acceptodds
Under review as a conference paper at ICLR 2027

CoSplat: Complementary Surface Geometry Learning for Generalizable Surface Reconstruction

15 views

Abstract

Feed-forward Gaussian Splatting enables efficient surface reconstruction from sparse views, yet the geometry of its predicted primitives remains ambiguous under rendering-driven supervision. Specifically, distinct primitive configurations can produce similar renderings, as geometric errors can be compensated for by adjustments to appearance parameters. To resolve this ambiguity, we propose CoSplat, a feed-forward 2DGS framework that explicitly constrains surfel geometry. We first introduce Geometry-Calibrated Normal Supervision to directly regularize primitive orientations and mitigate geometric drift caused by geometry-appearance compensation. Since local normal constraints are insufficient to recover globally coherent geometry from sparse views, we further propose Confidence-Guided Tangent-Plane Depth Propagation, which translates reliable local orientations into surface-aligned depth hypotheses and validates them using multi-view geometric evidence, thereby promoting globally coherent surface reconstruction. Experiments on DTU show that CoSplat achieves state-of-the-art surface reconstruction accuracy among feed-forward Gaussian methods while retaining efficient inference. Zero-shot evaluations on DTU, BlendedMVS, and ETH3D further demonstrate strong cross-dataset generalization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Feedforward 3D Reconstruction, Surface Reconstruction, Gaussian Splatting, Sparse View Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_tusxOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents CoSplat, a feed-forward framework for generalizable surface reconstruction from sparse-view images using 2D Gaussian Splatting. The authors address a key limitation of existing feed-forward Gaussian-based reconstruction methods, where rendering supervision alone may not provide sufficiently accurate geometric constraints, potentially leading to inaccurate surface reconstruction despite plausible novel-view synthesis. To address this issue, the paper introduces two complementary components, Geometry-Calibrated Normal Supervision (GCNS) and Confidence-Guided Tangent-Plane Depth Propagation (TDP). GCNS combines monocular normal priors with geometry-derived normals to provide more reliable normal supervision. TDP improves geometric accuracy by propagating depth hypotheses based on local tangent planes and evaluating their reliability using normal consistency and multi-view geometric confidence. The proposed framework jointly improves surfel orientation and position, enabling more accurate surface reconstruction without requiring expensive per-scene optimization. Extensive experiments on DTU, BlendedMVS, and ETH3D demonstrate competitive reconstruction accuracy and strong generalization capabilities.

Critical strengths
  • The paper addresses an important limitation of existing feed-forward Gaussian-based reconstruction methods, where rendering supervision alone may not sufficiently constrain surface geometry. The motivation is clearly presented, and the proposed approach effectively incorporates complementary geometric constraints to improve reconstruction accuracy.

  • The proposed GCNS and TDP modules are well designed and complement each other effectively. GCNS leverages monocular normal priors while maintaining consistency with predicted geometry, whereas TDP utilizes the resulting normal information to refine depth estimates. Their integration provides a coherent framework for jointly improving surfel orientation and position.

  • The proposed method achieves strong reconstruction accuracy while maintaining efficient feed-forward inference. Experiments across multiple datasets demonstrate its effectiveness and generalization capability. Furthermore, the ablation studies provide convincing evidence of the individual contributions of GCNS and TDP, as well as the benefits of combining these two components.

Critical weaknesses

I did not identify any critical weaknesses in this paper. However, I have a few minor questions and suggestions for clarification, as detailed in the questions for authors.

Questions for authors
  • The experimental results demonstrate the effectiveness of LmonoL_{\text{mono}}, and the calibration process provides an intuitive way to supervise normals without explicit depth supervision during pretraining. However, LgeoL_{\text{geo}} and LmonoL_{\text{mono}} appear to be coupled through the predicted geometry rather than acting as independent supervision terms. Could the authors clarify whether any specific gradient detachment strategy was employed during the calibration process?

