Instant4D: 4D Gaussian Splatting in Minutes

Thank you very much for your recognition and valuable feedback. We will revise this paper based on your suggestions. Please feel free to point them out if any of our explanations are unclear.

Weakness:

1:Our Pipeline Design

Our motivation is rooted in the recent advances in deep visual SLAM systems, which have significantly improved speed [1], robustness [2], and generalization [3] in geometry recovery from casually captured monocular videos. Traditional 4DGS methods initialize from the sparse point cloud calibrated from COLMAP. Our methods show that in the casual video setting, the training speed will be greatly improved by introducing our initialization strategy.

To best our knowledge, our work is one of the first to tightly integrate deep visual SLAM with 4D Gaussian Splatting, bridging the gap between geometry-centric SLAM research and appearance-centric neural rendering. We fully leverage the deep visual SLAM tool, achieving fast, robust, and efficient 4D reconstruction in monocular settings. We believe this direction is valuable for the community, as it provides a strong baseline, enables real-time rendering, and could be used as a dataset to train feedforward novel view synthesis models for casual video. We will open-source our code to further facilitate our research community.

Question:

1: More Statistics about the Result

We agree that more statistical analysis would strengthen our claims. In the revised manuscript, we will include standard deviation and confidence intervals for the ablation study, and test the statistical significance against baseline methods. We believe these additions will better reflect the consistency and statistical reliability of our improvements.

Here, we provide a detailed ablation study with mean and standard deviation on the Dycheck dataset by repeating for 5 times:

2: Evaluation Compared with Online 4DGS Methods

Thank you for providing these useful suggestions. Due to limited time, we are unable to provide the quantitative comparison here. We will properly cite related work and include the comparison with 3DGStream [4] in our final version.

For qualitative comparison, one advantage of our methods is the initialization strategy. 3DGStream [4] requires a 5000-iteration initialization for the first frame, while in our method all Gaussians will receive gradient updates at the beginning. Besides, our other advantage is the native 4D Gaussian representation, which does not require per-frame neural transformation.

3: Evaluation with COLMAP Initialization & Analysis

Here is the experiment evaluated on the NVIDIA dataset with COLMAP-based camera pose:

---

Note that for our method initialized from COLMAP, we turn off the motion-aware Gaussian design, enable densification, and set the number of iterations to 5000. For the Ours runtime, we include the calibration time with deep visual SLAM. For COLMAP's runtime, we exclude the calibration time, which is around 5 times longer than the current runtime. We will add more analysis on the efficiency of our initialization strategy, with the metrics listed in question 1.

Finally, we want to express our sincere gratitude. Our paper will definitely be better with your suggestions. Thank you!

---

Reference:

[1] Jianing Yang, Alexander Sax, Kevin J. Liang et al. Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass

[2] Jianyuan Wang, Minghao Chen, Nikita Karaev et al. VGGT: Visual Geometry Grounded Transformer

[3] Zhengqi Li, Richard Tucker, Forrester Cole et al. MegaSaM: Accurate, Fast, and Robust Structure and Motion from Casual Dynamic Videos

[4] Jiakai Sun, Han Jiao, Guangyuan Li et al. 3DGStream: On-the-Fly Training of 3D Gaussians for Efficient Streaming of Photo-Realistic Free-Viewpoint Videos

Thank you for your in-depth discussions and suggestions for our paper

Weakness:

1: Qualitative Evaluation & Comparison with Other Methods

For the visualization, we have addressed this issue by adding camera movement. We promise to update the visualization part on our official website very soon after the reviewer discussion (following the rebuttal rules). For the train/test split, we believe Gaussian Splatting/NeRF was originally designed for interpolation between different viewpoints, therefore these per-scene methods naturally do not support extrapolation with large camera movement. Our train/test splitting could be viewed as temporal interpolation.

For comparison with other methods, Shape-of-Motion [4] requires an additional tracking model to optimize, and the preprocessed file will reach 150 GB even for a 10-second video. Shape-of-Motion [4] also requires users to segment the mask themselves. Our method is fully automated. MoSca [5] utilizes depth, point track, and optical flow to reconstruct the scene and camera parameters. However, this design significantly increases the training time because of the complexity of the optimization. Due to time constraints, we are unable to present a complete evaluation. In the future, we will continue experiments on that benchmark.

