acceptodds
Withdrawn from ICLR 2027

GS-Mesh: Implicit Surface Bridging for Collaborative Gaussian-Mesh Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

3D Gaussian Splatting enables high-quality novel-view synthesis and real-time rendering, but its discrete primitives do not directly provide explicit and continuous surface representations. Existing methods either extract meshes through post-processing, preventing mesh feedback to Gaussians, or introduce additional geometric representations that require explicit coupling. To address these limitations, we propose GS-Mesh, a collaborative Gaussian-mesh reconstruction framework based on implicit surface bridging. Our key observation is that surface-aligned Gaussians naturally form attribute-rich oriented surface samples, whose positions, rotations, and opacities can directly induce a continuous implicit surface. Based on this observation, we instantiate Implicit Moving Least Squares (IMLS) on Gaussian primitives and combine it with differentiable surface extraction and mesh rendering, establishing a Gaussian-native implicit bridge for collaborative Gaussian-mesh optimization. We further introduce IMLS-guided densification to improve surface support in under-sampled regions and coarse-to-fine surface optimization to progressively recover coherent geometry and fine details. Experiments on standard surface reconstruction and novel-view synthesis benchmarks demonstrate accurate and detailed mesh reconstruction while retaining competitive rendering quality. The code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3DGS, 3D Reconstruction, Surface Reconstruction

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

This paper proposes GS-Mesh, a method that leverages differentiable mesh optimization within Gaussian-Splatting reconstruction to improve continuous surface reconstruction while maintaining high-fidelity novel-view synthesis (NVS). The contributions include: (1) bridging discrete Gaussians to continuous mesh surfaces through implicit moving least squares (IMLS) surface instantiation from Gaussian-native attributes; (2) IMLS-based densification; and (3) coarse-to-fine surface optimization. The paper conducts experiments on three datasets to validate surface reconstruction and NVS quality.

Critical strengths

This is a well-motivated paper working on a highly significant problem. Reconstructing a continuous surface from discrete 3D Gaussians through mesh extraction is an important problem with direct applications. Involvement of mesh optimization in GS reconstruction is also very intuitive. The paper is also easy-to-follow with clear definition of the claimed novelties. The authors promise code release upon publication, enabling the reproduction of the work.

Critical weaknesses

[Claimed Contributions and Novelties]

• (Major) To my best understanding, the first novelty is the integration of several ideas already introduced/explored by prior works. Specifically, (1) the use of Gaussian-Native parameters to obtain an auxiliary surface is explored by GSRec [1]; (2) IMLS surface representation is explored by IMLS-Splatting; and (3) Mesh involvement in optimization is explored by IMLS-Splatting, SuGaR, and MILo. The paper needs to do a better job of distinguishing itself from these works. Also, regarding the third novelty: Coarse-to-fine surface optimization appears in the baseline work IMLS-Splatting and is an established training paradigm in the literature. To be considered as a meaningful novelty, it should involve some form of special adaptation to IMLS representation or GS-Mesh methodology in general.

• (Minor) The paper needs to cite and discuss a closely related prior work named GSRec [1] (ECCV '24). Moreover, this prior work has overlapping ideas with the proposed GS-Mesh in terms of constructing an IMLS surface directly from Gaussians. The authors are encouraged to elaborate the distinction between GS-Mesh and this prior work.

[Methodology]

• (Minor) To address under-sampled geometry issue, the paper identifies geometrically sparse regions on the IMLS surface and adds new Gaussians there. However, this approach can unnecessarily overpopulate some local regions and increase Gaussian counts potentially without any benefit. Perhaps the more intuitive and straightforward way of identifying under-sampled regions is to use some sort of supervision from normal and depth targets?

[Experiments and Ablations]

• (Major) The experiments do not seem to establish the superiority of GS-Mesh over baselines: GS-Mesh achieves similar results to the baseline on DTU dataset (with higher compute time) and achieves second-best result on the BlendedMVS dataset. Also, the benefit of Gaussian-Native parameterization to induce a bridge needs to be explored. What is the benefit of using these attributes as compared to SDF-based parametrization of baseline work MILo? or other potential parametrizations?

• (Minor) While the ablation establishes the benefits of IMLS-guided densification, it does not compare this densification strategy with other potential approaches including using direct supervision for targeted densification as discussed above.

[1] Wu, Qianyi, Jianmin Zheng, and Jianfei Cai. "Surface reconstruction from 3d gaussian splatting via local structural hints." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

Questions for authors

I invite the authros to focus on addressing the major points above in their rebuttal and consider minor points for future works or further improvements. This work needs improvement in terms of major weaknesses mentioned above, and so I lean towards rejection. I thank the authors for putting effort into the presentation of their work both through the paper and the supplementary demo video.

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

This paper proposes a method that combines IMLS with 3DGS for geometric reconstruction. While the idea is interesting, there are still some issues with the manuscript. The proposed method still falls short of state-of-the-art (SOTA) methods in both geometric reconstruction and view synthesis tasks.

