acceptodds
Withdrawn from ICLR 2027

GRE-GS: Geometry Repair and Gaussian Extrapolation for Sparse-View 3D Gaussian Splatting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

3D Gaussian Splatting (3DGS) achieves real-time, high-quality novel-view synthesis through explicit Gaussian primitives. However, under sparse-view settings, limited observations provide insufficient geometric support, yielding unreliable geometry and incomplete scene coverage. To address this issue, we present GRE-GS (Geometry Repair and Gaussian Extrapolation for Sparse-View Gaussian Splatting), a reconstruction framework that explicitly separates the two failure modes of sparse-view 3DGS—unreliable geometry and incomplete coverage—and addresses them with geometry repair and Gaussian extrapolation, respectively. For regions covered by training views, GRE-GS repairs the initial point cloud generated by a pretrained geometric model. It completes missing surfaces via locally aligned depth propagation and corrects depth errors via structural conflict correction. For uncovered regions, GRE-GS extrapolates new Gaussians from the trained Gaussian representation—extending their positions using projected DA3 depths and synthesizing colors with an MLP conditioned on neighboring Gaussian appearance. We also introduce a training-view safety mechanism that constrains and filters newly added Gaussians to keep them from disrupting the existing representation. Extensive experiments on Mip-NeRF360, Tanks and Temples, and DTU demonstrate that our method substantially improves novel-view rendering quality across different sparse-view settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Sparse-View Novel View Synthesis, Geometry Repair, Gaussian Extrapolation, Monocular Depth Priors

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XRgFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GRE-GS, a sparse-view 3D Gaussian Splatting framework combining geometry repair and Gaussian extrapolation. It first improves initialization through structural correction and local geometry completion using MASt3R point clouds and DA3 depth priors. It then extends a frozen Gaussian model into low-coverage regions using virtual-view probes, geometry propagation, and learned neighboring-color fusion. A footprint-aware safety guard limits interference with training views. Experiments on Mip-NeRF360, Tanks and Temples, and DTU report improved average novel-view rendering metrics.

Critical strengths
  • The paper addresses two meaningful challenges: unreliable geometry in observed regions and incomplete coverage in unobserved regions.
  • Experiments across three datasets show improved novel-view rendering performance over the common-prior baseline.
Critical weaknesses
  • The validation in Sec. 3.2.2 uses anchors surrounding the hole and does not directly assess whether the fitted depth mapping generalizes to its interior, particularly when the missing region contains different surfaces or structural changes. Additional evidence is needed to establish the reliability of this extrapolation.

  • The necessity of the local geometry completion module is insufficiently established. Since the targeted holes lie within observed regions, multi-view depth prediction and fusion provide a natural alternative. A controlled comparison, ideally within the same missing-region masks and with matched point counts, is needed to demonstrate the additional benefit of the proposed local alignment and validation.

  • The extrapolation relies heavily on local geometric and appearance support. When projected DA3 samples are unavailable, inverse-depth plane extension may favor simple surface continuation, while convex fusion of neighboring Gaussian colors limits the recovery of novel textures. The reported results do not sufficiently demonstrate the ability to reconstruct complex unseen structures or detailed appearance, beyond filling relatively simple boundary gaps.

  • The benefits of the proposed extrapolation over methods using diffusion priors remain unclear. Since it extends local geometry and blends neighboring colors, its ability to recover complex unseen structures and textures may be limited. Comparisons with relevant generative approaches in rendering quality, computational cost, and preservation of observed views would help clarify its advantages.

  • The method section is rather brief, and several key steps require further explanation for clarity and reproducibility:

    (1) Sec. 3.3.1: The retained region (R) and its associated virtual view should be explicitly defined. The computation of (s_R) and (z_R^G), and the fitting and extension of the inverse-depth plane, also need clarification.

    (2) Sec. 3.2.1 and Sec. 3.3.1–3.3.2: Terms such as “supported surface” and “boundary-supported” lack clear operational definitions. Although the appendix provides some thresholds, the criteria for geometric support and reliable occlusion remain insufficiently specified.

  • The discussion of experimental results is limited: Table 3 is not explicitly referenced in the main text, and Figs. 3 and 4 receive only generic descriptions without explaining the specific improvements. Additionally, the caption of Fig. 4 incorrectly swaps the descriptions of the upper and lower rows.

  • Tables 4 and 5 partially report identical worst training-view PSNR changes despite using three-run averages and a single-run diagnostic, respectively. The aggregation procedure and reason for these identical values should be clarified.

  • The layout could use space more effectively to improve figure readability. Table 4 could also be consolidated by removing redundant metric-difference columns, making comparisons easier.

