acceptodds
Withdrawn from ICLR 2027

GASplat: Geometry-aware Distractor-free 3D Gaussians Splatting Under Sparse-View Settings

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

3D Gaussian Splatting (3DGS) achieves real-time photorealistic 3D static scene reconstruction, yet its performance degrades significantly when sparse input images also contain dynamic transient objects. Existing distractor-free reconstruction methods misclassify static hard-to-reconstruct regions as distractors, sparse-view reconstruction methods don't block gradients of distractors and their vanilla combination enhances both static and transient regions. To tackle this new problem, we propose GASplat, a novel framework reconstructs static scene with sparse input containing distractors. Specifically, with aim of better distinguishing static and transient regions to identify distractors with memorization effect, we generate coarse masks through pretrained image matching model and geometry-aware filtering. Then, we enhance initial point clouds of high confident static regions with coarse masks. Next, we recover point clouds and masks of far background regions mitigating limitations of pretrained model. Finally, in order to precisely block gradients from transient objects and refine distractors masks, we design a progressive training schedule to reconstruct static scenes. Extensive experiments on distractors datasets under sparse view settings demonstrate that our approach can detect transient objects given sparse input images, outperforming current state-of-the-art methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Scene Reconstruction, Distractor-free, Sparse-view, Epipolar geometry

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qd9POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper appears to violate the official ICLR formatting guidelines by modifying the margins to fit more content.

A 3D Gaussian Splatting (3DGS) reconstruction framework specifically designed for challenging real-world scenes with sparse input views and transient dynamic distractors.

Critical strengths
  1. Novel Problem Setting: The paper addresses 3DGS reconstruction in challenging real-world scenes with both sparse input views and transient distractors. This is a practical problem that existing distractor-removal and sparse-view reconstruction methods may struggle to handle simultaneously.

  2. Adaptive Distractor Suppression: Instead of using manually tuned thresholds, the method employs a Gaussian Mixture Model (GMM) to generate static and dynamic pseudo-labels from RGB residuals and uncertainty estimates. An MLP then predicts distractor masks at the pixel level, while a progressive masking strategy at different training stages helps reduce the impact of transient objects on gradient updates.

Critical weaknesses
  1. Heavy Reliance on Pre-trained Models: The pipeline depends on multiple pre-trained models, including image matching, Depth Anything 3, DINOv3, and SAM3. Errors in challenging scenarios, such as extreme lighting, severe occlusion, or unusual viewpoints, may propagate through mask generation and pseudo-label construction, compromising reconstruction quality.

  2. High Computational Cost and Pipeline Complexity: The extensive use of 2D foundation models introduces substantial overhead. Feature extraction, mask generation, and background detection alone take approximately 77 seconds for 12 views, excluding the subsequent 3DGS optimization. This limits the method’s practicality in resource-constrained or time-sensitive settings.

  3. Limited Robustness to Illumination Changes: The method assumes that the appearance of static objects and environmental illumination remain largely unchanged. This assumption may not hold for images captured under different lighting or weather conditions, limiting its robustness in diverse in-the-wild scenarios.

Questions for authors

Beyond the weaknesses mentioned above, could the authors provide additional video visualizations to better demonstrate the method’s robustness, particularly under challenging conditions such as extreme lighting, rapid motion, and other difficult scenarios?

1 comment
AuthorsOpenReview ↗

Response to Reviewer qd9P

We thank Reviewer qd9P for your thoughtful comments and spending time to review our paper. We respond to your questions point by point as follows.

Formatting issue.

A: We sincerely apologize for it. In fact, this was an unintentional formatting error due to some incorrect LaTeX preamble settings that we had overlooked rather than a deliberate attempt to accommodate more content. We have fixed problems in the revised manuscript which meets the requirements of ICLR formatting.

[W1] Discussion on Pre-trained Models.

