acceptodds
Under review as a conference paper at ICLR 2027

TwinDropGS: Rethinking Dropout Regularization for Sparse-View Gaussian Splatting with Twin Consistency

Abstract

Sparse-view novel view synthesis is a fundamental yet challenging problem in 3D reconstruction, where limited observations often lead to severe overfitting and unreliable scene representations. Recent dropout-based 3D Gaussian Splatting methods alleviate this issue by randomly removing Gaussian primitives during training. However, these methods typically optimize only a single stochastic Gaussian subfield at each iteration, leaving the discrepancy across different dropout realizations unobserved and unconstrained. In this work, we rethink dropout regularization from a consistency-centric perspective and propose TwinDropGS, a Twin Dropout-induced Consistency Regularization framework for sparse-view Gaussian Splatting. To make dropout-induced stochastic discrepancy explicitly observable and regularizable, we first construct twin Gaussian subfields from a shared Gaussian field while ensuring complete joint Gaussian coverage. Building on it, we introduce an error-guided low-frequency consistency regularization that emphasizes regions with larger reconstruction errors and aligns their low-frequency luminance responses, avoiding excessive constraints on ambiguous high-frequency details. Additionally, a progressive complementary consistency scheduling strategy gradually increases both subfield complementarity and consistency strength throughout training, enabling stable structure formation in early optimization and stronger stochastic regularization in later stages. Extensive experiments demonstrate that TwinDropGS consistently improves reconstruction quality and training stability over existing dropout-based 3DGS methods while retaining a simple and plug-and-play design. The source code will be released at http://github.com.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Dropout Regularization, Consistency Regularization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_3wcsOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes TwinDropGS, a twin dropout-based consistency regularization framework for sparse-view 3D Gaussian Splatting. The method constructs two complementary Gaussian subfields and enforces error-guided low-frequency consistency between their renderings. A progressive scheduling strategy further adjusts dropout and consistency strength during training. Experiments on multiple benchmarks demonstrate consistent improvements in novel-view synthesis quality.

Critical strengths

S1. Clear Motivation. The paper identifies the lack of consistency constraints across different dropout realizations and provides a well-motivated solution.

S2. Simple and Effective Design. The proposed framework is straightforward and does not require additional pretrained models or external geometric priors.

S3. Consistent Performance Improvements. The method achieves consistent improvements across multiple sparse-view benchmarks, including LLFF, Mip-NeRF360, and Blender.

Critical weaknesses
  1. Significant Computational Overhead. TwinDropGS approximately doubles the training time compared with DropGaussian (151s vs. 74s, Tab 3). Although TwinDropGS-S reduces this overhead, its reconstruction quality is lower than that of the full model. The quality-efficiency trade-off remains a concern.

  2. Limited Comparison with Feed-forward Sparse-view Reconstruction Methods. The experiments focus primarily on optimization-based methods, without comparisons against recent feed-forward sparse-view reconstruction approaches. Including representative feed-forward baselines would provide a more comprehensive understanding of the method's performance and practical trade-offs

Questions for authors

Q1. Geometric Accuracy and Structural Consistency. Does enforcing rendering consistency between Gaussian subfields actually improve the underlying 3D geometry? Could the authors provide quantitative geometric evaluations or additional structural analyses to support this claim?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 3wcs

We sincerely thank you for recognizing the clear motivation, simplicity, and consistent performance improvements of our method. We also appreciate your constructive comments and suggestions. In response to your concerns, we have conducted additional experiments on geometric accuracy, computational efficiency, and comparisons with feed-forward methods. We provide detailed point-by-point responses below.

1. Geometry Accuracy and Structural Consistency.

TwinDropGS enforces consistency between two complementary subsets of the same Gaussian field, reducing structural instability caused by stochastic dropout. Since inaccurate geometric structures are more sensitive to different subsets, consistency regularization suppresses floating Gaussians, incorrect occlusions, and local geometric noise, thereby improving reconstruction quality. Following the evaluation protocol of Co-Adaptation, we further evaluate depth quality on LLFF 3-view. The results confirm TwinDropGS's effectiveness in improving geometric accuracy and structural consistency.