  • TDP employs a fixed propagation radius of r∈r \in {1,21,2} in image space. While this is a reasonable design choice, the spatial extent covered by a fixed pixel-level neighborhood may vary with image resolution, potentially affecting the quality of tangent-plane-based depth propagation. Could the authors clarify whether this formulation is sensitive to input image resolution?

3 comments
AuthorsOpenReview ↗

Response to Questions

Dear Reviewer,

We sincerely thank you for taking the time to review our paper and for your thoughtful questions of our work. We address your questions below.

Response to Question 1

Thank you for asking about gradient flow in GCNS. We detach the depth-derived normals before low-pass filtering. The resulting geometric reference serves as the target for LgeoL_{geo} and determines the rotation used to calibrate the monocular normals. We also detach the calibrated target for LmonoL_{mono}. These detachments prevent gradients from either normal loss from flowing back to the depth prediction through its target branch.

Response to Question 2

Thank you for asking about the choice of propagation radii in TDP. In principle, the quality of tangent-plane propagation can be sensitive to input resolution, as the fixed pixel radius covers different physical extents on the surface as the image resolution changes, which can affect the quality of the propagated depth hypotheses. Since resolution also affects reconstruction quality in the absence of TDP, it is difficult to isolate the effect of the propagation radius from changes in overall reconstruction quality. We would also like to explain why we chose these propagation radii:

  1. We use tangent-plane depth propagation to turn the predicted normals into depth hypotheses that encourage surface continuity and improve reconstruction accuracy.
  2. A radius of one pixel can yield too few useful depth hypotheses. Inspired by the adaptive checkerboard sampling of depth hypotheses in the PatchMatch MVS method of ACMM [1], we include candidates from a slightly wider neighborhood to broaden the local depth search. We use confidence scores to downweight unreliable propagation hypotheses.
  3. We use these hypotheses for local depth refinement. Over longer distances, tangent-plane propagation can introduce errors on curved surfaces or across discontinuities. We therefore limit the propagation radii to r∈{1,2}r\in\{1,2\} pixels and use a feature pyramid to fuse features at multiple scales into full-resolution descriptors for cross-view validation of the depth hypotheses.

Our choice of radii is empirical, and this dependence on input resolution is a limitation of the current design.

Thank you again for your time and feedback. We will revised our text later, and our code and model weights will be available after the final decisions.

Reference

[1] Q. Xu and W. Tao. “Multi-Scale Geometric Consistency Guided Multi-View Stereo.” CVPR 2019.

Thank you for the clarifications. My questions have been adequately addressed. Regarding Q2, I believe an additional ablation study with different propagation radii rr would further strengthen the justification for the chosen values.

Thank you for your constructive feedback on the choice of propagation radius. During the design of TDP, we evaluated reconstruction performance with different propagation radii, as shown in the table below.

Setting243740556365698397105106110114118122Mean
r=10.761.871.300.851.171.260.681.011.160.740.820.580.470.871.000.97
r=20.741.961.210.831.031.160.630.971.100.710.810.470.440.850.960.92
r=30.741.981.220.831.051.170.640.981.100.700.800.480.450.860.960.93

The results for r=2 and r=3 are very similar, and the difference may fall within experimental variability. This is partly because reprojection and normal consistency suppress unreliable propagated depth hypotheses. We will release our code after the review process to provide additional implementation details.

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

CoSplat addresses geometric ambiguity in feed-forward Gaussian surface reconstruction from sparse calibrated views. The method combines Geometry-Calibrated Normal Supervision (GCNS) with Confidence-Guided Tangent-Plane Depth Propagation (TDP). GCNS uses depth-derived geometry to define a local orientation reference and transfers monocular normal deviations into that reference through a minimal rotation. TDP then uses the learned normals to generate surface-aligned depth hypotheses, filters them using normal and reprojection consistency, and validates them through multi-view feature matching. Experiments show improved DTU surface reconstruction in both fine-tuned and zero-shot settings, supported by component-level ablations.

Critical strengths
  1. Motivation is claer and reasonable. This paper assigns different roles to depth-derived geometry and monocular normals rather than treating them as interchangeable absolute targets.

