acceptodds
Withdrawn from ICLR 2027

GeoSupSplat: Generalizable Sparse-View Surface Reconstruction via Geometry-Supervised Gaussian Splatting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Feed-forward Gaussian splatting enables fast reconstruction from sparse images, but photorealistic rendering could not yield an accurate surface. Geometry supervision in this setting faces two coupled challenges: pseudo-depth and other pretrained cues may be ambiguous or inconsistent across views, while the opacity used for RGB compositing does not always represent geometric visibility. We present , a feed-forward framework for reconstructing surfaces from two sparse images using 2D Gaussian surfels. Our key insight is that appearance visibility should not fully determine geometric confidence. GeoSupSplat therefore assigns each surfel separate appearance and geometry opacities, allowing RGB and depth to be composited independently while sharing the same center, scale, and orientation. Furthermore, we propose a set of novel geometry supervision techniques that enable the model to learn high-quality geometry. These include scale-invariant depth, depth-to-normal, and keypoint-based 3D consistency losses, which transfer relative depth structure, local surface orientation, and cross-view agreement without imposing the teacher's unknown scale. We evaluate our model on the Re10K dataset and two zero-shot cross-dataset benchmarks: ScanNet and Replica. Our model outperforms state-of-the-art methods in mesh quality and mesh-rendering tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
2D Gaussian, 3D Reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_gsGEOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents GeoSupSplat, a feed-forward framework for sparse-view surface reconstruction from two input images using 2D Gaussian surfels. The method introduces separate appearance and geometry opacities over shared surfel geometry, together with scale-invariant depth, depth-to-normal, and keypoint-based 3D consistency supervision to improve geometric reconstruction. GeoSupSplat_Generalizabl The approach is evaluated on RealEstate10K and zero-shot transfer to ScanNet and Replica, where it reports improved surface reconstruction and mesh-rendering performance over several feed-forward Gaussian baselines.

Critical strengths
  1. Clear motivation and coherent method design. The paper identifies the mismatch between appearance visibility and geometric visibility in feed-forward Gaussian surface reconstruction, and proposes a weakly decoupled representation together with complementary geometry supervision objectives. The overall methodological pipeline is easy to understand.

  2. The paper evaluates the method on RealEstate10K as well as zero-shot transfer to ScanNet and Replica, and includes ablations on opacity decoupling, geometry supervision, extraction strategies, camera prediction, and depth/normal quality, providing a relatively comprehensive empirical study.

Critical weaknesses
  1. The quantitative results contain internal inconsistencies that materially undermine confidence in the empirical conclusions. In Table 3, the shared-opacity configuration reports CD/PSNR/LPIPS of 0.494 / 18.28 / 0.268, whereas the corresponding ablation in Table 9 reports the same CD and PSNR but LPIPS = 0.260. More seriously, Table 7 reports Precision = 0.439, Recall = 0.194, and F1 = 0.408 for pixelSplat on Replica. Under the paper's stated definition F1=2PR/(P+R)F1=2PR/(P+R), these values give F1≈0.269F1\approx0.269, not 0.408. Even if F1 is computed per sample before averaging, F1i≤2RiF1_i\leq2R_i implies F1‾≤2R‾=0.388\overline{F1}\leq2\overline{R}=0.388, which is still incompatible with the reported 0.408. Since the main empirical claims rely heavily on quantitative superiority, these inconsistencies raise a substantial concern about the reliability of the reported results.

  2. The paper's scale-related interpretation of the geometry supervision is not fully supported by its mathematical formulation. Eq. (9) is described as scale-invariant because globally rescaling the predicted depth D^→aD^\hat D\rightarrow a\hat D is absorbed by the optimal alignment scale. However, rescaling the pseudo-depth as D~→kD~\tilde D\rightarrow k\tilde D yields Ld→k2Ld\mathcal{L}_d\rightarrow k^2\mathcal{L}_d, so the magnitude of the supervision still depends on the arbitrary numerical scale of the pseudo-depth. A related inconsistency appears in the keypoint-based 3D loss: with Xa=Ca+daraX_a=C_a+d_ar_a and Xb=Cb+dbrbX_b=C_b+d_br_b, globally rescaling the predicted depths changes only the ray-depth terms while the camera baseline Ca−CbC_a-C_b remains fixed. Therefore, this loss constrains scene geometry relative to the normalized camera baseline and is not globally scale-invariant. These properties are more restrictive than the paper's claim that the geometry supervision enforces consistency without constraining scene scale.