MethodMAE↓RMSE↓AbsRel↓
DNGaussian2.58024.27850.0732
DropGaussian2.17093.87440.0584
TwinDropGS (Ours)1.97883.74310.0516

2. Computational Cost and Efficiency.

Although TwinDropGS introduces additional training overhead compared with DropGaussian, it achieves a favorable quality-efficiency trade-off. Under matched training budgets, TwinDropGS-S (74s, 20.99 dB) outperforms DropGaussian (74s, 20.33 dB), while full TwinDropGS (151s, 21.20 dB) surpasses DropGaussian trained for 20k iterations (157.8s, 20.39 dB). Moreover, as shown below, TwinDropGS achieves superior reconstruction quality with substantially less training time than other competitive methods, including FSGS, CoR-GS, and NexusGS. Compared with the full TwinDropGS, TwinDropGS-S reduces training time by approximately 51% (74s vs. 151s), while maintaining competitive reconstruction quality with only a 0.21 dB decrease in PSNR.

MethodIter.PSNR↑SSIM↑LPIPS↓Time (s)↓
3DGS10K19.170.6460.26875
DropGaussian10K20.330.7090.20174
FSGS10K20.430.6820.248276
CoR-GS10K20.360.7100.202373
NexusGS30K21.070.7380.177473
NexusGS (10K)10K20.660.7290.181167
TwinDropGS-S5K20.990.7440.19274
TwinDropGS (Ours)10K21.200.7580.180151

3. Comparison with Feed-Forward Methods.

Following DropGaussian's evaluation protocol, we compare TwinDropGS with representative feed-forward methods on Replica 2-view. TwinDropGS achieves the best PSNR, SSIM, and LPIPS, demonstrating superior reconstruction quality. Although pre-trained feed-forward methods offer faster inference, TwinDropGS achieves superior reconstruction quality under sparse-view inputs. We have added the corresponding discussion and quantitative comparisons in Appendix A of the revised manuscript.

MethodTypePSNR↑SSIM↑LPIPS↓
pixelSplatFeed-forward26.240.8290.229
MVSplatFeed-forward26.160.8400.173
FreeSplatFeed-forward26.980.8480.171
DropGaussianOptimization27.760.8660.179
TwinDropGS (Ours)Optimization28.190.8780.170

Thank you for the detailed responses and additional experiments. The authors have adequately addressed my concerns regarding geometric accuracy, computational overhead, and comparisons with feed-forward methods. The additional results provide sufficient evidence to support the proposed method.

Therefore, I maintain my original rating.

AuthorsOpenReview ↗

Response to Reviewer 3wcs

Dear Reviewer,

Thank you for confirming that our additional experiments have adequately addressed your concerns. We would like to further emphasize that TwinDropGS goes beyond conventional dropout strategy design by rethinking sparse-view Gaussian optimization from a consistency-centric perspective. Rather than treating stochastic Gaussian subfields as independent training realizations, our framework makes dropout-induced discrepancies explicitly observable and regularizable, establishing a new perspective on dropout-based regularization. Importantly, this simple, prior-free, and plug-and-play design achieves consistent improvements in reconstruction quality, geometric accuracy, and training stability, with additional compute-matched experiments confirming that the gains are not merely due to increased computation. More broadly, we believe this consistency-centric perspective offers a generalizable regularization principle with potential applications beyond sparse-view reconstruction, such as dynamic scene reconstruction and other 3D representation learning tasks.

We sincerely appreciate your valuable suggestions and continued support. Please let us know if you have any remaining questions or suggestions. We would greatly appreciate your consideration of a more favorable score.

Thank you again for your valuable time and consideration.