2: Covariance Design

We choose to simplify the scaling matrix into one shared shape scale $s_{xyz}$ and one temporal scale $s_t$. We find that the isotropic design regularizes optimization and improves rendering quality as shown in Table 4.

From our observations, this design can reduce overfitting and improve rendering quality from novel views. We assume the flickering behavior is caused by the lighting effects and texture changes in the bear's fur. We also tried the anisotropic Gaussians formulation, which does not solve this flickering behavior. Based on the overall trade-off, the simplified design is feasible. We will add this comparison for more visualization.

3: Technical Contribution

Our motivation is rooted in the recent advances in deep visual SLAM systems, which have significantly improved speed [1], robustness [2], and generalization [3] in geometry recovery from casually captured monocular videos. Traditional 4DGS methods initialize from the sparse point cloud calibrated from COLMAP. Our methods show that in the casual video setting, the training speed will be greatly improved by introducing our initialization strategy.

Our work is, to our knowledge, one of the first to tightly integrate deep visual SLAM with 4D Gaussian Splatting, bridging the gap between geometry-centric SLAM research and appearance-centric neural rendering. We fully leverage the deep visual SLAM tool, achieving fast, robust, and efficient 4D reconstruction in monocular settings. We believe this direction is valuable for the community, as it provides a strong baseline, enables real-time rendering, and could be used as a dataset to train feedforward novel view synthesis models for casual video. We will open-source our code to further facilitate our research community.

4: Typo

We have corrected them in our manuscript. We will do a second proofreading of our manuscript.

Finally, we sincerely thank you for your questions. Your insights have indeed helped us further refine our paper.

---

Reference:

[1] Jianing Yang, Alexander Sax, Kevin J. Liang et al. Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass

[2] Jianyuan Wang, Minghao Chen, Nikita Karaev et al. VGGT: Visual Geometry Grounded Transformer

[3] Zhengqi Li, Richard Tucker, Forrester Cole et al. MegaSaM: Accurate, Fast, and Robust Structure and Motion from Casual Dynamic Videos

[4] Qianqian Wang, Vickie Ye and Hang Gao et al. Shape of Motion: 4D Reconstruction from a Single Video

[5] Jiahui Lei, Yijia Weng, Adam Harley et al. MoSca: Dynamic Gaussian Fusion from Casual Videos via 4D Motion Scaffolds

We thank Reviewer vtZE for more detailed feedback, and we would love to respond further to these valuable comments.

• We would like to emphasize that all the presented outputs in the paper are rendered from novel viewpoints provided by the official evaluation sets. Additionally, we believe that there is a misunderstanding on the "rendered videos from training viewpoints". We hope to draw the attention of Reviewer vtZE around our rendered videos on our website. In fact, our rendering roughly followed the camera trace of the given video but we also rendered novel views for presentation in each video.

• To further validate the effectiveness of our method, we evaluated it on the DyCheck dataset under the same settings, comparing it with the most recent approaches including MoSca [1], RoDyGS [2], Shape-of-Motion [3]): | Model | Training Time (min) | Rendering (FPS) | PSNR (dB) | |----------|----------------|---------------|---------| | MoSca [1] | 63 | 37 | 25.03 | | RoDyGS [2] | 60 | 67 |17.37| | Shape-of-Motion[3]| 74 | 131 | 23.84 | | Ours | 11.2 | 120 | 24.52 |

  Note that:

  • Bold denotes the best result in a column and Italic indicates the second best.
  • For fair comparison, we implemented all these methods (including MoSca [1], which adopted ground-truth depth in the official code) using estimated depth. Training time includes preprocessing.

  Analysis: As shown in the table, our method reduces the training time by over 80%, supporting the goal of fast casual-video reconstruction. Simultaneously, it remains competitive -- ranking second in terms of both rendering speed and quality.

• As Reviewer vtZE discussed, the design of "Isotropic Gaussian" degrades the center of primitive as: $\mu_{xyz|t}=\mu_{1:3}$ . Obviously, timestamp $t$ cannot affect the spatial center $\mu_{xyz}$. However, we argue that, modeling the relationship between opacity and timestamp by Equation $o_t = o \times \mathcal{N}(t, \mu_4, \Sigma_{4,4})$, along with more primitives kept by an aggressive grid pruning strategy, helps capture the motion information both temporally and spatially.
• Under the same settings, the empirical results (refer to Table 4) prove the effectiveness of our design, as below:

• As shown in Figure 2, we apply the output of Mega-SAM [5] to the Geometric Recovery process. As mentioned by Reviewer vtZE, this process performs back-projection using the output camera parameters and video depth.

