acceptodds
Under review as a conference paper at ICLR 2027

StratGS: Layered Surface Recovery for Transparent Objects from Gaussian Splatting

Abstract

Reconstructing transparent and semi-transparent scenes requires recovering multiple surfaces along a viewing ray. In nested scenes, dominant shell contributions can suppress the learning of interior structures. Reducing their contributions to a single depth can discard weak surface evidence or place geometry between distinct interfaces. We present StratGS, a Gaussian Splatting framework that combines nested-scene optimization with multiple-surface mesh reconstruction. After learning the whole scene, our nested refinement fixes the shell and background Gaussians and caps the shell's rendering contribution while optimizing the interior representation. Range-Aware Depth Extraction (RADE) retains local depth estimates from multiple search ranges for independent multi-view reconstruction. Independent TSDF integration produces fine meshes for all queries and a coarse candidate from the widest query using a broader truncation distance. To control the addition of coarse geometry, Fusion Using Scale-aware Exclusion (FUSE) preserves all fine triangles as a fixed coverage reference and clips the coarse mesh to retain fragments beyond that region. To evaluate coverage of transparent enclosures and the objects inside them, we introduce TransNest, eight synthetic multi-view scenes with separate ground-truth geometry for both regions. On existing benchmarks, the resulting meshes achieve mean Chamfer distances of 0.52 on αSurf and 1.553 mm on TransLab, and a mean one-way CD (s→d) of 3.276 mm on ClearPose. These are the lowest mean errors among the compared methods. On TransNest, StratGS achieves the lowest CD on all eight scenes and the lowest mean ground-truth-to-mesh distances for both enclosures and interiors.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Geometry Reconstruction, Multi-Layer Geometry, Neural Rendering, Transparent Object Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.80 over 5 reviews, above 85% of the others.
Reviewer_dNS5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper puts forward StratGS, a Gaussian Splatting framework built for surface reconstruction on transparent and semi‑transparent scenes. It especially targets the nested‑scene issue: the outer transparent shell messes up the model’s ability to learn inner structures.

The method has three main parts. First, after regular joint optimization, the authors add a two‑stage nested‑scene refinement step: lock down the shell and background Gaussians that have already been learned, limit how much the shell contributes to rendering, initialize brand‑new interior Gaussians using SfM, and keep optimizing only the inner representation. Second, they propose Range‑Aware Depth Extraction (RADE). Instead of squashing all Gaussian evidence along one ray into just one depth value, it pulls out multiple candidate depths by using several search ranges that keep getting wider step‑by‑step. Third comes FUSE: run TSDF integration independently for each depth query to get high‑quality fine meshes. Meanwhile generate a coarse mesh with a larger truncation distance, and only fill in areas missed by the fine reconstruction using this coarse mesh.

On top of that, the paper introduces TransNest, which contains eight synthetic nested scenes. It provides separate ground‑truth geometry for both enclosing shells and inner objects. This lets us evaluate how well the outer transparent structure and inner objects get reconstructed separately.

Experiments are run on αSurf, TransLab, ClearPose, TransNest and DTU benchmarks. StratGS achieves the lowest average geometric error across the first three transparent‑scene benchmarks, as well as the lowest CD on all eight TransNest scenes. Especially on TransNest, StratGS hits a mean CD of 3.970 mm, compared to 8.499 mm from TSGS. The interior GT‑to‑mesh distance drops from TSGS’s 12.317 mm down to 4.580 mm.

Critical strengths

This paper doesn’t only look at whether the transparent shell gets reconstructed. It also explicitly checks whether structures behind or inside that shell can be recovered. That’s harder to solve and way more useful for real‑world applications compared to evaluating only the outer transparent surface. They test across multiple datasets: αSurf, TransLab, ClearPose, TransNest and DTU, instead of relying only on one benchmark. Main results hold fairly consistently across tests. The nested‑refinement ablation gives solid experimental backing. The evaluation setup for TransNest is well‑thought‑out. Ground‑truth data for enclosures and interiors are given separately. That way you can’t hide bad inner‑object reconstruction behind an overall CD metric.