  2. Strong performance with useful efforts toward fair comparison. CoSplat outperforms the compared feed-forward Gaussian methods while remaining close to SparseSurf's three-view accuracy. Using common input resolution and mesh extraction for feed-forward Gaussian methods, matched fine-tuning settings, and depth-range-corrected zero-shot baselines strengthens these comparisons.

Critical weaknesses
  1. Quantitative evidence for broader generalization remains limited. The numerical surface-reconstruction results are concentrated on DTU under different input-view configurations. The RE10K-to-DTU zero-shot experiment is meaningful, but the evaluations on BlendedMVS and ETH3D provide only qualitative examples. Consequently, the evidence for generalization beyond DTU's object-centric scenes is less conclusive than the paper's broader claims suggest. Quantitative evaluation on an additional dataset would strengthen this aspect of the contribution.
Questions for authors
  1. Can the cross-dataset results be supported by quantitative geometry evaluation? Please report reconstruction metrics for a representative set of scenes from at least one additional dataset, together with several of the existing baselines under the same sparse-view input protocol. BlendedMVS or ETH3D would be natural choices given the qualitative results already presented; a scene-level benchmark such as ScanNet++ would also be informative. The BlendedMVS evaluation in Mesh Splatting [1] or the scene-level evaluation in G4Splat [2] could provide references for dataset selection and metric reporting, with the protocol adapted to CoSplat's sparse calibrated-view setting.

  2. Can the comparison include closely related methods that are currently absent from the quantitative evaluation? FatesGS [3] would be a useful optimization-based reference because it addresses sparse-view geometry through depth and feature consistency. SurfelSplat [4] is also relevant because it also targets generalizable sparse-view reconstruction with feed-forward Gaussian surfels. A comparison under matched input-view settings would help position CoSplat among closely related approaches.

  3. Clarify the experimental settings in ablation study. In ablation study, please specify the loss terms, relative weights, and validity masks used for naive joint supervision in Table 5, including whether the depth-derived targets receive gradients. A sensitivity check on the relative loss weights would help establish whether the observed degradation reflects conflicting supervision targets rather than an unfavorable weighting choice.

References [1] Zhang R, Wu J, Chen J. Mesh Splatting for End-to-end Multiview Surface Reconstruction[J]. arXiv preprint arXiv:2601.21400, 2026. [2] Ni J, Chen Y, Yang Z, et al. G4splat: Geometry-guided gaussian splatting with generative prior[C]//International Conference on Learning Representations. 2026, 2026: 138924-138951. [3] Huang H, Wu Y, Deng C, et al. FatesGS: Fast and accurate sparse-view surface reconstruction using gaussian splatting with depth-feature consistency[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2025, 39(4): 3644-3652. [4] Dai C, Zhang S, Chen M, et al. SurfelSplat: Learning Efficient and Generalizable Gaussian Surfel Representations for Sparse-View Surface Reconstruction[J]. Advances in Neural Information Processing Systems, 2026, 38: 173477-173504.

6 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

Dear Reviewer,

We sincerely appreciate the time you took to review our paper and your constructive feedback. We respond to each of your concerns and questions below.

Response to Weakness 1 and Question 1

Thank you for raising this concern about cross-dataset generalization and for suggesting additional quantitative evaluation. Following your suggestion, we follow BlendedMVS evaluation in Mesh Splatting[1], we evaluated 18 BlendedMVS scenes. For each scene, we formed view triplets using the official camera parameters and sampled 10 triplets without replacement. Using the ground-truth depth maps, we report the absolute relative depth error and the mean angular error of normals derived from depth. For ETH3D, we used the same triplet sampling protocol on 13 scenes. The valid pixels in the laser-scan-derived ground-truth depth maps provide limited neighborhood coverage for reliable normal estimation, so we report the absolute relative depth error and the mean absolute depth error. All methods use the same test views, and we report metrics averaged across scenes for each dataset in Table R1, where CoSplat consistently outperforms baseline methods on both datasets.