• However, instead of a naive addition, we further explore the mechanism of Mega-SAM [5] and fuse its intermediate probability map $\hat{m} \in \mathbb{R}^{\frac{H}{8} \times \frac{W}{8}}$ into our framework. Specifically, we convert probablity map into motion mask to determine whether each Gaussian primitive is from static or dynmaic element and to assign suitable values of s_t. This ensures to initialize compact static background and dynamic objects with less overlappings. The empirical results prove the effectiveness of this fusion: | Component | PSNR ↑ | SSIM ↑ | |------------------------|--------|--------| | w/ Motion Aware Gaussian | 24.52 | 0.834 | | w/o Motion Aware Gaussian | 21.11 | 0.721 |

• For more clear illustration to readers, we will improve Figure 2 by describing how probability map flows in the Geometric Recovery process and the pipeline.

---

Reference:

[1] Jiahui Lei, Yijia Weng, Adam Harley et al. MoSca: Dynamic Gaussian Fusion from Casual Videos via 4D Motion Scaffolds

[2] Yoonwoo Jeong, Junmyeong Lee, Hoseung Choi et al. RoDyGS: Robust Dynamic Gaussian Splatting for Casual Videos

[3] Qianqian Wang, Vickie Ye and Hang Gao et al. Shape of Motion: 4D Reconstruction from a Single Video

[4] Zeyu Yang, Hongye Yang, Zijie Pan et al. Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting.

[5] Zhengqi Li, Richard Tucker, Forrester Cole et al. MegaSaM: Accurate, Fast, and Robust Structure and Motion from Casual Dynamic Videos

Thank you for your in-depth discussions and constructive suggestions for our paper. Our responses to the weaknesses and questions are listed below.

Major Weakness:

1: Runtime Analysis with Other Work

Before clarification, we show a "detailed" Table 1 for the comparison of efficiency:

---

• For fair comparison, we implement a popular and competitive method InstantSplat [6] and the baseline 4DGS [7] and compare our method with them using the same server settings.
• The implementation of Casual-FVS [5] has not been open-sourced, so we can hardly report the runtime with the same server settings. We will not compare our method with it in Table 1 for our camera-ready version, but we will keep citing it.
• Our method performs over 1000x faster than some previous NeRF methods like HyperNeRF [1], DynamicNeRF [2] and RoDynRF [3]. Though these methods used different GPUs from ours, we believe the obvious efficiency superiority of our method can show its competitiveness.

2: Discussion on Depth Estimation in our Work

For a detailed discussion, we list the updated Table 4 below:

---

We acknowledge that the speedup and the performance increase are related to the underlying depth estimation algorithm, while we believe our contributions such as the pruning strategy and simplified 4D Gaussian representation are orthogonal to the specifics of the initial depth estimation.

Our ablation study shows that the isotropic Gaussian formulation regularizes the training and increases the rendering quality by 1.52 dB. To solve the problem that the Gaussians outside of the frustum tend to disappear, we design the motion-aware Gaussian strategy, improving the render result by 3.41 dB PSNR compared with the baseline. Experiments show that our method achieves a 30× speed-up and reduces training time to within two minutes. As a result, our improvement does not merely stem from the depth estimation.

3: Comparison between the Optimized Depth and the Actual Depth

We conducted a new experiment on the Dycheck dataset to compare the depth estimated by Unidepth [8] with the ground-truth depth. We will also incorporate the table below in our supplementary material for our camera-ready version.

The comparison shows that the first-order optimization reduces error in Abs-Rel by 46.1% and in Log-RMSE by 16.2%, and increases $\delta_{1.25}$ accuracy by 30.8%.

4: Evaluation under Different Depth Setting

Thank you for emphasizing the importance of depth estimation. As shown in DepthSplat [9], better depth leads to improved novel view synthesis quality. The current evaluations are done under short real-world video and AI generated videos that are featured with foreground moving objects and static background, which does not involve different depth settings. We will add more discussion about the depth setting.