Critical weaknesses
  1. The cumulative, expanding‑range search setup still includes strong peaks from surfaces up front. Just grabbing the maximum‑weight window within each range doesn’t force later searches to pick deeper surfaces.
  2. A key preprocessing step: splitting shell / interior / background for nested refinement is not explained thoroughly. If manual region marking or scene‑specific information is required, the method becomes less general‑purpose. If it works fully automatically, they should lay out exactly how the algorithm runs.
  3. In Related Work they talk about NU‑NeRF as a nested transparent‑object reconstruction approach. But NU‑NeRF results are missing from the TransNest comparison in Table 2. This alone isn’t a reason to reject the paper, but authors should either add those numbers or explain why a fair comparison under the TransNest protocol isn’t feasible.
  4. FUSE prevents coarse‑mesh outputs from duplicating geometry already covered by fine meshes. However, that reference fine mesh itself comes from multiple fine meshes generated via separate RADE queries. If different thresholds end up reconstructing the exact same physical surface, readers want to know: do they have any mechanism to remove duplicates or merge overlapping fine outputs?
Questions for authors
  1. For Equation (4): wisw_i^s is calculated purely from GsG_s. When rendering jointly with GinG_{in} and GbG_b, how exactly do those shell weights (after attenuation) participate in front‑to‑back compositing? Could you write out the complete mathematical definition for the Stage‑II renderer?
  2. Before Stage II, how exactly do you tell shell, interior and background regions apart? Is this process 100 % automatic? Do you need manually set bounding boxes, masks, ground‑truth meshes, or other scene‑specific inputs?
  3. Could you run ablation studies for RADE and FUSE on TransNest? Please report CD, F1@5mm, Outer GT‑to‑mesh distance and Inner GT‑to‑mesh distance. Testing them on TransNest makes far more sense for what RADE and FUSE were built for, compared to only testing on TransLab.
  4. How sensitive is your method to hyper‑parameters τ1,τ2,τ3,h,β,δf,δc,r\tau_1,\tau_2,\tau_3,h,\beta,\delta_f,\delta_c,r? Are these parameters fixed for every scene within one dataset, or tuned separately scene‑by‑scene?
  5. Related Work introduces NU‑NeRF as a method built for nested transparent objects. Why are there no NU‑NeRF numbers in the TransNest Table 2? Are there issues with evaluation protocols or implementation?
  6. Figure 3 notes that both TSGS and StratGS use TransNormal‑2 priors for TransNest experiments. For baseline TSGS results on αSurf, TransLab and ClearPose — do those also use the exact same TransNormal‑2 priors? If not, adding a controlled comparison for TSGS + TransNormal‑2 would help separate performance gains brought by the normal prior versus improvements coming from StratGS itself.
4 comments

We sincerely thank the reviewer for the time and care invested in this careful and constructive review, and we are especially grateful for the recognition of interior recovery, the benchmark evaluation, and TransNest's separate ground truth. These comments have directly helped us sharpen both the evidence and the presentation, and we address each of them in turn below.

W1 & W4. Repeated queries and surface evidence

FUSE does not deduplicate the fine meshes. It removes only coarse regions near them. Later queries can repeat earlier peaks: query 3 repeats query 2 on 74.40% of jointly valid rays, while combined straight-ray GT-group coverage rises from 23.96% to 41.73%.

Q1. Stage-II shell compositing

We keep the standard 3DGS front-to-back order. For a ray, let Tis=∏j<i, gj∈Gs(1−αj)T_i^s=\prod_{j<i,\,g_j\in G_s}(1-\alpha_j) be the shell-only transmittance before shell Gaussian ii. The shell weights are w~is=αiTismin⁡(1,β∑j∈GsαjTjs),β=0.2.\widetilde w_i^s=\alpha_iT_i^s\min\left(1,\frac{\beta}{\sum_{j\in G_s}\alpha_jT_j^s}\right),\qquad \beta=0.2. If the accumulated shell weight is below β\beta, it is unchanged. Otherwise all shell weights are scaled proportionally. Interior and background Gaussians keep the standard 3DGS rule. Thus the shell remains visible but receives at most 0.20.2 total opacity along a ray.

In implementation, the scaled shell weights are converted back to effective alphas, and transmittance is recomputed over all depth-ordered Gaussians, including the interior and background:

