In the ablation study, Concatenation does not calculate the bpp of the depth map and does not use depth map-based disparity estimation to align features between two views. Instead, it directly concatenates the features of the reference view and the target view. The rate-distortion curves in Figure 5 show that 3D-GP-LMVIC improves compression efficiency by approximately 33% compared to Concatenation. This demonstrates that incorporating the depth map is beneficial.

R4[BiSIC codec time and computational complexity analysis]

---

Thank you for your suggestion. The open-source code of BiSIC does not provide separate implementations for encoding, decoding, or entropy coding, so we only measured its overall computational complexity. Since BiSIC is an improvement over LDMIC with a more complex autoencoder structure and entropy model, we infer that BiSIC's encoding and decoding time is longer than that of LDMIC, and BiSIC-Fast's encoding and decoding time is longer than that of LDMIC-Fast.

Additionally, we evaluated the Multiply-Accumulate Operations (MACs), model parameters, coding speed, and memory usage of various learning-based image codecs under the same experimental environment and settings as described in Section 4.2 of this paper for measuring encoding and decoding speed. Furthermore, we tested a lightweight version of 3D-GP-LMVIC, where the feature channel dimensions of the autoencoder and entropy model were reduced from (128, 192) to (100, 128). The experimental results are shown in the table below.

The MACs of the encoder in our method are relatively large, but they should still be smaller than those of BiSIC. Since BiSIC uses a symmetric encoder-decoder structure, we estimate that the MACs for its encoder and decoder each account for approximately half of the total MACs. Furthermore, in addition to designing an image codec, we also designed a depth map codec, which is another reason for the larger MACs and model parameter count. However, overall, the MACs of 3D-GP-LMVIC are still lower than those of the current SOTA scheme, BiSIC. The lightweight version of 3D-GP-LMVIC ranks in the middle among the compared methods in terms of MACs and model parameters, while its memory usage is quite low at only 1532M, similar to LDMIC-Fast.

The table below shows the BDBR comparison of the lightweight 3D-GP-LMVIC and other methods relative to MV-HEVC on the DeepBlending dataset.

3D-GP-LMVIC (100, 128) outperforms other multi-view image codecs in both PSNR and MS-SSIM.

In summary, the computational complexity of the proposed 3D-GP-LMVIC is within an acceptable range, overall comparable to that of the SOTA BiSIC and slightly better than BiSIC-Fast. SASIC and LDMIC-Fast have certain advantages in terms of computational complexity, but they fall short of the proposed method in terms of compression performance. Moreover, we found that reducing the channel count of 3D-GP-LMVIC can significantly reduce computational complexity (the number of model parameters is reduced by approximately 37%, and the memory usage is only 1532M) while maintaining satisfactory compression performance.

Thanks to you for the valuable comments. We are appreciated for recognizing the strengths of our paper in terms of performance. We will alleviate your remaining concerns as follows:

R1[How 3D Gaussian Splatting derives geometric priors and its advantages]

---

1. 3D Gaussian Representation and Geometry Modeling
3D Gaussian Splatting uses a large number of views to train a dense 3D Gaussian representation that captures the entire scene’s geometry. This allows for novel view synthesis by ensuring the representation contains sufficient geometric information. Without accurate geometric priors, rendering from arbitrary viewpoints may produce inconsistencies.

2. Depth Estimation with 3D Gaussian Representation
Depth estimation involves tracing a ray from the camera center through a pixel and identifying all intersecting 3D Gaussians. The depth is determined using a transmittance-based approach: the first 3D Gaussian reducing transmittance below 0.5 is taken as the pixel’s depth. This method accounts for occlusions and ensures depth accuracy.

3. Advantages Over Traditional Methods
Traditional disparity estimation relies on local similarity between stereo images, which works well for small viewpoint changes. However, in wide-baseline setups with significant geometric differences, such methods often fail. In contrast, the proposed 3D Gaussian-based disparity estimation method align 3D points across views, handling large viewpoint variations effectively. Since 3D Gaussian Splatting models the entire scene using a large number of views, it comprehensively captures 3D spatial relationships and provides accurate geometric information for disparity estimation. Appendix E of this paper compares the alignment performance of traditional disparity estimation methods and the proposed method for wide-baseline multi-view data. Experimental results demonstrate the superiority of the proposed method.