Table R1. Cross-dataset depth and normal evaluation.

MethodBlendedMVS AbsRel ↓BlendedMVS Normal MAE (°) ↓ETH3D AbsRel ↓ETH3D AbsDiff (m) ↓
MVSplat0.065135.700.3261.643
MeshSplat0.065334.910.221.196
DepthSplat0.059335.510.1210.554
IDESplat0.118042.790.0730.377
SurfSplat0.060234.630.0780.392
Ours0.049031.140.0640.358

We will include this table in the revised main text and provide the experimental details in the appendix.

AuthorsOpenReview ↗

Response to Question 2

Response to Question 2

Thank you for suggesting these baselines. For FatesGS[2], we followed the settings provided in its official code and used the same DTU input views as CoSplat in each setting. Tables R2 and R3 report the three-view and two-view results, respectively.

Table R2. Per-scene CD (mm) on DTU with three input views.

Method243740556365698397105106110114118122MeanTime
FatesGS0.671.941.170.771.281.230.631.050.980.690.750.480.410.780.900.9214 min
Ours0.741.961.210.831.031.160.630.971.100.710.810.470.440.850.960.920.6 s

Table R3. Per-scene CD (mm) on DTU with two input views.

Method243740556365698397105106110114118122Mean
FatesGS1.973.422.381.044.961.841.454.672.962.881.872.271.451.481.302.40
Ours1.572.301.821.311.681.761.171.391.711.211.480.800.791.461.221.44

Later we will update these in revised main text.

Under our original evaluation setting, the released SurfelSplat[3] checkpoint yielded meshes with limited reconstruction accuracy. We therefore followed SurfelSplat's DTU input-view setting and used views 00, 22, and 99 for all methods, and we also follow them report D2S on the DTU test set in Table R4.

Table R4. Per-scene D2S (mm) on DTU using input views 00, 22, and 99.

Method243740556365698397105106110114118122Mean
UFORecon1.152.421.672.551.902.731.551.492.160.952.221.981.402.112.321.91
FatesGS0.772.351.431.001.312.060.851.241.060.831.220.580.640.991.321.18
SurfelSplat1.231.691.630.901.241.141.121.181.130.790.841.020.980.841.041.12
Ours0.730.840.470.311.070.870.441.030.930.430.310.600.290.420.380.61
AuthorsOpenReview ↗

Response to Question 3

Response to Question 3

Thank you for asking about the ablation settings. In naive joint supervision, we directly supervise the normal predictor per pixel with both depth-derived and monocular normals. For each target, we compute the per-pixel cosine distance 1−n^⊤n1-\hat{\mathbf{n}}^\top\mathbf{n}. We add the two losses with equal weights of 11 and multiply their sum by the overall normal loss coefficient of 0.010.01 used in the main paper. For the validity mask, we check whether neighboring depths are within the valid range and whether their relative difference from the center depth satisfies ∣D−Dneighbor∣/D<0.05\lvert D-D_{\mathrm{neighbor}}\rvert/D<0.05. We detach the depth-derived normal targets to prevent gradients from flowing through those targets into the depth predictor.

Table R5. Per-scene CD (mm) for the normal supervision ablation on DTU.

Setting243740556365698397105106110114118122Mean CD (mm)
Monocular supervision1.182.161.751.121.381.930.921.211.330.821.060.750.850.991.181.24
Local depth supervision1.112.101.731.011.141.930.871.131.300.760.990.680.830.911.101.17
Naive joint supervision1.382.442.031.192.282.200.991.271.610.800.981.020.791.011.331.42

As shown in Table R5, per-scene results for the ablation with different supervision indicated that the overall performance drop under naive joint supervision is driven by a subset of scenes. To further examine whether the degradation in these scenes is linked to conflicting supervision targets, we conducted the following experiments.