\qquad C=\sum_i c_i\,\bar\alpha_i\prod_{j<i}\left(1-\bar\alpha_j\right)+T\,C_{\mathrm{bg}},$$ where $\bar\alpha_i=\bar\alpha_i^s$ for $i\in G_s$ and $\bar\alpha_i=\alpha_i$ otherwise, $c_i$ is the color of Gaussian $i$, $T=\prod_i(1-\bar\alpha_i)$, and $C_{\mathrm{bg}}$ is the background color. Together with the shell-weight formula above (Eq. (4)), these expressions for the effective alphas (Eq. (5)) and the final composition (Eq. (6)) give the complete Stage-II renderer. ## Q2. Automatic separation of shell, interior, and background The separation is fully automatic. It uses only the input RGB images, cameras, the Stage-I representation, and the SfM point cloud—no bounding boxes, masks, or GT meshes. Given these inputs, the three regions are separated in two steps. **Interior.** We first select SfM points that Stage I does not explain: $$\mathcal P_{\mathrm{in}}=\{p\in\mathcal P_{\mathrm{SfM}}:\ R(p)\land B(p)\land M(p)\},$$ where $R(p)$ is a high Stage-I rendering residual, $B(p)$ places the point behind the first depth peak, i.e. behind the shell, and $M(p)$ requires depth, normal, and color support from multiple viewing directions. Accepted points initialize the new interior group $\mathcal{G}_n$. **Shell, replacement, and background.** Among the Stage-I Gaussians, multi-view depth, normal, color, and SfM evidence identify the supported shell Gaussians $\mathcal{G}_s$. Gaussians with behind-shell support but no consistent front-side evidence are removed as conflicting interior representations $\mathcal{G}_r$. All remaining Gaussians, including the background, form the retained context group $\mathcal{G}_c=\mathcal{G}_0\setminus(\mathcal{G}_s\cup\mathcal{G}_r)$. The Stage-II representation is therefore $\mathcal{G}_s\cup\mathcal{G}_c\cup\mathcal{G}_n$, with $\mathcal{G}_r$ excluded before Stage II. **Training permissions.** $\mathcal{G}_s$ keeps position, shape, and opacity fixed and updates color only. $\mathcal{G}_c$ keeps position and shape fixed, while color and opacity may be updated. $\mathcal{G}_n$ is fully trainable and is the only group subject to densification and pruning. Stage II optimizes the full image jointly rendered from $\mathcal{G}_s\cup\mathcal{G}_c\cup\mathcal{G}_n$ with the RGB term in Eq. (3). **Controls.** Retaining $\mathcal{G}_r$ and optimizing it with the same $\mathcal{G}_n$ changes CD/F1 from **3.97/0.834** to **4.60/0.812**, while Inner rises from 4.58 to 6.73 mm. Additionally unfreezing the position and shape of $\mathcal{G}_c$ changes CD/F1 from **3.97/0.834** to **5.00/0.754**. So fixing the geometry of $\mathcal{G}_c$ mainly protects the overall reconstruction. **Table 1. Controls on Stage-II group assignment and permissions (TransNest).** | Setting | CD (mm) ↓ | F1\@5mm ↑ | Inner (mm) ↓ | | ----------------------------------------------------------- | --------: | --------: | -----------: | | Retain $\mathcal{G}_r$ and optimize it with $\mathcal{G}_n$ | 4.60 | 0.812 | 6.73 | | Unfreeze position and shape of $\mathcal{G}_c$ | 5.00 | 0.754 | **4.54** | | Full method (Ours) | **3.97** | **0.834** | 4.58 |

Q3. RADE/FUSE ablation

Following your suggestion, we added the full TransNest ablation with trained models, cameras, and extraction regions fixed.

The results show that both modules contribute: with FUSE off, enabling RADE reduces CD by 51.89% and improves F1 by 113.73%; with RADE on, enabling FUSE reduces CD by 71.36% and improves F1 by 67.47%.

Table 2. RADE/FUSE ablations on TransNest.

RADEFUSECD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
OffOff28.810.23358.5812.22
OffOn26.370.28153.649.32
OnOff13.860.49824.326.08
OnOn3.970.8342.624.58

Q4. Hyperparameter sensitivity

We use one parameter configuration per dataset and do not tune them scene by scene. ClearPose contains more complex real scenes and more noise, so we use five thresholds. The other datasets use three. We also add a new ablation on TransNest: increasing the count from three to five reduces CD by 11.6% and improves F1 by 1.3% (Table 3).

Table 3. Threshold-count comparison on TransNest, with trained models and other extraction settings fixed.

ThresholdsCD↓ (mm)F1↑
33.970.834
53.510.845

Q5. NU-NeRF comparison

We add NU-NeRF comparisons on TransNest. Our mean reconstruction error is lower (Table 4). On the same TransNest scenes and an A100 GPU, the complete NU-NeRF two-stage run takes at least 40 hours, whereas StratGS takes about 30 minutes. We will add these runtime results to the supplement.

Table 4. Mean reconstruction results on TransNest.

MethodCD (mm) ↓F1@5mm ↑
NU-NeRF24.820.224
StratGS3.970.834

Q6. Normal-prior comparison

We report TSGS and StratGS directly for each benchmark in Table 5. For ClearPose, TransLab, and αSurf, we additionally provide paired controls that change only the normal prior.

Table 5. TSGS–StratGS comparison under the listed normal priors.