A: We acknowledge that these challenging scenarios are meaningful but they are out of scope in this paper, which focuses on distractor-free sparse-view 3DGS reconstruction. Extreme lighting [1], severe occlusion [2], or unusual viewpoints [3] can be individually treated as a research direction, which is difficult to let one paper concern all the extreme cases.

In addition, our method has a certain degree of robustness under such extreme conditions. Rather than directly relying on pretrained model as fixed supervision, our coarse-to-fine progressive mask generation framework incorporates a subsequent mask refinement stage, allowing inaccurate predictions to be progressively corrected during optimization.

[W2] Computational Cost.

A: In the appendix of our manuscript, the 77 seconds covers the entire reconstruction pipeline including the 3DGS optimization described in Sec. 3.3. The 3DGS optimization takes 59 seconds accounting for approximately 76% of total runtime, while the preprocessing module only takes about 18 seconds accounting for about 24% of total runtime. Hence, the preprocessing module introduces relatively little computational overhead compared with the 3DGS optimization.

[W3] Illumination Changes scenarios.

A: We think it is reasonable to mainly focus on removing distractors and assume there is no huge environmental illumination changes in our paper. Existing studies on 3D reconstruction in the wild can be categorized into two main research directions: removing distractor [4, 5] and handling illumination changes [6, 7]. Our work falls into the former one.

In addition, our method exhibits some robustness to illumination variations. We assign each image a learnable embedding which is fed into the mask prediction MLP. It can learn the appearance of each image including different lighting and thus alleviate the influence of illumination changes.

[Q1] Video visualizations.

A: We have now provided videos of reconstructed 3D scenes in the supplementary material.


References:

[1] Cui, Ziteng, Xuangeng Chu, and Tatsuya Harada. "Luminance-gs: Adapting 3d gaussian splatting to challenging lighting conditions with view-adaptive curve adjustment." CVPR 2025.

[2] Liu, Shiyong, et al. "Occlugaussian: Occlusion-aware gaussian splatting for large scene reconstruction and rendering." ICCV 2025.

[3] Li, Jiahe, et al. "Evpgs: Enhanced view prior guidance for splatting-based extrapolated view synthesis." CVPR 2025.

[4] Fu, Chuanyu, et al. "Robustsplat: Decoupling densification and dynamics for transient-free 3dgs." ICCV 2025.

[5] Wang, Xu, et al. “DualSplat: Robust 3D Gaussian Splatting via Pseudo-Mask Bootstrapping from Reconstruction Failures.” CVPR 2026.

[6] Cui, Ziteng, Xuangeng Chu, and Tatsuya Harada. "Luminance-gs: Adapting 3d gaussian splatting to challenging lighting conditions with view-adaptive curve adjustment." CVPR 2025.

[7] Zhang, Dongbin, et al. "Gaussian in the wild: 3d gaussian splatting for unconstrained image collections." ECCV 2024.

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

The paper targets sparse-view 3DGS in scenes with transient distractors. It uses RoMa v2 matching with epipolar and cycle-consistency checks to find static regions and triangulate an initial point cloud, fills distant areas with Depth Anything 3, and refines masks using DINOv3/SAM3 features, uncertainty and GMM pseudo-labels. It is evaluated on NeRF On-the-go and RobustNeRF at five view counts against sparse-view and distractor-free baselines, and most metrics improve clearly.