4. Benefits for Compression Efficiency
Accurate disparity estimation using 3D Gaussians enables precise alignment of features across views, effectively eliminating redundant information. This significantly improves compression efficiency for wide-baseline multi-view datasets.

R2[Why use Mip-NeRF 360 and TnT datasets instead of stereo image datasets]

---

1. Motivation and Practical Applications
As mentioned above, we believe that previous multi-view image compression methods have already achieved satisfactory disparity estimation on stereo image datasets. However, their performance on wide-baseline setups for multi-view datasets is still insufficient. Therefore, we mainly focus on designing a multi-view image compression framework for wide-baseline setups to achieve more accurate disparity estimation and effectively exploit this estimation to eliminate redundancy across views. Additionally, we believe that multi-view image compression under wide-baseline setups is essential for practical applications, as a single scene in such setups often consists of dozens to hundreds of images, posing significant challenges for data storage and transmission.

2. Limitations of Stereo Images for Learning Geometry in 3D Gaussian Representation
We found that for stereo image datasets with only two viewpoints, the 3D Gaussian representation struggles to learn accurate geometric information. In contrast, multi-view data under wide-baseline setups provides rich scene information sampled from widely varying viewpoints, which helps the 3D Gaussian representation learn accurate geometry. In stereo image datasets, the cameras are often positioned close to each other with similar orientations, making it difficult for the 3D Gaussian representation to correct geometric errors through large viewpoint variations and learn the correct geometric structure. The table below presents the alignment performance of various disparity estimation methods on the Cityscapes and KITTI Stereo datasets under the same experimental settings as the alignment experiments described in Appendix E of this paper.

Experiments have shown that the 3D Gaussian representation cannot learn accurate geometry based on stereo image data with only two viewpoints. Additionally, the disparity estimation modules in current multi-view image codecs, such as Patch Matching (PM), already perform well on stereo image datasets. Therefore, we did not conduct further tests on stereo image datasets.

Thank you once again for dedicating your valuable time to reviewing our paper and providing constructive comments!

As the end of the discussion period approaches, we kindly ask if our responses, which address all the questions and concerns you raised, have satisfactorily resolved your queries. Your feedback would be greatly appreciated, and we would be delighted to engage in further discussions if needed.

Sincerely,
The Authors

Thanks to you for the valuable comments. We are appreciated for recognizing the strengths of our paper. We will alleviate your remaining concerns as follows:

R1[Use existing depth map estimation method based on 3D Gaussian reconstruction for disparities estimation]

---

Our depth map estimation based on 3D Gaussian representation differs from the original 3D Gaussian Splatting method [1]. We also found that this difference leads to improved accuracy in disparity estimation. Specifically, the original 3D Gaussian Splatting method calculates the depth of each pixel using a weighted averaging approach:

$$d = \sum_{i=1}^{M} T_i\alpha_i z_i,$$

where $T_i, \alpha_i, z_i$ represent the transmittance, opacity, and depth of the i-th point, respectively. In our work, we adopt a median-based estimation approach:

$$d = z_{i^*}, \hspace{0.5em} \text{where} \hspace{0.5em} i^* = \min ( i \mid T_i < 0.5 ).$$

Since the transmittance decreases as light passes through the points, starting from an initial value of 1, we take the depth of the point at which the transmittance drops below 0.5 as the estimated depth value.

We tested the alignment performance of the original 3D Gaussian Splatting depth estimation method under the same experimental settings as in Appendix E of this paper, and combined the results with Table 3, as shown in the table below:

3D-GP-A outperforms 3D-GP-A (original) in PSNR and MS-SSIM, indicating that the proposed depth estimation method improves the alignment accuracy compared to the original depth estimation method used in 3D Gaussian Splatting.

[1] Kerbl B, Kopanas G, Leimkühler T, et al. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 2023.