DatasetComparisonbidirectional CD↓F1↑
TransNestTSGS + TN2 → StratGS + TN28.50 → 3.970.611 → 0.834
ClearPoseTSGS original prior → TSGS + TN2 → Ours13.44 → 12.61 → 9.710.366 → 0.378 → 0.623
αSurfTSGS original prior → TSGS + TN2 → Ours0.90 → 1.03 → 0.520.774 → 0.794 → 0.723
TransLabTSGS original prior → TSGS + TN2 → Ours1.69 → 1.66 → 1.550.959 → 0.959 → 0.959

αSurf uses CD (×100) and [email protected]; the other datasets use CD in mm and F1@5mm. The last three paired controls fix the other TSGS settings and change only the normal prior. On αSurf, TN2 improves F1 by 0.020 but increases CD by 0.13, so it is not an overall gain on this benchmark.

Thank you again for your constructive and helpful reviews. As the discussion period draws to a close, we’d welcome any follow-up questions about this analysis or the method. If our response has addressed your concerns, we hope you’ll revisit your assessment in light of the additional evidence. Thank you again for your time and help in improving the quality of this paper.

Thank you very much for your reply. It has addressed my concerns perfectly.

Reviewer_GkBAOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies surface reconstruction for transparent scenes. In these scenes, a single ray can pass through multiple surfaces, such as a transparent shell and the object inside it. Reducing them to a single depth per ray can lose weak surfaces or place geometry between them. In nested scenes, the shell can also suppress the learning of the interior during training. Built on TSGS, the proposed method has three parts. First, for nested scenes, it trains the full scene and then freezes the shell. It limits the shell’s rendering weight and trains the interior Gaussians separately. Second, it uses several accumulated-weight thresholds to extract multiple depths per ray, and the depths from each threshold are fused into a separate TSDF mesh. Third, it builds a coarse mesh with a larger truncation distance and adds only the regions not covered by the fine meshes. The authors also introduce TransNest, a synthetic dataset with eight nested scenes and separate ground truth for the shell and the interior. The method achieves the best mean results on αSurf, TransLab, ClearPose, and TransNest.

Critical strengths
  1. This paper gives a clear analysis of why nested transparent scenes are hard. It points out two separate causes. During training, the shell can take most of the rendering weight and suppress the interior. During extraction, a single depth per ray can drop weak surfaces or place geometry between them. The method is designed to address both.

  2. The authors evaluate the shell and the interior separately on the new TransNest dataset. StratGS gets the best results on all scenes, with clear gains on both the shell and the interior. Table 5 further shows that capping the shell weight in Stage 2 clearly improves the interior reconstruction.

  3. The authors build the TransNest dataset, which provides separate GT meshes for the shell and the interior, so the two can be evaluated separately.

Critical weaknesses
  1. The performance of the base model is unclear. The main text does not say whether Stage 2 is used on the three datasets in Table 1. But Supplement A.1 only describes two training stages for TransNest, and refers to “standard training configurations” for αSurf, TransLab, and ClearPose. This implies Stage 2 is not used on these three datasets. Please confirm this. If so, the authors should report a base model on these datasets without RADE, FUSE. This means the same training setup, but with a single depth per ray fused into one TSDF mesh. This would show how much RADE and FUSE really add. Right now, Table 4 only reports ablations on TransLab, and the gains are small (1.60 without RADE, 1.56 without FUSE, and 1.55 for the full method). There is also no setting that removes both.

  2. Also, the biggest gain in Tab.1 is on ClearPose, which uses the one-way CD_{s→d}. The paper does not say which direction this is. Based on Supplement C.1, it seems to measure the distance from the GT to the prediction and only reflects coverage. In addition, Supplement A.3 shows that ClearPose uses five thresholds and doubled truncation distances, which gives six candidate meshes. So I suspect the large gain on ClearPose mainly comes from the extra coverage of fusing more surfaces, even if some of them are noisy. Please clarify the direction of CD_{s→d}. It would also help to report precision, recall, and F-score with a distance threshold inside the evaluation box, as done with F1 on TransNest. These are less affected by incomplete GT.

  3. The DTU results do not evaluate any of the proposed components. Supplement A.2 says that DTU does not use the nested refinement, RADE, or FUSE. Please clarify this in the main text, or report results with the full pipeline on DTU.

  4. The experiment section mostly restates the numbers in the tables, with little analysis. For example, it does not explain why GLINT performs poorly, why StratGS is worse than some baselines on certain scenes, or why the gains differ so much across datasets.