Questions for authors
  1. The quantitative results contain internal inconsistencies that undermine confidence in the empirical conclusions.

    In Table 3, the shared-opacity configuration reports CD/PSNR/LPIPS of 0.494 / 18.28 / 0.268, whereas the corresponding shared-opacity ablation in Table 9 reports the same CD and PSNR but LPIPS = 0.260.

    More seriously, Table 7 reports Precision = 0.439, Recall = 0.194, and F1 = 0.408 for pixelSplat on Replica. According to the paper's own definition,

    F1=2PRP+R,F1=\frac{2PR}{P+R},

    directly substituting the reported Precision and Recall gives

    F1=2×0.439×0.1940.439+0.194≈0.269,F1=\frac{2\times0.439\times0.194}{0.439+0.194}\approx0.269,

    rather than the reported 0.408. Moreover, even if F1 were computed per sample and then averaged separately from Precision and Recall, each sample satisfies F1i≤2RiF1_i\leq 2R_i, which implies

    F1‾≤2R‾=2×0.194=0.388.\overline{F1}\leq 2\overline{R}=2\times0.194=0.388.

    The reported F1 of 0.408 therefore remains incompatible with the reported Recall of 0.194 under the metric definition given in the paper. Since the paper's main claims rely heavily on quantitative superiority, these internal inconsistencies materially weaken confidence in the reliability of the experimental evidence.

  2. The scale-related interpretation of the geometry supervision is not fully consistent with the mathematical formulation.

    The paper motivates Eq. (9) as a scale-invariant depth objective. Let the rendered depth be D^\hat D and the pseudo-depth be D~\tilde D, with

    \frac{\sum_{u\in\mathcal M}\hat D_u\tilde D_u} {\sum_{u\in\mathcal M}\hat D_u^2+\epsilon}, \qquad \mathcal L_d = \frac{1}{|\mathcal M|+\epsilon} \sum_{u\in\mathcal M} (s^\star\hat D_u-\tilde D_u)^2 .$$ Ignoring the numerical stabilizer $\epsilon$, rescaling the prediction as $$\hat D \rightarrow a\hat D$$ gives $$s^\star \rightarrow \frac{s^\star}{a},$$ so the aligned prediction $s^\star\hat D$ remains unchanged. However, rescaling the pseudo-depth as $$\tilde D \rightarrow k\tilde D$$ gives $$s^\star \rightarrow ks^\star,$$ and consequently $$\mathcal L_d \rightarrow k^2\mathcal L_d.$$ Therefore, while the optimum is invariant to a global rescaling of the predicted depth, the magnitude of the supervision still depends on the numerical scale of the pseudo-depth. This is relevant because the paper explicitly motivates the loss by the unknown scale of the pseudo-depth. A related issue appears in the keypoint-based 3D consistency loss. For two matched pixels, Eqs. (12)--(14) compare $$X_a=C_a+d_a r_a,\qquad X_b=C_b+d_b r_b,$$ through $\|X_a-X_b\|_1$. If the predicted depths are globally rescaled by $k$, $$X_a-X_b = (C_a-C_b)+k(d_a r_a-d_b r_b).$$ Since the camera baseline $C_a-C_b$ is fixed in the normalized camera frame, this objective is not invariant to a global rescaling of the predicted depths. Thus, although it does not impose metric world scale, it does constrain scene geometry relative to the normalized camera baseline. This is more restrictive than the paper's statement that the objective enforces cross-view agreement without constraining absolute scene scale.
  3. It is difficult to disentangle the contribution of the proposed method from the strong external geometric priors used by the system.
    GeoSupSplat fine-tunes a DA3-LARGE backbone with a DINOv2 ViT-L trunk, while also using a separate frozen DA3 model for pseudo-depth supervision and RoMa for cross-view correspondences. The resulting prediction network contains 456.5M parameters, compared with 12.0M for MVSplat and 13.3M for MeshSplat. Although the paper includes several internal backbone and teacher ablations, the main comparison does not control for the substantial difference in pretrained geometric priors and model capacity. Consequently, it remains unclear how much of the reported reconstruction advantage should be attributed specifically to the proposed weakly decoupled representation and geometry losses.