Critical strengths
  1. A reasonable design for a real problem. Sparse views produce large residuals in static but hard regions, which confuses residual-based distractor detection. Building a geometric static initialization first and then refining masks is, I think, a well-targeted fix and indeed a valuable research topic.
  2. Broad evaluation with clear gains. Two datasets, five view counts and ten baselines cover a good range. On On-the-go with 12 views, PSNR is 16.92 versus 14.57 for the next best method.
Critical weaknesses
  1. Design gains are not separated from strong priors and training protocol. The method uses RoMa v2, DA3, DINOv3 and SAM3, while the baselines do not. RobustSplat gets only 6.96 dB on On-the-go with six views, below plain 3DGS at 12.18 dB, which needs an explanation of its budget and tuning, although this alone does not prove a reproduction error. Table 6 builds its combined comparison mainly on this weak baseline, so I can't confirm the advantage over properly configured baselines.
  2. Some key assumptions and reported results are not well checked. The paper does not say whether camera poses come from all images or only the sparse inputs, mask quality is not measured, and geometric checks cannot catch distractors that stay still or move along epipolar lines. Without per-scene results or variance over view sampling, it is hard to judge how stable the gains are. The text also states a 0.88 dB drop for removing progressive training while Table 5 shows 0.40 dB, and the "best on all metrics" claim fails for 36-view LPIPS in Table 2.
  3. Importantly, no supplementary video for a 3D reconstruction paper. The main failure modes of sparse-view 3DGS, such as floaters, wrong geometry, view-inconsistent artifacts and leftover distractors, show up when the camera moves. I don't think they can be judged from a few selected still images. With no video and no project page, I have no way to check the 3D consistency behind the metrics. I see this as a major weakness for a 3D paper.
Questions for authors

See major weakness. Why is RobustSplat below plain 3DGS at six views? How are camera poses and intrinsics obtained, do they use views outside the training input, and does the epipolar convention in Eq. 6 match the code? Could you add a baseline with matching priors, report quantitative mask quality, and provide novel-view videos next to the baselines? Please also explain the 0.88 versus 0.40 dB gap and the 20.27 versus 20.21 average in Table 5, and say whether view sampling was repeated. Finally, DroneSplat (CVPR 2025) already handles distractors with limited views. How does your setting differ?

1 comment
AuthorsOpenReview ↗

Response to Reviewer iTko

We thank Reviewer iTko for your detailed comments and spending time to review our paper. We respond to your questions point by point as follows.

[W1-1] Discussion of priors and comparison with baseline.

A: The baselines also use pretrained model, RobustSplat uses DINOv2 to extract features and OCSplat uses SAM2 and DINOv2 as prior. We have conducted following experiments to show the influence of RoMa v2 and DA3:

Table 1: Quantitative PSNR comparisons on corner scene in NeRF On-the-go dataset

method6-view9-view12-view24-view36-view
RobustSplat6.969.4710.8516.9719.75
RobustSplat + Roma v2 & DA313.3514.1414.5815.7416.87
ours13.7415.7116.9219.7220.59

As shown in the table, vanilla combination of strong prior with baseline cannot get satisfactory results.

[W1-2 & Q1] RobustSplat vs 3DGS in experiments.

A: We have checked our implementation and ensure that it is not a reproduction error. RobustSplat assumes the dense input, while the sparse input lets the algorithm cannot distinguish the distractor regions and hard-to-reconstruct static regions using memorization effect in the early stage (line 69-70 in manuscript). This leads the mask prediction MLP tend to predict 00 and insufficient optimization results in inaccurate scene reconstruction, which in turn causes the MLP to incorrectly assign almost all zero values. RobustSplat cannot be optimized due to zero masks and 3DGS at least can be optimized by part of static regions which is the reason why RobustSplat underperform than original 3DGS under sparse view setting.

[W2-1 & Q2] Key assumptions and epipolar convention.

A: We follow the assumption of previous papers [1,2] that the camera poses are known from the dataset which are estimated from all images. The mask quality is hard to measure due to no ground truth mask from the dataset, we indirectly evaluate mask quality by novel view rendering performance. The mentioned scenarios are good points but distractors in real-world typically appear randomly, sparse-view observations make it more difficult to capture their complete motion trajectories. We have checked our implementation code that the epipolar convention in Eq. 6 match the code. As we haven't uploaded any source codes, we don't know the meaning of your question about epipolar convention.

[W2-2 & Q4] Results analysis.