Best regards,

The Authors

Reviewer_WVDJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TwinDropGS, a dropout-based regularizer for sparse-view 3DGS. Instead of rendering a single randomly dropped subset of Gaussians per iteration, it renders two complementary subsets of the same Gaussian field and penalizes the difference between their low-frequency luminance, with more weight on pixels with larger reconstruction error. The dropout ratio and the consistency weight increase over training. The method improves over existing dropout-based methods on LLFF, Mip-NeRF360 and Blender, and it can be added to other sparse-view 3DGS methods.

Critical strengths
  • A simple and well-motivated idea. Regularizing the agreement between two dropout realizations is a natural extension of dropout-based sparse-view 3DGS. It requires no external priors or auxiliary networks.
  • Consistent improvements. The method improves over its baseline on all three benchmarks and view settings. It also improves other sparse-view 3DGS methods when added to them, which supports its plug-and-play claim, and it reduces the variance of results across random seeds.
  • Clear presentation. The paper is well organized and easy to follow. Each component is motivated by a specific question: what to match, where to emphasize, and when to regularize.
Critical weaknesses
  • The main baseline is reported below its published results, which makes the improvement appear larger. DropGaussian is both the codebase and the most direct baseline, so its numbers determine how much the proposed components add. In all four settings of Tables 1–2, the DropGaussian results are lower than those in its original paper (Park et al., 2025). Against the published results, the gain of TwinDropGS on LLFF 3-view is roughly halved, and on Mip-NeRF360 24-view and Blender it shrinks to about 0.1 dB, comparable to the seed-to-seed variation of DropGaussian reported in the appendix (Table 8). A reproduction gap is not unusual, but the paper does not mention it, and it is unclear whether the other baselines were re-run or copied from their papers. Please state the source of each baseline result, report the published DropGaussian results alongside the reproduced ones, and explain the difference.

  • The LLFF comparison omits two recent methods that the paper itself relies on. Table 1 includes neither D2GS (Song et al., ICLR 2026) nor NexusGS (Zheng et al., CVPR 2025), although both report LLFF 3-view results. The paper follows D2GS for its training pipeline and multi-seed protocol (l. 185, 737) and includes NexusGS in Table 2. Both methods use pretrained priors, so a prior-free method need not outperform them, but Table 1 already includes prior-based methods such as DNGaussian and FSGS. Their reported results are close to those of TwinDropGS: D²GS is higher in PSNR, and both are slightly better in LPIPS. This would be a favorable result for a prior-free method, but the current Table 1, which marks TwinDropGS as best on every metric, overstates its standing. Please add both methods, indicate which methods use priors.

  • The ablation does not separate the effect of the consistency loss from the added computation. In Table 4, the twin-branch variant trained without the consistency loss already accounts for more than half of the total gain. This variant only averages the reconstruction loss over two dropout masks per iteration, at nearly twice the training time. Much of the gain may therefore come from extra computation and lower-variance gradients rather than from consistency, which weakens the consistency-centric framing. TwinDropGS-S improves on DropGaussian at equal training time, but it is not ablated, so the source of its gain remains unclear. Please repeat the Table 4 ablation under the shortened schedule, and report DropGaussian trained with compute matched to the full TwinDropGS.

References

  • Park, H., Ryu, G., and Kim, W. DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting. CVPR 2025.
  • Song, M., Lin, X., Zhang, D., Li, H., Li, X., Du, B., and Qi, L. D2GS: Depth-and-Density Guided Gaussian Splatting for Stable and Accurate Sparse-View Reconstruction. ICLR 2026.
  • Zheng, Y., Jiang, Z., He, S., Sun, Y., Dong, J., Zhang, H., and Du, Y. NexusGS: Sparse View Synthesis with Epipolar Depth Priors in 3D Gaussian Splatting. CVPR 2025.
Questions for authors
  • In Eq. 8, the gradient is stopped through branch B, so only the Gaussians rendered in branch A receive gradients from the consistency loss in each iteration. What is the motivation for this one-sided design rather than a symmetric loss, and how does a symmetric version compare?
5 comments
AuthorsOpenReview ↗

Response to Reviewer WVDJ

We sincerely thank you for recognizing the simplicity, sound motivation, consistent performance improvements, and clear presentation of our method. We also appreciate your constructive suggestions and address your concerns point by point below, with additional experiments and clarifications.