  4. The primary Re10K surface benchmark provides limited evidence for the generality of the reconstruction claims.
    RealEstate10K does not provide ground-truth surfaces, so the paper constructs reference geometry using dense COLMAP reconstructions from only 50 selected scenes. The paper further states that candidate scenes are chosen to have sufficient camera motion, parallax, and texture for dense reconstruction. This introduces a nontrivial selection bias toward scenes that are favorable to classical multi-view reconstruction, while the evaluation target itself remains reconstructed pseudo-ground-truth rather than true surface geometry. Given the relatively small benchmark size, this limits how strongly the reported Re10K gains support broad claims about generalizable surface reconstruction.

  5. The claimed appearance/geometry decoupling is only partial even at the supervision level.
    The central motivation is that appearance visibility should not determine geometric confidence, yet the geometry-supervision mask explicitly requires both appearance coverage AA and geometry coverage BB to exceed thresholds. Thus, whether a pixel receives geometric supervision still depends on the appearance-opacity pathway. This does not invalidate the method, but it weakens the conceptual interpretation that appearance and geometric confidence are cleanly separated.

  6. The strongest reconstruction claims rely on a scale-and-translation aligned evaluation protocol.
    The default surface metrics remove a global scale and translation before evaluation. Under the raw protocol, GeoSupSplat still obtains the best CD and precision, but it does not achieve the best recall or F1. Therefore, the broad claim of uniformly superior reconstruction should be interpreted specifically as superiority in aligned shape quality rather than complete reconstruction accuracy in the original coordinate frame.

  7. Minor Presentation and Notation Issues.
    The manuscript contains a noticeable number of presentation and notation issues. For example, there are several grammatical errors and awkward expressions, such as "which assigning each surfel separate appearance and geometry opacities" in the abstract, "Our method achieve state-of-the-art performance" in the contribution summary, and "the weakly decoupled model produce sharper RGB renderings" in the caption of Fig. 2. The capitalization of "Gaussian" is also inconsistent across the manuscript. There are also several notation and definition issues. In Eq. (17), the full objective includes the camera-loss weight λc\lambda_c, but the implementation details only specify (λrgb,λd,λn,λk)=(1.0,0.1,0.1,0.3)(\lambda_{\mathrm{rgb}},\lambda_d,\lambda_n,\lambda_k)=(1.0,0.1,0.1,0.3), leaving λc\lambda_c unspecified. In addition, several symbols are reused for different meanings: vv denotes a view index in the main method but is later reused as the TSDF voxel size; cc is used for Gaussian color/SH coefficients and later as a correspondence index; and ss denotes Gaussian tangential scale as well as the global alignment scale. The stabilizer ϵ\epsilon also appears in multiple equations without an explicit numerical value, and ci(x)c_i(x) in the RGB rendering equation is not formally defined as the view-dependent color obtained from the spherical-harmonic coefficients. These issues do not by themselves determine my overall recommendation, but they reduce the precision, readability, and reproducibility expected from an ICLR submission.

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

This paper presents GeoSupSplat, which predicts 2D Gaussians for surface reconstruction from sparse views. It uses a DA3 backbone and separates the color and depth opacities. The model is trained with depth, normal, and geometry losses. The results show some improvements in geometry and mesh rendering. However, the novelty over previous geometry supervision methods is not clear, and the evaluation mainly covers indoor scenes with two input views.

Critical strengths
  1. I think separating the color and depth opacities makes sense. Tables 3 and 9 show better geometry and mesh rendering.
  2. The meshes look cleaner, especially on flat surfaces, and the local geometry looks more consistent. The authors also report depth and normal results and test on other datasets. These results give more support for the geometry improvements.
Critical weaknesses
  1. I do not see much novelty in the geometry losses. Similar supervision ideas have been used in MeshSplat, PM-Loss, and GBR. The authors should explain what is new in these losses.
  2. The experiments mainly use static indoor scenes and two input views. I would like to see more results on different scenes and with different numbers of input views.
  3. I am not sure how color and depth stay consistent after separating the opacities. Sharing Gaussian positions and shapes does not guarantee that the two paths use the same visible surface. The authors should discuss when different opacity weights are reasonable and when they may cause inconsistencies.

references: [1] MeshSplat: Generalizable Sparse-View Surface Reconstruction via Gaussian Splatting [2] Revisiting Depth Representations for Feed-Forward 3D Gaussian Splatting (PM-Loss) [3] GBR: Generative Bundle Refinement for High-fidelity Gaussian Splatting with Enhanced Mesh Reconstruction