5、The normal prior constrains the outermost surface, and the depth–normal consistency loss passes this constraint to the depth. The NCC loss from PGSR constrains surface position through photometric consistency. In transparent regions, the image mostly shows the object behind, so this loss actually fits the inner surface better. Both losses act on the same rendered depth, and the design does not separate them, so they may conflict. Notably, Supplement A.1 shows that NCC is turned off in the first stage on TransNest, while it is 0.15 on the other datasets. This is not explained. More importantly, Stage II needs geometric constraints the most, but it only uses the RGB loss. Even if the normal prior does not suit the interior, the flattening and multi-view terms could still be applied to the interior Gaussians. In the first two rows of Figure 3, the inner surfaces seem less detailed and less smooth than those of PGSR. The Inner metric in Table 2 only measures coverage, so it does not reflect this.

Questions for authors

Overall, this paper studies the reconstruction of multiple surfaces in transparent scenes. The problem is well motivated, and TransNest allows the shell and the interior to be evaluated separately. On this dataset, the interior reconstruction clearly improves. However, the TransNest scenes are quite simple, and the visual comparisons are limited. Only two TransNest scenes are shown, and there are no visual results for TransLab, the ablations, or failure cases. In addition, it is hard to tell where the gains on the other benchmarks come from, and some design issues are not addressed. These are the main reasons for my current score.

8 comments

We sincerely thank the reviewer for the time and care invested in this detailed and constructive review, and we are especially grateful for the recognition of our analysis of training and extraction failures and of our separate shell/interior evaluation. These comments have directly helped us sharpen both the evidence and the presentation, and we address each of them in turn below.

W1.1. Base training and extraction settings

TransNest is designed specifically for nested structures. We use two-stage optimization that treats the outer shell and interior separately, including separate interior training and shell-contribution control, followed by RADE and FUSE for geometry extraction. αSurf, TransLab, and ClearPose use single-stage training followed by RADE and FUSE, without this nested optimization. DTU uses the base opaque-object reconstruction configuration.

W1.2. Base control without RADE/FUSE

TransLab is a relatively simple transparent-object dataset with few nested structures, which helps explain the small gains from RADE and FUSE in the original ablation. We therefore add ablations on the more challenging TransNest, as well as αSurf and ClearPose (Tables 1–3), including a base version with both RADE and FUSE disabled. Compared with the base version with both RADE and FUSE disabled, the full method reduces mean CD by approximately 86.22% on TransNest and 66.01% on αSurf, and mean GT-to-prediction distance by 30.87% on ClearPose.

Table 1. RADE/FUSE ablations on TransNest.

RADEFUSECD (mm) ↓F1↑Outer (mm) ↓Inner (mm) ↓
OffOff28.810.23358.5812.22
OffOn26.370.28153.649.32
OnOff13.860.49824.326.08
OnOn3.970.8342.624.58

Table 2. Fixed-field RADE/FUSE ablations on αSurf.

RADEFUSECD (×100) ↓[email protected] ↑
OffOff1.530.466
OffOn0.950.510
OnOff0.770.598
OnOn0.520.723

Table 3. Fixed-field extraction controls on TransLab and ClearPose.

RADEFUSETransLab CD (mm) ↓F1↑ClearPose GT→pred. ↓F1↑
OffOff2.120.9094.730.569
OffOn1.600.9104.690.569
OnOff1.560.9253.810.606
OnOn1.550.9593.270.623

W2. ClearPose evaluation and extra surfaces

Thank you for pointing out the limitations of one-way evaluation. This setting follows TSGS's ClearPose evaluation protocol. CDs→d\mathrm{CD}_{s\to d} is the mean nearest-neighbor distance from GT to reconstruction, which mainly measures coverage. ClearPose contains challenging real scenes with complex lighting, refraction, and occlusion between transparent objects. We therefore use five thresholds and a coarse candidate with doubled truncation distance to recover missed surfaces.

For the complete FUSE reconstructions, we report four-scene means with the same GT samples and evaluation boxes, using a 5 mm threshold. Compared with TSGS, our method reduces bidirectional CD by 27.75% and improves F1 by 70.22% (Table 4).

Table 4. Bidirectional distance, precision, and recall on ClearPose.

MethodBidirectional CD (mm) ↓Precision ↑Recall ↑F1↑
TSGS13.440.2930.4780.366
StratGS9.710.5010.8330.623
Improvements vs. TSGS27.75% ↓70.99% ↑74.27% ↑70.22% ↑

W3. Scope of the method on DTU

We report DTU results to verify that our method preserves reconstruction performance on opaque scenes. We therefore use the base configuration without nested refinement, RADE, or FUSE. We now state this setting in blue in Sec 4.2 “Opaque-object reconstruction” part of the main paper.