R2[Compare to other depth estimation methods (Colmap and MVSFormer++)]

---

Thank you for your valuable suggestions. The primary motivation of our work is to design a compression framework for multi-view image data under wide-baseline setups. We agree that exploring potentially better depth estimation methods is crucial, but we also believe that developing a dedicated compression framework for this specific scenario is important. Beyond depth estimation, our compression framework also includes a depth map compression model to remove geometric redundancy between views, a context transfer module for aligning features across viewpoints while masking out non-overlapping areas between views, and a multi-view sequence sorting algorithm to ensure higher similarity between adjacent views. We believe that designing this pipeline is meaningful, as it provides a reference for future researchers to develop compression frameworks for wide-baseline multi-view image data or to further enhance specific modules within this context.

Additionally, we have followed your suggestion and conducted alignment experiments using COLMAP and MVSFormer++ for depth map estimation under the same experimental settings as described in Appendix E of this paper. The results are combined with Table 3, as shown below:

3D-GP-A outperforms COLMAP and MVSFormer++ in both PSNR and MS-SSIM, indicating its effectiveness in disparity estimation.

R3[No visual comparisons in the main text]

---

Thank you for your suggestion. We have added Figure 5 in the main text, showcasing a visual comparison between the proposed 3D-GP-LMVIC, LDMIC, and BiSIC. The results demonstrate that 3D-GP-LMVIC preserves more texture details and achieves higher reconstruction quality while consuming fewer bits. You can find the visual comparison results in the revised paper.

R4[The network architecture figure has too many details]

---

Thank you for your suggestion. We have revised Figure 3 by simplifying the modules proposed by others and highlighting our proposed method. You can find the corresponding changes in the revised paper.

R5[Compression and decompression time for a set of images]

---

For the Tanks & Temples train scene, there are a total of 301 images. We conducted tests on a platform equipped with an Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz and an NVIDIA RTX A6000 GPU. On average, encoding each image takes 0.19 seconds, and decoding takes 0.18 seconds. Additionally, training a 3D Gaussian representation for 30,000 iterations requires approximately 13 minutes and 35 seconds. Therefore, compressing this image set takes approximately 14 minutes and 32 seconds, while decompression requires about 54 seconds.

We also tested reducing the number of 3D Gaussian representation training iterations to 7,000, which resulted in a training time of 2 minutes and 22 seconds. The total compression time could thus be reduced to 3 minutes and 19 seconds. Additionally, we evaluated the alignment performance of depth estimation based on the 3D Gaussian representation trained with 7,000 iterations under the same experimental settings as described in Appendix E of this paper. The results were combined with Table 3, as shown in the table below:

3D-GP-A (7000 iterations) achieves performance comparable to 3D-GP-A (30000 iterations) in terms of PSNR and MS-SSIM.

R6[Why use Mip-NeRF 360 and TnT datasets instead of KITTI and InStereo2K?]

---

1. Motivation and Practical Applications
We aim to design a multi-view image compression framework tailored for wide-baseline setups to achieve accurate disparity estimation and effectively eliminate inter-view redundancy. Unlike KITTI and InStereo2K datasets, which contain stereo images with small, mostly horizontal disparities captured by closely positioned cameras, wide-baseline setups in datasets like Mip-NeRF 360 and TnT feature irregular view relationships and less consistent disparities. These characteristics make existing disparity estimation methods, such as homography transformation and patch matching, less effective, as shown in Table 3 on the TnT dataset. Wide-baseline setups are also critical for practical applications, where scenes often consist of dozens to hundreds of images, creating significant challenges for storage and transmission.

2. Limitations of Stereo Images for Learning Geometry in 3D Gaussian Representation
We found that for stereo image datasets with only two viewpoints, the 3D Gaussian representation struggles to learn accurate geometric information. In contrast, multi-view data under wide-baseline setups provides rich scene information sampled from widely varying viewpoints, which helps the 3D Gaussian representation learn accurate geometry. In stereo image datasets, the cameras are often positioned close to each other with similar orientations, making it difficult for the 3D Gaussian representation to correct geometric errors through large viewpoint variations and learn the correct geometric structure. The table below presents the alignment performance of various disparity estimation methods on the Cityscapes and KITTI Stereo datasets under the same experimental settings as the alignment experiments described in Appendix E of this paper.

Experiments have shown that the 3D Gaussian representation cannot learn accurate geometry based on stereo image data with only two viewpoints. Additionally, the disparity estimation modules in current multi-view image codecs, such as Patch Matching (PM), already perform well on stereo image datasets. Therefore, we did not conduct further tests on stereo image datasets.