Minor Weakness:

1 & 2

We have fixed this problem and will do a thorough proofreading.

Question:

1: Zoom-in & Zoom-out Effect

With significant zoom-in, we find that our method tends to display clear boundaries without outlier Gaussians that are frequent in 3D Gaussian Splatting and other dynamic reconstruction work. We do not find unusual visual effects other than this. With significant zoom-in, we observe some voidness between Gaussians. This is not observed in the baseline methods.

2: Deblurring for Motion and Defocus

Motion blurring and defocus blurring are challenging problems. These problems are not quite obvious when evaluating since we typically have a higher video frame rate compared with the camera and object motion. While current dynamic methods don't focus on this problem, we are open to exploring this issue.

3: Application with Video Depth Model?

We agree that better depth estimation will improve the result. However, we find that a video-depth prediction model might not be the best solution in our method. For instance, the NVIDIA dynamic dataset is featured with a 12-camera teleporting setup, which is incompatible with the assumptions in the video depth model.

Finally, we sincerely thank you for your questions. Your insights have indeed helped us further refine our paper.

---

Reference

[1] Keunhong Park, Utkarsh Sinha, Peter Hedman et al. HyperNeRF: A Higher-Dimensional Representation for Topologically Varying Neural Radiance Fields.

[2] Chen Gao, Ayush Saraf, Johannes Kopf et al. Dynamic View Synthesis from Dynamic Monocular Video.

[3] Yu-Lun Liu, Chen Gao, Andreas Meuleman et al. Robust Dynamic Radiance Fields.

[4] Guanjun Wu, Taoran Yi, Jiemin Fang et al. 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering.

[5] Yao-Chih Lee, Zhoutong Zhang, Kevin Blackburn-Matzen et al. Fast View Synthesis of Casual Videos with Soup-of-Planes.

[6] Zhiwen Fan, Wenyan Cong, Kairun Wen et al. InstantSplat: Sparse-view Gaussian Splatting in Seconds.

[7] Zeyu Yang, Hongye Yang, Zijie Pan et al. Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting.

[8] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis et al. UniDepth: Universal Monocular Metric Depth Estimation.

[9] Haofei Xu, Songyou Peng, Fangjinhua Wang et al. DepthSplat: Connecting Gaussian Splatting and Depth.

We thank reviewer K6kc for the valuable feedback.

1: Depth-Dependent Module Analysis

Originally, we applied MegaSAM [7] for the Geometric Recovery process, using UniDepth [3] as depth estimator and performing the consistent depth optimization (denoted as Ours). To assess whether our method overfits to a specific depth estimator, we disabled the consistent depth optimization and replaced the depth estimator UniDepth [3] with its alternatives VideoDepthAnything [4] and Cut3R [5], denoted as Ours-VideoDepthAnything and Ours-Cut3R, respectively. For reference, we also provide a UniDepth version without consistent depth optimization, denoted as Ours-UniDepth.

The results on the Dycheck dataset are shown below:

---

The results show that across various depth estimators, the proposed method consistently outperforms previous state-of-the-art methods, RoDynRF [1] and RoDyGS [2] by a large margin. This fact demostrates that the improvements of the proposed method are not tied to a specific depth estimator, and the method does not overfit to a particular depth setting.

2: Does your first-order optimization address this issue to prevent the system from overfitting to specific scenes?

Our approach performs per-scene optimization to enable real-time rendering at inference. We do not expect the model to generalize depth predictions across unrelated scenes.

3: UniDepth itself exhibits inconsistent performance in large-scale scenes.

In our deep visual SLAM part, we do not rely solely on Unidepth [3] to generate sufficiently accurate depth. Instead, we apply additional optimization to acquire consistent video depth, as shown in the following table.

On the Dycheck dataset, the optimization reduces Abs-Rel and Log-RSME errors by 46.1%, 16.2% respectively, indicating that the refined depth can be quantitatively closer to the ground truth. Additionally, $Delta_{1.25}$ indicates that 89.6% of pixels in the optimized depth are within ±25% of the ground-truth depth. This experiment showcases that although Unidepth produces scale shifts, our optimization yields a temporally stable depth aligned with the ground truth scene geometry.