With the overall normal loss coefficient fixed, the relative weights assigned to the two supervision signals directly determine the relative weight of each target where the signals disagree. To show how this weighting affects reconstruction quality, we initialized from the Base + 2DGS checkpoint used in the main-component ablation, obtained by pretraining on RE10K while leaving the normal module untrained. We kept the overall normal loss coefficient fixed at 0.010.01 and varied the relative weights of the depth-derived and monocular supervision terms. We set the two weights to

w_{\mathrm{mono}}=\frac{2r}{1+r}, \qquad w_{\mathrm{depth}}+w_{\mathrm{mono}}=2,$$ with $r\in\{0.1,0.3,1,3,10\}$. Table R6 reports CD for each setting. As shown in Table R6, varying $r$ has a noticeable effect on reconstruction quality. Since **$r$ partly reflects the relative weighting of conflicting supervision targets**, we attribute the degradation to the underlying conflicts between the two normal fields. **Table R6. Sensitivity to the relative weights of depth-derived and monocular normal supervision**. Per-scene values are CD in mm. | Setting | 24 | 37 | 40 | 55 | 63 | 65 | 69 | 83 | 97 | 105 | 106 | 110 | 114 | 118 | 122 | Mean CD (mm) | | -------------------------------------- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | -----------: | | Local depth supervision ($w_{mono}=0$) | 1.60 | 2.36 | 1.71 | 1.09 | 1.34 | 2.04 | 1.02 | 1.38 | 1.25 | 0.83 | 1.02 | 1.10 | 0.63 | 1.08 | 1.26 | 1.31 | | $r=0.1$ | 1.55 | 2.63 | 1.80 | 1.29 | 1.49 | 2.05 | 0.97 | 1.42 | 1.35 | 0.77 | 0.94 | 1.05 | 0.74 | 1.01 | 1.40 | 1.36 | | $r=0.3$ | 1.60 | 2.58 | 1.87 | 1.33 | 1.46 | 2.09 | 1.02 | 1.47 | 1.35 | 0.80 | 1.03 | 1.10 | 0.79 | 1.06 | 1.44 | 1.40 | | $r=1$ | 1.61 | 2.67 | 1.90 | 1.35 | 2.13 | 2.09 | 1.02 | 1.52 | 1.39 | 0.81 | 0.99 | 1.10 | 0.78 | 1.07 | 1.43 | 1.46 | | $r=3$ | 1.70 | 2.66 | 1.88 | 1.34 | 1.54 | 2.13 | 0.90 | 1.48 | 1.32 | 0.81 | 1.02 | 1.09 | 0.83 | 1.03 | 1.36 | 1.41 | | $r=10$ | 1.64 | 2.57 | 1.86 | 1.32 | 1.46 | 2.08 | 1.02 | 1.45 | 1.35 | 0.82 | 0.99 | 1.09 | 0.78 | 1.06 | 1.41 | 1.39 | | Monocular supervision ($w_{depth}=0$) | 1.56 | 2.33 | 1.71 | 1.07 | 1.35 | 2.01 | 1.02 | 1.36 | 1.24 | 0.82 | 1.00 | 1.07 | 0.63 | 1.09 | 1.24 | 1.30 | **Thank you again for your time and feedback. We will revised our text later, and our code and model weights will be available after the final decisions.**
Reviewer_ta3QOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CoSplat, a feed-forward 2D Gaussian splatting framework for surface reconstruction from sparse calibrated views. The method couples two geometric constraints. Geometry-Calibrated Normal Supervision (GCNS) treats depth-derived normals as an absolute orientation anchor and transfers only the angular residual of a frozen monocular normal prior after a minimal rotation aligns the two low-frequency bases. Confidence-Guided Tangent-Plane Depth Propagation (TDP) converts the learned normals into tangent-plane inverse-depth hypotheses, filters them with chained normal confidence and cross-view reprojection confidence, and refines the coarse depth with a local matching volume and a lightweight 3D CNN. The model is pretrained on RealEstate10K and briefly fine-tuned on DTU. BlendedMVS and ETH3D are used for qualitative zero-shot evaluation, and inference takes about 0.6 s.