Thank you once again for dedicating your valuable time to reviewing our paper and providing constructive comments!

As the end of the discussion period approaches, we kindly ask if our responses, which address all the questions and concerns you raised, have satisfactorily resolved your queries. Your feedback would be greatly appreciated, and we would be delighted to engage in further discussions if needed.

Sincerely,
The Authors

Thank you again for reviewing our paper. We have addressed all the issues you raised and believe our responses sufficiently resolve your concerns.

Regarding your comment on the lack of novelty in our depth estimation approach based on 3D-GS, we first demonstrated that our method outperforms COLMAP and MVSFormer++ in alignment accuracy. Furthermore, we identified the key distinctions between our approach and other multi-view image compression methods: (1) leveraging 3D-GS to model spatial geometric relationships and perform disparity estimation between views; (2) integrating a context transfer module based on disparity and disparity masks within the image compression model; (3) eliminating geometric redundancy between views to improve compression efficiency; and (4) proposing a view-to-view distance metric to reorder views with high correlation in adjacent positions, thereby enhancing compression efficiency.

We validated the effectiveness of these innovations through alignment and ablation experiments. Additionally, we refined the description and flowchart in Section 3.2 to highlight the novelties within the image compression model. We also implemented a complete encoding and decoding pipeline and will release the code to facilitate further research and practical applications. We believe our work introduces a new perspective to the multi-view image domain by shifting the focus from sparse-view image compression (e.g., stereo images) to wide-baseline setups, which capture more comprehensive scene information from diverse viewpoints. As 3D applications become increasingly prevalent, this type of data will grow in popularity, and efficient compression for such large-scale data is of significant importance.

We hope you will recognize the novelty of our approach and consider revising your score. Should you have any further questions, we would be happy to provide additional clarifications.

Thanks to you for the valuable comments. We are appreciated for recognizing the strengths of our paper in terms of good motivation, novelty, and effectiveness. We will alleviate your remaining concerns as follows:

R1[Test on high resolution images]

---

We conducted a survey of various multi-view datasets and found that the 1080p images from the Tanks&Temples dataset, which we have tested, are already among the higher-resolution images in these datasets. Our method has demonstrated superior performance compared to existing multi-view image compression schemes on the Tanks&Temples dataset in terms of the BDBR metric relative to MV-HEVC, as shown in the experimental results in the table below.

If you could provide some higher-quality datasets, we would be willing to conduct further testing.

R2[Computational complexity analysis]

---

We evaluated the Multiply-Accumulate Operations (MACs), model parameters, coding speed, and memory usage of various learning-based image codecs under the same experimental environment and settings as described in Section 4.2 of this paper for measuring encoding and decoding speed. The open-source code of BiSIC does not provide separate encoder and decoder implementations, so we only measured its overall computational complexity. Additionally, we tested a lightweight version of 3D-GP-LMVIC, where the feature channel dimensions of the autoencoder and entropy model were reduced from (128, 192) to (100, 128). The experimental results are shown in the table below.

The MACs of the encoder in our method are relatively large, but they should still be smaller than those of BiSIC. Since BiSIC uses a symmetric encoder-decoder structure, we estimate that the MACs for its encoder and decoder each account for approximately half of the total MACs. Furthermore, in addition to designing an image codec, we also designed a depth map codec, which is another reason for the larger MACs and model parameter count. However, overall, the MACs of 3D-GP-LMVIC are still lower than those of the current SOTA scheme, BiSIC. The lightweight version of 3D-GP-LMVIC ranks in the middle among the compared methods in terms of MACs and model parameters, while its memory usage is quite low at only 1532M, similar to LDMIC-Fast.

The table below shows the BDBR comparison of the lightweight 3D-GP-LMVIC and other methods relative to MV-HEVC on the DeepBlending dataset.

3D-GP-LMVIC (100, 128) outperforms other multi-view image codecs in both PSNR and MS-SSIM.