4: Elaborate on the comparison of voidness between Gaussians across different scenes and with other methods.

We thank reviewer K6kc for the suggestions, we will include more visualization and analysis around the zoom-in and zoom-out viewpoint across various scenes, compared with the baseline 4DGS [6], and previous state-of-the-art methods RoDyGS [2] and MoSca [8].

---

[1] Yu-Lun Liu, Chen Gao, Andreas Meuleman et al. Robust Dynamic Radiance Fields.

[2] Yoonwoo Jeong, Junmyeong Lee, Hoseung Choi et al. RoDyGS: Robust Dynamic Gaussian Splatting for Casual Videos

[3] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis et al. UniDepth: Universal Monocular Metric Depth Estimation.

[4] Sili Chen, Hengkai Guo, Shengnan Zhu et al. Video Depth Anything: Consistent Depth Estimation for Super-Long Videos

[5] Qianqian Wang, Yifei Zhang, Aleksander Holynski et al. Continuous 3D Perception Model with Persistent State

[6] Zeyu Yang, Hongye Yang, Zijie Pan et al. Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting.

[7] Zhengqi Li, Richard Tucker, Forrester Cole et al. MegaSaM: Accurate, Fast, and Robust Structure and Motion from Casual Dynamic Videos

[8] Jiahui Lei, Yijia Weng, Adam Harley et al. MoSca: Dynamic Gaussian Fusion from Casual Videos via 4D Motion Scaffolds

Thank you for sharing your precious advice for our paper. Our discussions of the weaknesses and concerns are listed below.

Weakness:

1: More Details about the Grid Pruning Implementation

Our grid pruning implementation follows the voxel-based filtering approach in Open3D [1] and PointCloudUtils [2]. Specifically, each point is discretized into voxel coordinates (via integer division by voxel size) and inserted into a hash map using these coordinates as keys within constant time complexity. Points within the same voxel are grouped, averaging their positions and attributes (RGB, covariance) into one representative Gaussian. We will clarify this implementation explicitly in our revision.

2: Comparison with "4D Gaussian Splatting for Real-Time Dynamic Scene Rendering" [A]

The comparisons to "4D Gaussian Splatting" [A] have been presented in our manuscript, denoted as "4DGS [31]" in Table 1 and Table 3.

On the NVIDIA dataset (Table 1):

On the Dycheck dataset (Table 3):

We apologize for the confusion regarding the method naming. For clarity, we will rename the 4DGS [37] as "Real-time 4DGS" in the camera-ready version to avoid ambiguity.

3: Experiment on the Neu3D dataset.

We conducted experiments on Neu3D below:

---

The primary focus of our work is the reconstruction of scenes from casual monocular videos, characterized by a single camera undergoing free movement. The Neu3D [6] dataset, however, features static multi-camera setups without camera movement. Our method tries to handle a general set of casual videos that does not require synchronized multi-view videos. Thus, some of our designs like initialization strategy and Gaussian simplification formulation are incompatible with Neu3D's static multi-camera setup [6]. However, the performance of our method is still on-par with the most competitive methods.

We will also add this evaluation in the supplementary material for our camera-ready version.

4: Detailed Comparison on how the Proposed Techniques Improve Training Speed Relative to the Baseline method

We do analyze how these proposed techniques improve training speed through the ablation study in Table 2. Following your suggestion, we now refactor this table as shown below for better clarity. The two major techniques related to the training speed are grid pruning and the color design.

In the table below, the Baseline method shares similar methods with Real-time 4DGS. "+Pruning" method denotes that instead of starting from the sparse point cloud calibrated from COLMAP, we apply our initialization strategy without requiring densification. As shown in the table, one benefit is that all Gaussians are present at the beginning and can immediately receive gradient updates. The second change is the trade-off for spherical harmonics function design. In this ablation study, experiments show that, given dense point cloud initialization, direct RGB color is sufficient to reconstruct monocular dynamic scenes. It is also worth noting that a third-order spherical harmonics function requires 48 parameters. In comparison, other parameters e.g. (position, rotation, scaling) take 13 parameters in total.

Updated Ablation Study Table 2:

• Our final version applies both of these techniques and thus achieves 4x less runtime and 13x less memory requirement, compared with the baseline method 4DGS.
• We will follow your suggestions, update our Table 2 in the camera ready, and include a detailed, reader-friendly analysis on how the proposed techniques improve training speed relative to the baseline method.