A: We have rerun the algorithm for 3 times as follows:

Table 1: Quantitative PSNR comparisons on NeRF On-the-go dataset

method6-view9-view12-view24-view36-view
Run 113.7715.7016.8519.8720.65
Run 213.6615.7816.8719.8520.59
Run 313.7415.8516.8719.6620.61
mean13.7215.7816.8619.7920.62
std0.050.080.010.110.03

As showed above, our method have stable results. We apologize for inaccuracies caused by calculation and typographical mistakes. We have carefully checked all the numerical values of tables in the paper. The table 1-4 are all correct, the PSNR of full model in table 5 and 6 should be 20.2120.21. The drop in the line 516 should be 0.400.40 dB. In line 477-478 of the manuscript, we only claim that our method achieves the best performance in 6-/9-/12-/24-view comparing with sparse-view reconstruction methods rather than "best on all metrics". We achieve best results on PSNR and SSIM under 36-view setting, only 0.010.01 higher value in LPIPS metric which don't affect the overall effectiveness of our approach. We don't repeat the view sampling to produce the sparse dataset followed by previous work [1,2].

[W3 & Q3] Supplementary video.

A: We have now provided videos of reconstructed 3D scenes in the supplementary material.

[Q5] Comparison with DroneSplat (CVPR 2025).

A: We summarize three key differences between our method and DroneSplat as follows:

Table 1: Comparisons with DroneSplat

methodInput assumptionsInitialization strategyOptimization pipeline
DroneSplatDrone imageryPretrained DUSt3R with downsamplingSAM2 to compute the average residual for each segmented object
OursGeneral nature imageMGCMT with distractors filteringCoarse-to-fine masking strategy in pixel level

References:

[1] Zhang, Jiawei, et al. "Cor-gs: sparse-view 3d gaussian splatting via co-regularization." ECCV 2024.

[2] Wang, Xu, et al. “DualSplat: Robust 3D Gaussian Splatting via Pseudo-Mask Bootstrapping from Reconstruction Failures.” CVPR 2026.

Reviewer_UYBCOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents GASplat, a framework for sparse-view distractor-free 3DGS optimization. The proposed framework first enhances initial point cloud and generates coarse mask using pretrained matcher and depth estimation model. Then 3d gaussians and mask predictor are optimized during progressive reconstruction and mask refinement process. The paper reports stronger performance on sparse-view distractor-free 3DGS settings than previous methods.

Critical strengths
  • Challenging and meaningful problem setup: most distractor-free 3DGS methods leverage SfM point clouds (which mainly consists of static scene components) produced by COLMAP with dense views, so I think it is important to investigate distractor-free reconstruction in settings where the initial point cloud is insufficient.
  • the proposed method is well engineered and outperforms sparse-view methods and distractor-free methods, and selected combination of those methods.
Critical weaknesses
  • The proposed MGCMT module and the subsequent training pipeline seem to be the integration of existing components such as epipolar geometry, matching, and mask prediction procedure, which may limit its methodological novelty.
  • The method relies on several pretrained models such as Roma v2, Depth Anything 3, DINOv3, and SAM3. It is unclear how each component contributes to the final performance.
  • While the paper compares against both sparse-view and distractor-free 3DGS methods, the evaluations do not separately assess sparse-view and distractor-free performance. It is unclear how much each contributes to the performance gains.
  • Qualitative results in Fig. 1 and Fig. 4 still shows significant artifacts caused by distractors.

For minor concerns:

  • Missing citations on recent works: for sparse-view 3dgs, SE-GS (ICCV'25), DDGS (ICLR'26) / for distractor-free 3dgs, DeSplat(CVPR'25), HybridGS(CVPR'25), AsymGS(NeurIPS'25), PDF-GS(CVPR Findings'26), 3DGS-HPC(ICML'26)
  • As pointed out by Reviewer iTko, The paper uses wider text area than the original template, and I think it would exceed the page limit if put in the original text width. I would also like to discuss this issue with other reviewers and AC.
Questions for authors
  • Could you clarify the SfM setting used in the experiments? Previous distractor-free 3DGS methods typically use COLMAP sparse reconstruction. How does the quality of the point cloud generated by the MGCMT module compare to that of COLMAP dense reconstruction?
  • How does the model perform in sparse-view benchmarks without distractors (i.e. LLFF, Mip-NeRF 360 in sparse-view setting) against previous sparse-view 3DGS methods?
1 comment
AuthorsOpenReview ↗

Response to Reviewer UYBC

We thank Reviewer UYBC for your constructive feedback and spending time to review our paper. We respond to your questions point by point as follows.