Critical strengths
  1. The GCNS calibration is simple and geometrically explicit. Eqs. 5 and 6 define a minimal rotation from the smoothed monocular base to the smoothed geometry base and show that the angular residual is preserved exactly. This avoids an additional learned fusion network and integrates cleanly with the DepthSplat backbone. The design is not a new learning principle, but it is a clean and interpretable way to combine two cues that are otherwise in conflict.

  2. The TDP ablation is rigorous and includes a useful negative result. Table 6 isolates the candidate domain, multi-view matching, normal confidence, and reprojection confidence. Propagated candidates with multi-view matching reduce CD from 1.23 to 1.09; adding both confidence terms reaches 0.92. Importantly, propagated candidates with the two confidence terms but without multi-view matching give 1.21, worse than the fixed-local baseline at 1.18. This supports the authors' claim that propagation requires multi-view validation rather than confidence alone.

Critical weaknesses
  1. The cross-dataset generalization claim exceeds the evidence. Only DTU is evaluated quantitatively. BlendedMVS is shown as a qualitative mesh comparison in Figure 7, and ETH3D as qualitative depth and normal maps in Figure 6. No Chamfer distance, depth error, or normal error is reported for either dataset. Both datasets provide ground-truth depth maps for their training or validation scenes, at least depth and normal errors are feasible. The abstract and conclusion nevertheless describe extensive experiments on DTU, BlendedMVS, and ETH3D and strong cross-dataset generalization. The claim should be softened or supported by a small quantitative table.

  2. The baseline set is incomplete for the central claim. MeshSplat(AAAI2026) is the closest feed-forward 2DGS surface reconstruction method. It takes two views, uses monocular normal supervision, and has official code and weights. It appears only in the three-view zero-shot Tables 3 and 7, and is absent from the fine-tuned Table 1 and the two-view Table 2. The current omission leaves the claim of state-of-the-art accuracy among feed-forward Gaussian methods unestablished. FatesGS(AAAI2025), a per-scene sparse-view surface reconstruction method with code, is also missing. And SparseSurf comparison reports 0.92 mm on DTU large-overlap, tied with CoSplat. Per-scene breakdowns for the additional comparisons and ablations in the appendix would help.

  3. The zero-shot comparison is not conducted under a unified protocol. Appendix A.3 applies depth-range rescaling only to DepthSplat and SurfSplat, reducing their mean CD from 9.58 to 1.72 and from 9.25 to 1.72, respectively, but does not apply the same correction to MVSplat or MeshSplat. The paper explains that the DPT upsampling heads of DepthSplat and SurfSplat carry an inverse-depth residual tied to the training range, but it does not establish that MVSplat and MeshSplat are free of a comparable mismatch. More importantly, MeshSplat is a two-view method but is evaluated in a three-view table, and the paper does not describe how three images are adapted to a two-view network. The 1.39 versus 3.61 comparison in Table 3 therefore cannot be read as a method advantage without clarification.

Questions for authors
  1. Eq (9), whether the corresponding reprojection confidence is differentiable during training. If confidence is used as a differentiable weight during training, a hard threshold becomes non-differentiable at the boundary, with the gradient being zero almost everywhere. Conversely, if confidence is detached, serving only as a weight during inference—then such abrupt transitions do not affect the gradient.

  2. The novelty of TDP is overstated relative to classical tangent-plane propagation. Eq. 10 is the standard tangent-plane inverse-depth construction. The same geometry appears in NTPP-MVSNet,"Joint Bilateral Propagation Upsampling for MVS", et al. The paper should cite the priors and frame the contribution around the confidence design and multi-view validation.

  3. The two-view instantiation of TDP is not described. Section 3.3 and Appendix A.2 refer to selected neighboring views and to correlations averaged across two selected source views, but Table 2 uses two input views, which leaves only one source view for the reference. The paper does not explain how the two source views are chosen when V=2, whether the same view is reused, whether the reprojection confidence degenerates to a single-view prior, or how the candidate range is constructed in this case. Because the two-view setting is the paper's main stress test, this is a substantive methodological gap.