5: Discussion on the contribution of our work.

Our motivation is rooted in the recent advances in deep visual SLAM systems, which have significantly improved speed [3], robustness [4] and generalization [5] in geometry recovery from casually captured monocular videos. Traditional 4DGS methods initialize from the sparse point cloud calibrated from COLMAP. From the deep visual SLAM system, we are able to acquire the point cloud sequence. However, we find the redundant gaussians greatly slow the overall training speed if we directly input the point cloud sequence into our optimization pipeline. To solve this problem, we apply a simple yet effective grid pruning method, which reduces the number of Gaussians by 92%. Experiments show that our method achieves a 30× speed-up and our method is the first to reduce optimization time for a casual video to minute-level, facilitating numerous downstream applications.

To further overcome challenges posed in the monocular setting, we also introduce an isotropic gaussian formulation, regularizing the expression and increasing the rendering quality by 1.52 dB. To solve the problem that the gaussians outside of the frustum tend to disappear, we design the motion-aware gaussian strategy, improving the render result by 3.41 dB PSNR compared with the baseline. Experiments show that our method achieves a 30× speed-up and reduces training time to within two minutes, facilitating numerous downstream applications.

6: Discussion on the COLMAP calibration and Visual SLAM calibration.

COLMAP is designed for static scenes from consistent videos. However, it commonly fails in dynamic and casually captured videos. Visual SLAM systems, in contrast, are more robust to motion blurring [4], fast to optimize [5] or feedforward [4] and avoid manual masking.

We provide the evaluation with COLMAP as calibration as shown below.

---

We also considered combining Visual SLAM's prediction with COLMAP's pose estimates. However, such pairing is rarely available in practice, as these systems operate independently and rely on different camera intrinsics, scales, and assumptions. WWe will clarify the grid pruning implementation explicitly in the revised manuscript, and we appreciate the opportunity to explain this more clearly.

Question:

1. Comparison with COLMAP calibration

COLMAP often outputs highly accurate camera poses, but at the same time requires more than 15 minutes optimization. When evaluating, the common practice is to use the COLMAP camera pose as ground truth. Therefore we argue that it is fair to compare with mainstream methods. It is also worth noting that for the runtime reported for COLMAP methods, we do not include the calibration time, while we include the calibration time in our method.

2. Evaluation on the Neu3D Dataset

Please refer to our discussion in weakness 3.

---

Reference

[1] Qian‑Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A Modern Library for 3D Data Processing.

[2] Francis Williams. Point Cloud Utils

[3] Jianing Yang, Alexander Sax, Kevin J. Liang et al. Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass

[4] Jianyuan Wang, Minghao Chen, Nikita Karaev et al. VGGT: Visual Geometry Grounded Transformer

[5] Zhengqi Li, Richard Tucker, Forrester Cole et al. MegaSaM: Accurate, Fast, and Robust Structure and Motion from Casual Dynamic Videos

[6] Tianye Li, Mira Slavcheva, Michael Zollhoefer et al. Neural 3D Video Synthesis from Multi-view Video

[7] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian et al. Mixed Neural Voxels for Fast Multi-view Video Synthesis.

[8] Liangchen Song, Anpei Chen, Zhong Li et al. NeRFPlayer: A Streamable Dynamic Scene Representation with Decomposed Neural Radiance Fields.

[9] Benjamin Attal, Jia-Bin Huang, Christian Richardt et al. HyperReel: High-Fidelity 6-DoF Video with Ray-Conditioned Sampling.

[10] Zeyu Yang, Hongye Yang, Zijie Pan et al. Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting.

Many thanks to the authors for their response regarding the main contribution. My concerns have been well addressed, and the rebuttal has clarified the method's contributions—specifically, the integration of deep visual SLAM into 4DGS reconstruction, the pruning of initial point clouds, and the introduction of the isotropic Gaussian formulation. The experimental results on the multi-camera Neu3D dataset further demonstrate the effectiveness of the proposed approach. I suggest that the authors include these details and experimental findings in the final version of the paper.

Additionally, I encourage the authors to provide a more thorough description of the implementation details for the Neu3D experiments, given that the method is originally designed for monocular videos.