W4. GLINT performance and cross-dataset variation

As shown in Figure 1 & Figure 4 of the paper, GLINT’s poor performance stems from missing outer surfaces and small objects, resulting in insufficient coverage. The datasets differ in reconstruction difficulty. TransNest contains multilayer nested structures, while ClearPose includes challenging real captures with complex lighting, reflections, refraction, and occlusion between transparent objects. Our method therefore shows larger gains on these datasets. TransLab is simpler and contains fewer nested structures, so the gains are smaller. In some scenes, PGSR reconstructs smoother interiors with higher local accuracy but misses the outer shell. We recover both the shell and interior, with the main gains in overall structural completeness and interior coverage. We have added this analysis to the revised manuscript.

W5.1. Stage-I losses

In nested transparent scenes, an image patch may mix contributions from the shell and interior, disrupting NCC-based cross-view matching. Our TransNest ablation shows that adding NCC increases reconstruction error and lowers F1 (Table 5). We therefore disable NCC in Stage I for these scenes.

Table 5. Stage-I NCC ablation on TransNest.

Stage-I NCCCD (mm) ↓F1↑
Off (weight 0)3.970.834
On (weight 0.15)6.330.715

W5.2. Normal supervision in Stage II

The normal prior we use predicts the first surface encountered along each viewing ray from RGB, without providing separate normals for interior surfaces behind the shell. It therefore cannot directly supervise the interior, so Stage II uses RGB reconstruction loss.

W5.3. Geometric constraints for the interior

Both constraints can be applied to interior Gaussians. We therefore added ablations of each term and their combination on TransNest (Table 6).

Table 6. Stage-II geometry loss ablations on TransNest.

Stage-II settingCD (mm) ↓F1↑Outer (mm) ↓Inner (mm) ↓
RGB3.970.8342.624.58
+ Flattening4.110.8322.594.50
+ Multi-view geometry3.570.8582.925.69
+ Both3.630.8532.905.69

The results show that multi-view geometry lowers overall CD but increases outer and inner errors, while flattening gives no overall improvement. We therefore retain RGB-only training as a balance between overall accuracy and recovery of both layers.

W5.4. Interior detail and smoothness

PGSR produces smoother interiors, but its reconstruction misses the outer shell and does not jointly recover the shell and its image contribution. In real observations, shell reflections, refraction, and occlusion interfere with interior reconstruction, making joint recovery more challenging. Our strength is recovering the shell while improving interior coverage.

Q1. Additional visual evidence

Thank you for pointing this out. In the revised manuscript, we have added more visual results to Figure 4, including additional TransNest and TransLab scenes. More visual results are available in the supplement.

Thank you again for asking how much interior detail StratGS recovers. Our response explains how Stage II supervises interior geometry and includes visual comparisons alongside the quantitative results. Before the discussion closes, we’d be happy to discuss any scene or result you feel deserves a closer look. If our response has addressed your concerns, we hope you’ll revisit your assessment in light of the additional evidence and consider raising your score.

Reviewer_i5oVOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes StratGS, a 3D Gaussian Splatting-based surface reconstruction framework for transparent and semi-transparent nested scenes. It has three components: (1) a two-stage nested-scene optimization that freezes shell/background Gaussians, caps the shell’s rendering contribution, and refines only the interior; (2) Range-Aware Depth Extraction (RADE), which preserves multiple depth candidates from progressively broader ray search ranges; and (3) Coverage-Constrained Fusion (FUSE), which keeps fine TSDF meshes as a fixed reference and supplements uncovered regions with a clipped coarse mesh. The authors also introduce TransNest, eight synthetic laboratory/robotic scenes with separate ground-truth meshes for enclosures and interiors. Experiments on αSurf, TransLab, ClearPose, TransNest, and DTU show strong improvements on transparent/nested benchmarks, though not on DTU.

Critical strengths
  1. The paper identifies a real and important failure mode in nested transparent reconstruction: the outer shell dominates joint optimization, suppressing gradients and evidence for interior surfaces. The two-stage freeze/attenuation strategy directly targets this issue, and the TransNest ablation is compelling: without nested refinement, CD is 6.545 mm and inner GT-to-mesh distance is 26.122 mm; with full refinement, these drop to 3.970 mm and 4.580 mm. This is a substantive, problem-specific contribution rather than a generic 3DGS variant.
  2. RADE and FUSE address a practical limitation of single-depth extraction in transparent scenes: multiple surfaces at different depths can be mixed or missed. Retaining multiple range-restricted depth candidates and using a fixed fine-mesh reference to clip coarse supplementation is a sensible design, and the method obtains the lowest mean CD on αSurf, TransLab, ClearPose, and all eight TransNest scenes among the compared methods.
Critical weaknesses