Q1: Clarification of Baseline Results.

(1) DropGaussian results.

According to the official DropGaussian implementation and the authors' clarification from its official GitHub repository, the results reported in the original paper were selected as the best performance across multiple independent runs (typically 5 runs). In contrast, following the evaluation protocol of DropAnSH-GS, all results in our paper are reported as the average over three independent runs. This difference in statistical reporting explains the discrepancy between the published results and our reproduction. For transparency, we provide both sets of results as follows.

MethodLLFF 3-viewLLFF 6-viewLLFF 9-viewMip-NeRF360 12-viewMip-NeRF360 24-viewBlender 8-view
DropGaussian (best of 5 runs)20.7624.7426.2119.7424.1325.42
DropGaussian (mean of 3 runs)20.3324.5825.8519.5723.7125.19

(2) More baseline results.

For baselines with publicly available implementations, we consistently report the average performance over three independent runs under the same evaluation protocol. For methods without publicly available implementations, we use the results reported in their original papers.

Q2: Comparisons with D²GS and NexusGS.

We have added both methods and explicitly indicated their external pretrained priors as follows. For completeness, we also evaluated D²GS using its official implementation following our evaluation protocol of averaging three independent runs. TwinDropGS achieves competitive performance without relying on pretrained models. These results demonstrate that effective consistency regularization in our TwinDropGS can alleviate the reliance on external pretrained priors in sparse-view reconstruction.

MethodExternal PriorPSNR↑SSIM↑LPIPS↓
NexusGSOptical flow21.070.7380.177
D²GS (Reported)Depth21.350.7460.179
D²GS (Reproduced)Depth20.780.7280.191
TwinDropGS (Ours)None21.200.7580.180

Q3: Consistency Regularization vs. Computational Cost.

(1) Motivation.

The dual-branch architecture serves as the foundation for consistency learning. Specifically, dual-branch reconstruction optimizes the reconstruction quality of each subfield individually, while consistency regularization explicitly enforces structural consistency between them, constituting two distinct yet complementary optimization mechanisms. Our consistency regularization with progressive scheduling provides an additional 0.41 dB PSNR gain (20.79 to 21.20 dB) with only 23s of additional training time, highlighting its crucial role in balancing reconstruction quality and computational cost.

(2) Shortened-schedule ablation.

We conduct the same ablation experiments for TwinDropGS-S, which show that it retains the same sources of gain as the full framework.

Two-B.EG-LF Cons.Prog. S.PSNR↑SSIM↑LPIPS↓Avg. Time(s)↓
×××20.260.6870.21835
✓××20.600.7170.20668
✓✓×20.870.7400.19975
✓✓✓20.990.7440.19274

(3) Compute-matched comparison.

TwinDropGS-S (5k) already outperforms DropGaussian (10k) at comparable training time (20.99 vs. 20.33 dB). We further extend DropGaussian to 20k iterations to match the computational budget of full TwinDropGS. DropGaussian improves by only 0.06 dB despite longer training (157.8s vs. 151.0s), confirming that our gains cannot be explained by additional computation alone.

MethodIter.Time(s)PSNR↑SSIM↑LPIPS↓
DropGaussian10k74.020.330.7090.201
DropGaussian20k157.820.390.7140.203
TwinDropGS-S5k74.020.990.7440.192
TwinDropGS10k151.021.200.7580.180

Q4. One-Sided vs. Symmetric Consistency.

Symmetric consistency allows both branches to simultaneously adapt to each other, potentially suppressing informative structural details in pursuit of agreement. In contrast, our one-sided design applies stop-gradient to one branch only within the consistency loss, while both branches remain supervised by their respective reconstruction losses. This avoids mutual adaptation through the consistency loss, discouraging convergence to overly smoothed solutions. Our experiments further validate the effectiveness of this design.