Critical strengths
  1. Incorporating Gaussian native attributes into implicit surface construction—thereby eliminating the need to maintain an independent learnable SDF—provides a well-founded motivation for the proposed approach, while also ensuring a clear mesh rendering feedback path. 2. The active voxel filtering mechanism offers practical benefits: under a 384³ configuration, the number of queries decreased from 56.6M to 0.794M, while the reported Chamfer distance remained unchanged; the appendix further presents runtime and memory usage across different resolutions.
  2. The evaluation includes geometric metrics on DTU and BlendedMVS datasets, novel-view rendering on NeRF Synthetic, component ablation studies, and discussions on scaling to larger scenes; the authors also candidly report the high computational costs associated with unbounded scenes.
Critical weaknesses
  1. The “coarse-to-fine” optimization method described in the paper (Section 3.5) is quite simple, and the distinction between it and IMLS-Splatting is not explicitly explained.

  2. The key difference between this paper and IMLS-Splatting is that the 3DGS branch is retained throughout the entire process and is jointly optimized with IMLS. The authors should place greater emphasis on the distinction between “two-branch optimization” and “using IMLS solely to extract the mesh without incorporating it into training”. It is recommended that relevant experiments be added.

Questions for authors
  1. Of the 15 scenes in the DTU dataset, 11 had metrics that were inferior to those of IMLS-Splatting, and the proposed method only significantly outperformed it in scan37. This indicates that this method does not offer general advantages over it, either in terms of efficiency (11 min vs 45 min) or geometric quality.

  2. In addition, the RaDe-GS metrics in Table 1 differ from the data in the 2026 edition of the TOG journal. Specifically, the chamfer distance (CD) for each scene is better in the TOG journal, with an average performance of 0.48 rather than 0.68. Given that this paper also uses regularization terms from external sources (your prior loss function), the authors should accurately report the best results achieved by RaDe-GS.

  3. The authors also did not consider methods such as PGSR and GeoSVR; incidentally, their metrics are 0.52 and 0.47 (according to the RaDe-GS paper). The performance of these methods is superior to that of IMLS-Splatting and the method proposed in the paper.

  4. Table 2 includes fewer methods for comparison than Table 1, so it is necessary to supplement the performance metrics for these methods (including PGSR, RaDe-GS, and others mentioned above). Furthermore, in Figure 4, the author did not include the Mesh-Splatting (seems to have better metrics) in the comparison of the visualization results.

  5. The view synthesis results in Table 3 fell short of expectations. In fact, 2DGS does not perform well on view synthesis tasks (significantly worse than RaDe-GS); however, the proposed method performs even worse than 2DGS. The author should analyze the reasons for this.

  6. In Table 4, “Base” is listed as 0.80. Does this mean that the baseline in this paper is 2DGS? Further clarification is needed.

  7. The “IMLS” row in Table 6 is identical to the “Ours” row in Table 2 for every scene. It is recommended dividing the ‘IMLS’ row into two cases: “two-branch optimization” and “using IMLS only to extract the mesh without participating in the optimization”.

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

This paper proposes GS-Mesh, a framework for joint Gaussian and mesh reconstruction. The main idea is to treat surface-aligned Gaussians as oriented points and use Implicit Moving Least Squares (IMLS) to construct a continuous surface directly from Gaussian attributes. The surface is then converted into a mesh through differentiable Marching Cubes, allowing mesh-rendering losses to optimize the underlying Gaussians. The authors also introduce IMLS-guided densification and coarse-to-fine optimization to improve surface coverage and geometric details. The method is evaluated on DTU, BlendedMVS, and NeRF Synthetic, with additional qualitative results on Tanks and Temples.

Critical strengths
  1. The paper is well written and easy to follow. The motivation is clear, and the overall framework is straightforward to understand.
  2. Using Gaussian attributes to directly construct an IMLS surface is a simple and intuitive idea. It avoids maintaining a separate implicit representation and allows mesh-rendering supervision to directly update the Gaussians.
Critical weaknesses
  1. The proposed method seems quite similar to IMLS-Splatting, as both construct an IMLS field from oriented points and use differentiable mesh extraction. The main difference seems to be that GS-Mesh directly uses Gaussian attributes to build the IMLS field. What are the actual benefits of this design compared with IMLS-Splatting? The paper does not clearly demonstrate significant improvements in reconstruction quality or efficiency.
  2. The improvements over existing methods are not very convincing. GS-Mesh performs similarly to IMLS-Splatting on DTU but takes much longer to train. It also falls behind some recent methods on BlendedMVS and NeRF Synthetic. Given the additional training cost, is the improvement really worth it?
  3. IMLS is commonly used for point cloud surface reconstruction, but its quality depends on the accuracy and distribution of the input points and normals. Since Gaussians can also be sparse, noisy, or unevenly distributed, wouldn't this make the IMLS field unstable and affect mesh quality? How robust is the method when the initial Gaussian geometry is inaccurate?
  4. The method seems to have limited scalability to large scenes. For unbounded scenes, it relies on a MILo-style adaptive tetrahedralization, which makes training much slower. How well does the method actually scale compared with other Gaussian-based reconstruction methods? Why are only qualitative results provided for large-scale scenes?
Questions for authors

See Weakness

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

Related papers

Discussion (0)

Sign in to comment.