The paper’s central claim—that RADE/FUSE and nested refinement recover fine multi-surface geometry—is not adequately supported by the experiments. The RADE/FUSE ablation is marginal (Table 4: full 1.55 mm vs. 1.56 mm without FUSE and 1.60 mm without RADE), with no per-scene variance or statistical test, and key hyperparameters are dataset-specific (three thresholds for αSurf/TransLab/TransNest but five for ClearPose, plus per-dataset h and r), so the necessity and robustness of these components remain unclear. TransNest contains only eight synthetic Blender scenes, and no real nested transparent captures are evaluated, which limits claims about generalization under real refraction, reflection, and calibration noise. Comparisons with TSGS/TSPE-GS/GLINT also need stronger evidence that priors, iterations, and compute are comparable. On DTU, StratGS reports mean CD 0.47 vs. 0.35 for TSGS and 0.37 for PGSR, so the method does not improve opaque reconstruction; this limitation should be explicitly discussed rather than only emphasizing transparent benchmarks. Finally, the unusually broad AI-use statement raises verification concerns that are not addressed.

Questions for authors
  1. Table 4 shows only 0.01 mm and 0.05 mm mean CD differences for FUSE and RADE. Please report per-scene results, standard deviations/significance, and TransNest ablations for RADE/FUSE. Also clarify how the number of thresholds, h, and r were selected (validation vs. per-benchmark tuning) and why ClearPose uses five thresholds while the other benchmarks use three.
  2. TransNest contains only eight synthetic scenes. What evidence supports generalization to real nested transparent objects, and how do the Blender materials/rendering settings match real glass/acrylic under capture noise? Please also discuss why DTU mean CD (0.47) is worse than TSGS (0.35) and PGSR (0.37).
  3. The AI-use statement says GPT-6-astra assisted with method and experiment design, coding, result analysis, and manuscript preparation. Please specify exactly which parts were AI-assisted and how the authors verified the correctness and originality of those contributions under the ICLR policy.
4 comments

We sincerely thank the reviewer for the time and care invested in this careful and constructive review, and we are especially grateful for the recognition of the inner-learning problem and of the motivation for multi-surface extraction. These comments have directly helped us sharpen both the evidence and the presentation, and we address each of them in turn below.

W1. Normal-prior comparison

The CD/F1 comparison among TSGS with its original prior, TSGS with TN2, and our method is shown in Table 1. On TransNest, both TSGS and ours use TN2.

Table 1. Normal-prior comparison.

DatasetMethodbidirectional CD ↓F1 ↑
αSurfTSGS original prior0.900.774
TSGS + TN21.030.794
Ours0.520.723
ClearPoseTSGS original prior13.440.366
TSGS + TN212.610.378
Ours9.710.623
TransLabTSGS original prior1.690.959
TSGS + TN21.660.959
Ours1.550.959
TransNestTSGS + TN2 (original prior)8.500.611
Ours3.970.834

αSurf uses CD (×100) and [email protected]; the other datasets use CD in mm and F1@5mm.

On the three datasets with a direct prior replacement control, TN2 improves the TSGS baseline on ClearPose and TransLab but does not close the CD gap to ours. On αSurf the effect is mixed: F1 rises from 0.774 to 0.794, above our 0.723, while CD rises from 0.90 to 1.03, roughly double ours, so the prior change is not an overall gain on this benchmark. On TransNest, the comparison uses the same TN2 prior for both methods.

TSPE-GS does not use an external predicted prior on TransNest. Its αSurf main-table results are taken directly from the original paper. GLINT retains the official DiffusionRenderer prior, which provides normals, depth, and material attributes. The prior configuration of all three methods is listed explicitly in the table.

W2. Compute budget and speed

We added a runtime comparison on scene 05.

Table 2. Runtime on TransNest scene 05 (minutes).

MethodPriorTraining / initializationExtraction / fusionTotal
2DGS—9.13.512.6
PGSR—26.42.528.9
GGGS—21.51.623.1
TSPE-GS—13.214.928.2
TSGSTN228.14.632.7
StratGSTN220.99.230.1
GLINTDiffusionRenderer222.82.5225.3
NU-NeRF—2493.710.82504.5

StratGS has a similar total runtime to the strong TSGS baseline, 30.1 versus 32.7 minutes. Its total time is about 1.2% of NU-NeRF's, corresponding to a reduction of about 98.8%.

Q1.1 Effects and stability of RADE and FUSE

TransLab is a relatively simple transparent dataset that rarely contains nested scenes, so we additionally add a two-factor ablation of RADE and FUSE on TransNest, keeping the trained models fixed and comparing the four on/off combinations:

Table 3. RADE/FUSE ablations on TransNest.