StrategyPSNR↑SSIM↑LPIPS↓
Symmetric21.090.7520.188
One-sided (Ours)21.200.7580.180
AuthorsOpenReview ↗

Response to Reviewer WVDJ

Dear Reviewer,

We hope this message finds you well. As the discussion period is nearing its end with less than two days remaining, we would like to ensure that we have addressed all of your concerns satisfactorily. If you find our responses satisfactory, we would greatly appreciate it if you could consider raising your score in recognition of our efforts. If there are any additional points or feedback, please let us know. Your insights are invaluable to us, and we are eager to address any remaining issues to further improve our work.

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

Sincerely,

The Authors

AuthorsOpenReview ↗

Response to Reviewer WVDJ

Dear Reviewer,

We sincerely appreciate your valuable time and effort in reviewing our paper. We hope that our detailed responses and additional experiments have satisfactorily addressed your concerns. If so, we would be very grateful if you could kindly consider updating your rating. Should you have any remaining concerns, questions, or suggestions, please do not hesitate to let us know. We would be more than happy to provide further clarifications during the remaining discussion period.

Thank you again for your time and consideration.

Best regards,

The Authors

Reviewer_dAG7OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper rethinks the limitations of existing dropout-based sparse-view reconstruction methods and proposes a Twin Dropout-induced Consistency Regularization framework for sparse-view Gaussian Splatting. TwinDropGS constructs twin stochastic Gaussian subfields from a shared Gaussian field and explicitly regularizes their consistency for more stable and robust sparse-view reconstruction. The method is simple but effective. Experiments demonstrate performance improvements on commonly used sparse-view reconstruction datasets.

Critical strengths
  1. The method is simple yet effective. It constructs twin stochastic Gaussian subfields from a shared Gaussian field and explicitly regularizes their consistency, a sound method that leads to more stable and robust sparse-view reconstruction.

  2. The experiments in the paper are relatively comprehensive and demonstrate the effectiveness of the proposed method.

  3. The presentation of the paper is good. It is easy to read, and the framework figure is clear and easy to understand.

Critical weaknesses
  1. The experimental settings could be further improved.
  • The number of training iterations can significantly affect sparse-view per-scene optimization. Instead of simply following previous settings, it would be better to explore different numbers of training iterations (e.g., 20k and 30k) and compare them with baselines under the same settings.
  • It is unclear which baseline results are reproduced and which are taken from previous papers. The sources of these results should be clarified. Since per-scene optimization is relatively inexpensive, I suggest rerunning closely related baselines, such as DropGaussian, on the same machine to ensure a fair comparison.
  1. The paper lacks a comparison with feed-forward methods. Although related work is discussed in the supplementary material, it does not include a quantitative comparison with feed-forward methods, including reconstruction quality, speed, and GPU memory usage. Although I understand that feed-forward methods belong to a different paradigm, such a comparison would help readers better understand the trade-offs inherent in sparse-view reconstruction across different paradigms.
Questions for authors
  1. The impact of the number of iterations. I am curious whether the conclusion of the paper would change if the baselines (Vanilla 3DGS and DropGaussian) and the proposed method were trained for more iterations, for example, 20k or 30k. Could you provide a comparison plot similar to Figure 4 that includes both baseline methods and the proposed method, and extends the training duration to 30k iterations?

  2. How does this compare to feed-forward methods? Also, could you include a simple baseline that uses VGGT point clouds for initialization (instead of SfM), and then performs optimization using the original 3DGS?

  3. What method does "PCDropGS" in Fig. 1(c) refer to? There might be a typo?

4 comments

any feedback on my questions?

AuthorsOpenReview ↗

Response to Reviewer dAG7

We sincerely apologize for our delayed response and for keeping you waiting. We greatly appreciate your patience and follow-up. Thank you for recognizing the simplicity and effectiveness of our method, the comprehensive experiments, and the clarity of our presentation. We also appreciate your constructive suggestions. We have carefully addressed your questions and provide detailed point-by-point responses below.