In summary, the computational complexity of the proposed 3D-GP-LMVIC is within an acceptable range, overall comparable to that of the SOTA BiSIC and slightly better than BiSIC-Fast. SASIC and LDMIC-Fast have certain advantages in terms of computational complexity, but they fall short of the proposed method in terms of compression performance. Moreover, in resource-constrained scenarios, we found that reducing the channel count of 3D-GP-LMVIC can significantly reduce computational complexity (memory usage is only 1532M) while maintaining satisfactory compression performance.

Thank you once again for dedicating your valuable time to reviewing our paper and providing constructive comments!

As the end of the discussion period approaches, we kindly ask if our responses, which address all the questions and concerns you raised, have satisfactorily resolved your queries. Your feedback would be greatly appreciated, and we would be delighted to engage in further discussions if needed.

Sincerely,
The Authors

Thank you again for reviewing our paper. We have addressed all the issues you raised, and we believe our responses sufficiently resolve your concerns. First, we have demonstrated the superior performance of the proposed method on the high-quality Tanks&Temples dataset with a resolution of 1080p. Second, we provided a detailed comparison of computational complexity among different methods, showing that our proposed method achieves comparable or slightly lower complexity compared to the SOTA BiSIC, while significantly reducing complexity in our lightweight version without compromising compression performance, which outperforms the SOTA methods.

We hope you will find our responses satisfactory and consider revising your score. Should you have any further questions, we are more than happy to provide additional clarification.

We deeply value your initial feedback and understand that you may have a busy schedule. However, we kindly request that you take a moment to review our responses to your concerns.

Regarding your first question, we have evaluated the performance of various methods on the high-quality 1080p Tanks&Temples dataset. The results demonstrate that the proposed 3D-GP-LMVIC outperforms existing multi-view codecs in terms of compression performance.

Regarding your second question, we have tested the MACs, model parameters, and memory usage of various methods. Experimental results show that the proposed method achieves slightly lower complexity compared to the SOTA BiSIC. Furthermore, our lightweight version not only outperforms BiSIC in terms of performance but also significantly reduces complexity. Its memory usage is only 1532 MB, approximately half of BiSIC’s, and is comparable to the most lightweight model among the baselines, LDMIC.

In conclusion, we kindly ask for your response once more, as we are eager to engage in further discussion with you.

Dear Reviewer MyG9,

Thank you for taking the time to review our paper and for providing valuable feedback. As the discussion phase is nearing its conclusion, we kindly request your response regarding our rebuttal and the updates we made to address your comments.

To summarize, here are the key points we addressed in response to your review:

1. Evaluation on High-Quality Datasets:

   • We demonstrated the superior performance of our proposed 3D-GP-LMVIC method on the 1080p Tanks & Temples dataset, surpassing existing state-of-the-art multi-view codecs.

2. Analysis of Computational Complexity:

   • We provided detailed comparisons of MACs, model parameters, and memory usage across methods. Our proposed method achieves slightly lower complexity than BiSIC while outperforming it in compression performance.
   • The lightweight version of our model significantly reduces complexity, with memory usage of only 1532M, roughly half of BiSIC, and comparable to the most lightweight baseline, LDMIC.

We believe these updates comprehensively address your concerns and strengthen the contributions of our work. Your insights and feedback on these points would be invaluable to further refining the paper.

We understand you may have a busy schedule but kindly ask for your input before the discussion phase concludes.

Thank you again for your time and consideration.

Best regards,
The Authors of "3D-GP-LMVIC: Learning-based Multi-View Image Compression with 3D Gaussian Geometric Priors"

Dear Reviewer MyG9,

As the discussion phase is coming to an end, we would like to kindly remind you of our responses to your valuable comments and suggestions. We have carefully addressed the concerns you raised, including:

1. Evaluating our method on the high-quality 1080p Tanks&Temples dataset, which demonstrates that our method consistently outperforms baselines.
2. Providing a detailed comparison of computational complexity (MACs, parameters, and memory usage), where our lightweight version shows clear advantages over the SOTA method BiSIC.

We deeply value your feedback and hope to hear your thoughts on these responses. Your insights would greatly help us refine our work further.

Thank you again for your time and effort in reviewing our paper. We sincerely appreciate your contributions to the discussion process.

Best regards,
The Authors of "3D-GP-LMVIC: Learning-based Multi-View Image Compression with 3D Gaussian Geometric Priors"