  4. The rendering trade-off is confined to the appendix. Table 8 reports PSNR 26.14, SSIM 0.863, and LPIPS 0.146 for CoSplat, compared with 26.77, 0.877, and 0.123 for DepthSplat-S and 27.80, 0.893, and 0.108 for IDESplat. The paper acknowledges in Appendix A.3 that the surface-constrained 2DGS representation offers less appearance flexibility and that opacity is not explicitly regularized, but the main text does not discuss this trade-off. Since the method is built on a rendering backbone and the drop is 0.63 dB relative to DepthSplat-S and 1.66 dB relative to IDESplat, the limitation should be stated in the main paper.

6 comments
AuthorsOpenReview ↗

Response to Weakness 1

Dear Reviewer,

We sincerely thank you for taking the time to review our paper and for your constructive feedback. We respond to each of your concerns and questions below.

Response to Weakness 1

Thank you for pointing out that our claim about cross-dataset generalization needed quantitative support. Following your suggestion, we evaluated CoSplat and the baselines using the ground-truth depth maps provided by BlendedMVS and ETH3D. For BlendedMVS, we report the absolute relative depth error and the mean angular error of normals derived from depth. For ETH3D, the laser-scan-derived ground-truth depth maps provide limited valid-pixel coverage for reliable normal estimation, so we report the absolute relative depth error and the mean absolute depth error. Table R1 presents the results. Since our main experiments use a three-view reconstruction setting, we use the official camera poses to form view triplets in each scene of both datasets. We sample 10 triplets per scene without replacement and report the metrics averaged across all scenes in each dataset. As shown in Table R1, CoSplat achieves the lowest error among the compared methods on all reported metrics across the two datasets.

Table R1. Cross-dataset depth and normal evaluation.

MethodBlendedMVS AbsRel ↓BlendedMVS Normal MAE (°) ↓ETH3D AbsRel ↓ETH3D AbsDiff (m) ↓
MVSplat0.065135.700.3261.643
MeshSplat0.065334.910.221.196
DepthSplat0.059335.510.1210.554
IDESplat0.118042.790.0730.377
SurfSplat0.060234.630.0780.392
Ours0.049031.140.0640.358

We will include this table in the revised main text and provide the experimental details in the appendix.

AuthorsOpenReview ↗

Response to Weakness 2

Response to Weakness 2

Thank you for identifying this gap in the comparisons supporting our main claim. We fine-tuned MeshSplat following the procedure described in the paper. Tables R2 and R3 report the results for MeshSplat and FatesGS under the three-view and two-view settings, respectively.

Table R2. Per-scene CD (mm) on DTU with three input views.

Method243740556365698397105106110114118122MeanTime
FatesGS0.671.941.170.771.281.230.631.050.980.690.750.480.410.780.900.9214 min
MeshSplat$_{\mathrm{ft}}$2.443.502.701.452.131.882.291.782.691.431.841.701.571.801.682.061 s
Ours0.741.961.210.831.031.160.630.971.100.710.810.470.440.850.960.920.6 s

Table R3. Per-scene CD (mm) on DTU with two input views.

Method243740556365698397105106110114118122Mean
FatesGS1.973.422.381.044.961.841.454.672.962.881.872.271.451.481.302.40
MeshSplat$_{\mathrm{ft}}$3.084.853.002.023.322.822.492.463.012.382.542.671.872.791.992.75
Ours1.572.301.821.311.681.761.171.391.711.211.480.800.791.461.221.44

We will update the experimental results in the main text.

Finally, we acknowledge that leading optimization-based methods such as SparseSurf and FatesGS can achieve higher surface reconstruction accuracy than our method in some scenes with favorable viewing conditions. This remains a limitation of our approach. But both methods rely on dense COLMAP reconstructions for initialization, which can limit the quality of the recovered surfaces under particularly challenging viewing conditions. Our method handles both typical and challenging view configurations robustly while balancing reconstruction accuracy and computational efficiency.