1. Training Iterations and Convergence.

Following your suggestion, we extend Vanilla 3DGS, DropGaussian, and TwinDropGS to 20K and 30K iterations under the same experimental settings on LLFF 3-view. As shown in the table below, TwinDropGS consistently outperforms both baselines across all training budgets. Notably, at 30K iterations, TwinDropGS maintains 21.21 dB PSNR, while DropGaussian decreases to 20.18 dB, suggesting that its stochastic dropout regularization may become less stable under prolonged optimization. Furthermore, Fig. 4(b) in the main manuscript presents the training PSNR curves on the challenging fortress scene on LLFF-3view. TwinDropGS converges rapidly and maintains consistently higher PSNR with smaller fluctuations, whereas DropGaussian and Vanilla 3DGS plateau at lower performance levels. These results demonstrate the effectiveness and training stability of TwinDropGS across different optimization budgets.

MethodIter.PSNR↑SSIM↑LPIPS↓
3DGS10K19.170.6460.268
20K19.500.6590.251
30K19.430.6570.259
DropGaussian10K20.330.7090.201
20K20.390.7140.203
30K20.180.6990.214
TwinDropGS10K21.200.7580.180
20K21.230.7570.179
30K21.210.7560.181

2. Baseline Reproduction and Fair Comparison.

For all experiments reported in the main manuscript, we follow consistent dataset splits, input-view configurations, and evaluation protocols to ensure fair comparisons. For baselines with publicly available code, we reproduce their results under the same experimental settings and report the average over three independent runs to reduce randomness. For methods without released code, we directly adopt the results reported in their original papers under the corresponding evaluation settings.

3. Comparison with Feed-Forward Methods.

Following the evaluation protocol of DropGaussian, we compare TwinDropGS with representative feed-forward methods, including pixelSplat, MVSplat, and FreeSplat, on Replica 2-view. As shown below, TwinDropGS achieves the best PSNR, SSIM, and LPIPS, demonstrating its superior reconstruction quality. Regarding computational efficiency, feed-forward methods offer fast inference without per-scene optimization, whereas TwinDropGS requires scene-specific optimization (151s on a single A800 GPU). Although feed-forward methods have advantages in inference speed, TwinDropGS achieves higher reconstruction accuracy, highlighting the trade-off between reconstruction quality and computational efficiency across different paradigms.

MethodTypePSNR↑SSIM↑LPIPS↓
pixelSplatFeed-forward26.240.8290.229
MVSplatFeed-forward26.160.8400.173
FreeSplatFeed-forward26.980.8480.171
DropGaussianOptimization27.760.8660.179
TwinDropGS (Ours)Optimization28.190.8780.170

4. VGGT-Based Initialization.

Following your suggestion, we additionally evaluate Vanilla 3DGS initialized with VGGT-predicted point clouds instead of SfM points on the LLFF 3-view dataset, while keeping the remaining training settings unchanged. As shown below, VGGT initialization achieves 18.96 dB PSNR, slightly lower than the 19.17 dB obtained with SfM initialization, suggesting that VGGT-based initialization does not necessarily improve reconstruction quality in texture-rich scenes under sparse-view supervision. Unlike approaches that focus on point-cloud initialization, TwinDropGS improves the optimization process through simple yet effective consistency regularization, achieving 21.20 dB PSNR without relying on additional geometric priors. These results highlight the effectiveness of our optimization-oriented design for sparse-view reconstruction.

MethodInitializationPSNR↑SSIM↑LPIPS↓
3DGSSfM19.170.6460.268
3DGSVGGT18.960.6340.269
TwinDropGSSfM21.200.7580.180

5. Typo in Fig. 1(c).

Thank you for pointing out this typo in our manuscript. We have corrected “PCDropGS” to “TwinDropGS” in Fig. 1(c) of the revised manuscript.

Thanks for your response, my concerns are solved, so I will keep my positive rating.

Dear Reviewer,

Thank you very much for revisiting our response and for your continued positive assessment. We sincerely appreciate your careful reading, constructive feedback, and recognition of our work. We are glad that our additional experiments and clarifications have satisfactorily addressed your concerns and strengthened the paper.