Questions for authors
  1. What is new in these geometry losses compared with similar ideas in previous work?
  2. Could the authors show more results on different scene types and with different numbers of input views?
  3. Could the authors give concrete examples of when separating color and depth opacities helps, for example on transparent or reflective surfaces? Why are different opacities reasonable in these cases, and is shared Gaussian geometry enough to keep color and depth consistent?
Reviewer_uoDmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents GeoSupSplat, a feed-forward framework for surface reconstruction from two images using pixel-aligned 2D Gaussian surfels. Its central design separates appearance and geometry opacities while sharing surfel positions, scales, and orientations. The geometry-rendering path is supervised through scale-invariant depth alignment, normals derived consistently from depth, and keypoint-based cross-view 3D consistency. The paper reports improved surface reconstruction and mesh-rendering quality on Re10K, with zero-shot transfer to ScanNet and Replica.

Critical strengths
  1. Clear and relevant motivation.

The distinction between a Gaussian's contribution to image appearance and its contribution to surface geometry is relevant to Gaussian-based reconstruction, particularly for transparent surfaces. Separate opacity channels provide a simple mechanism for accommodating these different objectives while retaining shared surfel geometry. Incorporating this mechanism into a feed-forward reconstruction framework is a reasonable design choice.

Critical weaknesses

1. Limited methodological novelty and insufficient differentiation from prior work.

The proposed separation of appearance and geometry opacities has direct precedents. Car-GS [1] assigns separate appearance and geometry opacities to 2D Gaussian primitives for RGB and depth/normal rendering, respectively. Geometry Gaussians [3] similarly introduces a dedicated geometry opacity alongside the opacity used for color rendering. StableGS [2] also separates geometric regularization from appearance rendering through a dual-opacity formulation, although its multiplicative parameterization differs from the independently predicted channels in GeoSupSplat. These closely related works are neither cited nor discussed in the manuscript. The paper therefore needs to clarify what its representation contributes beyond adapting an existing opacity-decoupling principle to generalizable sparse-view reconstruction.

The other principal contribution is the combination of depth, normal, and cross-view consistency supervision. These objectives have a reasonable motivation, but their combination is not accompanied by a sufficiently clear account of the methodological advance.

2. Reported metrics and ranking claims are inconsistent with the stated evaluation definition.

Equation (25) defines F1 as the harmonic mean of precision and recall. However, Table 1 reports precision 0.188, recall 0.207, and F1 0.233 for pixelSplat. The reported F1 exceeds their arithmetic mean, 0.1975, which is incompatible with this definition. Computing F1 separately for each scene and then averaging cannot explain this discrepancy if precision, recall, and F1 use the same samples and aggregation weights: the harmonic-mean upper bound is preserved under such averaging.

A similar inconsistency appears for MVSplat on Re10K in Table 7: precision and recall are 0.430 and 0.207, respectively, yet F1 is 0.352, exceeding their arithmetic mean of 0.3185. In addition, the accompanying text claims that GeoSupSplat achieves the best Re10K F1, although the table lists 0.351 for GeoSupSplat and 0.352 for MVSplat. These discrepancies undermine confidence in the reported comparisons. The authors should clarify the metric computation and correct the affected values and claims before the empirical conclusions can be assessed reliably.

References

[1] Li C, Wang J, Wang X, et al. Car-gs: Addressing reflective and transparent surface challenges in 3d car reconstruction[J]. arXiv preprint arXiv:2501.11020, 2025.

[2] Wang L, Ren Q, Liao K, et al. StableGS: A floater-free framework for 3D Gaussian splatting[J]. arXiv preprint arXiv:2503.18458, 2025.

[3] Zhou H, Lähner Z. Geometry Gaussians: Decoupling Appearance and Geometry in Gaussian Splatting[J]. arXiv preprint arXiv:2606.05124, 2026.

Questions for authors
  1. Can the authors reconcile the precision, recall, and F1 values and verify the reported rankings?

Please specify the aggregation procedure and explain the inconsistencies identified above. Corrected tables, together with per-scene metrics or the relevant evaluation code, would help determine whether these are reporting errors or differences in metric computation, and whether they affect the main conclusions.

  1. What is the specific methodological contribution beyond prior opacity-decoupling approaches and the combination of geometric supervision terms?

Please explain the relationship to existing works and identify the advances required to adapt these ideas to feed-forward sparse-view reconstruction. How were the relative weights of the depth, normal, and cross-view consistency losses selected, and how sensitive is the reconstruction performance to these choices?

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

Related papers

Discussion (0)

Sign in to comment.