Questions for authors
  • Are points generated by local geometry completion checked for cross-view depth, visibility, or photometric consistency? If not, how does the method prevent single-view completion from introducing conflicting geometry?
Reviewer_myeuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

GRE-GS addresses geometric inaccuracies and incomplete scene coverage in sparse-view 3DGS through geometry repair and Gaussian extrapolation. It leverages pretrained geometric priors to improve novel-view synthesis. Experiments demonstrate improvements over several optimization-based 3DGS baselines.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. Experiments demonstrate improvements over several optimization-based 3DGS baselines.
Critical weaknesses

1. Limited Novelty: Numerous existing studies have leveraged initial point clouds and pretrained geometric priors to improve sparse-view 3DGS reconstruction. For example, dense reconstruction using COLMAP has been explored. More recently, feed-forward methods such as ReSplat[1] can directly predict Gaussian primitives, which can serve as either the final reconstruction or the initialization for optimization-based methods, potentially avoiding complex pipelines.

2. Insufficient Evidence for the Independent Contributions of the Proposed Mechanisms: The method relies heavily on pretrained geometric priors. Although the authors provide ablation studies using the same MASt3R and DA3 priors to demonstrate performance improvements, they have not ruled out the possibility that these gains result from an increased number of Gaussians. Furthermore, robustness across different geometric priors has not been evaluated. Therefore, the independent effectiveness of the proposed mechanisms remains insufficiently validated.

3. Unclear Computational Overhead: The additional costs introduced by stronger initialization are not adequately discussed, including increased GPU memory consumption and runtime.

[1] Xu H, Barath D, Geiger A, Pollefeys M. ReSplat: Learning Recurrent Gaussian Splatting[C]//European Conference on Computer Vision (ECCV). 2026.

Questions for authors
  1. Robustness: Would the proposed method still achieve consistent performance improvements if MASt3R were replaced with alternative geometric models such as DUSt3R or VGGT, or if DA3 were replaced with other depth estimation models?

  2. Additional Computational Overhead: Could the authors report the additional costs introduced by the enhanced initialization, particularly in terms of storage requirements and rendering time?

  3. Comparison with Feed-Forward Methods: How does the proposed method compare with recent feed-forward Gaussian reconstruction approaches? Given that this method already relies on powerful pretrained geometric priors, could directly predicting Gaussians using a feed-forward model achieve better performance under sparse-view settings?

Reviewer_4VypOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes GRE-GS, a sparse-view 3D Gaussian Splatting framework that addresses inaccurate geometry and incomplete scene coverage. It combines MASTER and DA 3 for geometry repair, followed by Gaussian extrapolation with learned appearance propagation and training-view safety constraints.

Critical strengths

The paper provides a clear motivation for separating geometry repair from extrapolation. Component-wise ablations and common-prior comparisons support the effectiveness of the proposed modules to some extent. And the safety-guard experiments demonstrate improved novel-view rendering with minimal degradation of training-view quality.

Critical weaknesses
  1. Limited novelty. The method primarily combines existing geometric foundation models, local depth alignment, point-cloud correction, and Gaussian extrapolation. The technical novelty of the individual components is relatively incremental.

  2. Improvements are partly attributable to stronger initialization. On DTU, the common-prior baseline already achieves 23.20 dB PSNR, compared with 20.80 dB for AD-GS. The complete method reaches 23.60 dB, meaning that the proposed repair and extrapolation contribute only an additional 0.40 dB. Stronger controlled comparisons are needed to isolate the proposed contributions.

  3. Insufficient geometric and efficiency evaluation. Despite emphasizing geometry repair, the paper does not provide quantitative 3D reconstruction metrics such as Chamfer distance. Moreover, the multi-stage pipeline requires additional pretrained models, scene-specific MLP training, and potentially hundreds of thousands of extra Gaussians, without reporting end-to-end runtime or memory consumption.

Questions for authors
  1. Novelty: Could the authors clarify the key technical contributions beyond combining existing geometric foundation models, local depth alignment, and Gaussian extrapolation? What distinguishes the proposed modules from closely related methods?

  2. Contribution of the proposed method: On DTU, the common-prior baseline already achieves 23.20 dB PSNR, compared with 23.60 dB for the full method. Could the authors further justify the significance of the improvements beyond stronger initialization?

  3. Geometry and efficiency evaluation: Given the emphasis on geometry repair, could the authors provide quantitative reconstruction metrics, such as Chamfer distance? Additionally, how does the proposed multi-stage pipeline compare with existing methods in terms of end-to-end runtime and memory consumption?

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

Related papers

Discussion (0)

Sign in to comment.