Thank you again for your valuable time and thoughtful evaluation!

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Contributions, Rebuttal, and Reviewer Feedback

Dear AC,

We sincerely thank you and the reviewers for the constructive feedback. We have carefully addressed all concerns through additional experiments, clarifications, and manuscript revisions. Below, we first highlight the core contributions of TwinDropGS, followed by a summary of our responses to the reviewers' concerns.

Core Contributions of TwinDropGS

We would like to emphasize that TwinDropGS goes beyond conventional dropout strategy design by rethinking dropout regularization from a consistency-centric perspective, making dropout-induced discrepancies explicitly observable and regularizable. Importantly, this simple, prior-free, and plug-and-play framework consistently improves reconstruction quality, geometric accuracy, and training stability, with compute-matched experiments confirming that the gains are not merely due to increased computation. More broadly, we believe this consistency-centric perspective has potential applications beyond sparse-view reconstruction, including dynamic scene reconstruction and other 3D representation learning tasks.

Summary of Rebuttal and Revisions

1. Baseline Fairness and Comparisons (WVDJ, dAG7)

  • Clarification: Explained the discrepancy between DropGaussian's published best-of-multiple-runs results and our mean-of-three-runs evaluation.
  • New Comparisons: Provided published and reproduced DropGaussian results, and added D²GS and NexusGS with their external priors explicitly identified.

2. Consistency Effectiveness and Computational Efficiency (WVDJ, 3wcs)

  • New Experiments: Conducted shortened-schedule component ablations, compute-matched comparisons, and one-sided versus symmetric consistency evaluations.
  • Results: TwinDropGS-S outperforms DropGaussian by 0.66 dB under identical training time. TwinDropGS outperforms compute-matched DropGaussian by 0.81 dB, demonstrating that the improvements cannot be attributed merely to additional computation.
  • Paper Update: Expanded Table 4 and added Table 6 in Sec. 4.5.

3. Training Stability and Extended Optimization (dAG7)

  • New Experiments: Extended Vanilla 3DGS, DropGaussian, and TwinDropGS to 20K and 30K iterations under identical settings.
  • Results: At 30K iterations, TwinDropGS achieves 21.21 dB, compared with 20.18 dB for DropGaussian and 19.43 dB for Vanilla 3DGS, demonstrating stable convergence and sustained reconstruction advantages.
  • Paper Update: Added Fig. 4(b) and corresponding analysis in Sec. 4.5.

4. Geometric Accuracy and Structural Consistency (3wcs)

  • New Experiment: Evaluated depth reconstruction quality on LLFF 3-view.
  • Results: TwinDropGS achieves the lowest MAE (1.9788), RMSE (3.7431), and AbsRel (0.0516), outperforming DNGaussian and DropGaussian.
  • Paper Update: Added Table 10 and corresponding analysis in Appendix B.3.

5. Feed-forward Comparisons (dAG7, 3wcs)

  • New Experiments: Compared TwinDropGS with pixelSplat, MVSplat, and FreeSplat on Replica 2-view.
  • Results: TwinDropGS achieves the best reconstruction quality on Replica 2-view (28.19 dB PSNR, 0.878 SSIM, and 0.170 LPIPS).
  • Paper Update: Expanded Appendix A and added quantitative comparisons in Appendix B.4 (Table 11).

Reviewer Feedback

We are encouraged that all three reviewers (dAG7, 3wcs, and WVDJ) have consistently acknowledged that their concerns have been satisfactorily addressed and expressed positive assessments of our work. Reviewers dAG7 and 3wcs maintained their positive ratings, while Reviewer WVDJ kindly raised the initial rating to positive following our detailed responses, clarifications and additional experiments.

We believe these additional experiments and revisions substantially strengthen the technical contributions and experimental evidence supporting TwinDropGS. We respectfully hope that the AC will consider the reviewers' positive feedback and our revisions in the final assessment.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.