[W1] Integration of components.

A: Our method is not simple integration of existing components. Epipolar geometry is for checking multiview consistency, matching is to obtain enhanced point clouds and mask prediction procedure is for removing impact of distractors. The engineering-oriented combination of existing components cannot lead to satisfy results because it will enhance point clouds in both static and distractors regions and cannot achieve progressive training (Line 84-86, 94-95 in the manuscript). Our method incorporates epipolar geometry into matching to identify high-confidence static regions and generate point clouds of these static regions. In addition, it can generate coarse masks as priors for later mask prediction procedure to achieve more robust progressive training.

[W2] Contribution of pretrained models

A: We have conducted experiments as follows:

Table 1: Quantitative PSNR comparisons on NeRF On-the-go dataset

method6-view9-view12-view24-view36-view
w/o Roma v212.6614.1215.5418.4219.89
w/o DA313.4715.2016.4419.3920.19
w/o DINO13.6715.7016.8119.7020.49
w/o SAM13.7915.7616.9019.6920.57
w/o DINO & SAM13.1314.5315.3117.8019.07
full13.7415.7116.9219.7220.59

The results show that Roma v2 and DA3 have great impact on performance, while individual DINO or SAM don't. In our design, DINO and SAM are used to provide pixel-wise general and entity features for uncertainty modeling and mask prediction. Because the MGCMT module has provided geometry prior, the following modules may not need too much strong features from pretrained models. Based on these findings, we have removed the features extracted by DINO in the revised manuscript to achieve a more lightweight design.

[W3 & Q2] Separately assess sparse-view and distractor-free performance.

A: We have conducted experiments as follows:

Table 1: Quantitative PSNR comparisons on Mip-NeRF 360 dataset

method12-view24-view
DropGaussian18.4422.09
EDGS18.3522.60
ours19.1022.44

Table 1: Quantitative PSNR comparisons on NeRF On-the-go dataset

methodall views
SpotLessSplats22.22
DeSplat22.32
ours22.63

As showed above, although our method is not specific designed to clean sparse-view input or dense-view with distractors, we can also have competitive results on sparse-view benchmarks without distractors.

[W4] Artifacts in figures.

A: Artifacts are caused by distractors which occlude the static scene and provide wrong gradients for the optimization of 3DGS. Sparse-view inputs provide lack sufficient observations of the scene, leaving regions behind distractors inadequately observed. Consequently, our method can mitigate artifacts but cannot eliminate them entirely.

[Minor concerns 1] Missing citations on recent works

A: We have included the citations in our revised manuscript.

[Minor concerns 2] Formatting issue.

A: We sincerely apologize for it. In fact, this was an unintentional formatting error due to some incorrect LaTeX preamble settings that we had overlooked rather than a deliberate attempt to accommodate more content. We have fixed problems in the revised manuscript which meets the requirements of ICLR formatting.

[Q1] SfM setting & comparison with COLMAP dense reconstruction

A: We follows SfM setting with the previous work RobustSplat. Specifically, it use the full data to do COLMAP sparse reconstruction including feature_extractor, exhaustive_matcher, mapper and image_undistorter. We have conducted experiments as follows:

Table 1: Quantitative PSNR comparisons on NeRF On-the-go dataset

method6-view9-view12-view24-view36-view
COLMAP dense------18.9120.06
ours13.7415.7116.9219.7220.59

As showed in the table, our methods outperform with COLMAP dense methods. In the 6,9,12-view of NeRF On-the-go dataset, the COLMAP dense reconstruction fail due to insufficient information for PatchMatch operation.

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

Related papers

Discussion (0)

Sign in to comment.