AuthorsOpenReview ↗

Response to Weakness 3

Response to Weakness 3

Thank you for identifying this issue with our evaluation protocol. In our initial zero-shot experiments, we followed the official DTU generalization configurations for MVSplat and MeshSplat. We recognize that the comparison protocol needed closer consideration, and we will update the experimental results in the revised paper.

MeshSplat builds on the MVSplat architecture, and the official MVSplat implementation supports multi-view (>2>2) input. With more than two input views, MeshSplat constructs cost volumes for individual reference–source view pairs and aggregates them to estimate a depth map for each reference view.

Table R4. Comparison of zero-shot results on DTU with 3 views setting. ∗ denotes results with range rescaled depth residuals

Method243740556365698397105106110114118122MeanTime
MVSplat3.564.814.102.604.813.533.184.442.962.653.042.663.412.983.223.460.4 s
MVSplat$^{\ast}$4.885.16.025.764.285.395.113.833.333.855.165.125.3555.974.940.4 s
MeshSplat3.884.504.492.923.364.103.323.764.102.094.131.925.712.703.153.611 s
MeshSplat$^{\ast}$2.664.114.244.393.34.533.133.022.822.163.462.842.873.424.653.441 s

To discuss why range rescaling may degrade performance (MVSplat*), we first group the depth prediction approaches used by feed-forward Gaussian splatting methods into three categories.

  1. MVSplat-based methods, including MeshSplat. These methods construct a cost volume at a coarse scale to predict coarse depth explicitly, then use a refinement network to predict a residual for the 2D depth map.
  2. DepthSplat-based methods, including SurfSplat, Ours. These methods extend MVSplat with a monocular network and use a DPT head that incorporates monocular features to predict the depth refinement residual.
  3. Coarse-to-fine methods such as IDESplat. Following an approach similar to MVSNet-based methods, these methods progressively narrow the depth search range from coarse to fine scales.

The first two categories refine 2D depth maps with a learned network, without fine-scale multi-view feature matching. This makes direct transfer from RE10K training to DTU testing difficult, primarily because the depth ranges of the two datasets differ substantially in scale when mapped into the inverse-depth space used by the networks. Uniform sampling in inverse depth produces more closely spaced depth hypotheses near the camera. With the same far bound of 100, the RE10K training range of [1,100][1,100] used by MVSplat and MeshSplat gives an inverse-depth span of 0.99, compared with 1.99 for the [0.5,100][0.5,100] range used by other methods. This difference will partly explain why residual rescaling yielded smaller and less consistent gains for MVSplat and MeshSplat.

To better reflect the original methods' reconstruction capabilities in the zero-shot setting, we resorted to range rescaling as a compromise, even though it only partially addresses the depth-range mismatch. For a comparison after fine-tuning, see Tables 1 and 2 in the main paper. For the BlendedMVS and ETH3D experiments above, we checked the GT depth ranges and found the RE10K bounds suitable for both datasets, so we used those same bounds.

Comments on the paper

1 comment

We sincerely thank all reviewers again for their constructive comments and questions. We also thank the Area Chair and everyone who took the time to read our manuscript. We have revised the manuscript in response to the feedback. The main updates are as follows.

  1. We added citations to prior work on tangent-plane propagation in the Method section. (section 3.3)
  2. We added results for FatesGS and fine-tuned MeshSplat to the main DTU comparisons. (table 1, table 2)
  3. We expanded the description of the evaluation metrics and implementation details in the Experiments section.
  4. We added quantitative zero-shot results for depth and normal estimation on BlendedMVS and for depth estimation on ETH3D. (table 4)
  5. Following the reviewers’ suggestion, we added an ablation that varies the relative weights of the two normal supervision signals to examine how conflicting normal targets affect reconstruction accuracy. (table 7)
  6. We added a discussion of CoSplat’s limitations to the Conclusion.
  7. We added further quantitative results and qualitative visualizations to the appendix. (table 11, figure 10)
  8. We clarified how gradients are detached in GCNS and TDP in the appendix.
  9. We made additional minor revisions. All changes, including those omitted from this list, are highlighted in blue in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.