RADEFUSECD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
OffOff28.810.23358.5812.22
OffOn26.370.28153.649.32
OnOff13.860.49824.326.08
OnOn3.970.8342.624.58

We fix the trained model with RADE enabled and compare FUSE off versus on using three predeclared seeds. Values with ± report the mean and sample standard deviation across the three runs.

Table 4. Three-seed Stage-II repetitions and matched-model FUSE controls.

ExtractionCD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
FUSE off13.29±2.0590.505±0.04823.506±3.9396.041±0.031
FUSE on4.022±0.0520.835±0.0022.599±0.0114.498±0.025

Across the three repeated runs, enabling FUSE reduces mean CD by 9.27 mm (about 70%), increases mean F1 by 0.330 (65.35%), and reduces Outer and Inner error by 20.907 mm (88.94%) and 1.543 mm (25.54%), respectively. The CD gap is about 4.5 times the seed-to-seed variation of the less stable setting (CD ±2.059 mm for FUSE off vs. ±0.052 mm for FUSE on).

Q1.2. How sensitive are the settings?

We choose parameters per dataset and use one configuration within each dataset. ClearPose contains more complex real scenes with more noise, so we use five thresholds. The other benchmarks use three. The window width hh is set per benchmark: 0.025 for αSurf/TransNest, 0.03 for TransLab and 0.003 for ClearPose, in their depth-coordinate units, and is fixed within each benchmark. The fusion exclusion radius is r=4vr=4v throughout, where vv is the TSDF voxel size. It is not tuned separately per scene. We add a three-versus-five-threshold comparison on TransNest. More thresholds improve mean CD and F1 (Table 5).

Table 5. Threshold-count comparison on TransNest, with trained models and other extraction settings fixed.

ThresholdsCD↓ (mm)F1↑
33.970.834
53.510.845

Q2.1. Real nested scenes

The main paper already evaluates real-capture generalization on ClearPose (Table 1 of the main paper): our mean GT-to-prediction distance decreases from 7.96 mm for the strongest baseline, TSGS, to 3.27 mm—a 58.92% reduction. ClearPose includes nested transparent objects and mutual occlusion under complex lighting, reflections, and refraction, providing real-scene evidence beyond synthetic data.

In addition, we evaluate three multilayer nested ClearPose scenes (Set8_scene4, Set9_scene7,8) with transparent outer covers to test interior recovery under outer-layer occlusion. Table 6 reports the latest completed results. Compared with PGSR, mean CD decreases by approximately 12.16% and F1 increases by 27.03% across these three highly challenging scenes.

Table 6. Mean reconstruction results on three additional real nested scenes.

MethodCD (mm) ↓F1 ↑
TSPE-GS14.220.359
2DGS16.240.309
PGSR13.730.333
StratGS (Ours)12.060.423

Q2.2. DTU results

On DTU, the mean CDs are 0.51 and 0.55 mm for TSGS (original prior) and PGSR, respectively. Our mean CD is 0.47 mm, outperforming both baselines. DTU contains opaque objects, so we use the base configuration without two-stage nested refinement, RADE's multi-threshold extraction, or FUSE.

Q3. AI use and author checks

AI assistance primarily supported the agentic construction of TransNest, including drafting and revising Blender scripts and running previews and consistency checks under scene designs and constraints we specified. AI also assisted implementation and result organization. Writing assistance mainly involved language polishing. We manually checked the equations and code and take responsibility for the method’s originality, interpretation of results, and final manuscript.

Thank you again for your questions about stability, computational cost, and performance on real nested scenes. Our response includes results across three random seeds, runtime comparisons, and an evaluation on real scenes. With the discussion period nearing its end, we’d be glad to clarify any of these results. If our response has addressed your concerns, we hope you’ll revisit your assessment in light of the additional evidence and consider raising your score.

Thank you for the detailed response and the additional experiments. The newly added three-seed FUSE control, runtime comparison, four-way RADE/FUSE ablation on TransNest, and results on three real nested ClearPose scenes address my earlier concerns about insufficient ablation, parameter sensitivity, and real-scene generalization well. Overall, the authors’ response resolves my main technical concerns, and I am inclined to accept this work.

Reviewer_PzSWOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

StratGS addresses the reconstruction of transparent surfaces together with structures visible behind them, particularly opaque objects enclosed by transparent shells. The paper combines three components: nested-scene refinement, which freezes shell and background Gaussians while attenuating shell contributions and optimizing newly initialized interior Gaussians; Range-Aware Depth Extraction (RADE), which extracts depth candidates from multiple cumulative-weight search ranges; and Fusion Using Scale-aware Exclusion (FUSE), which preserves independently reconstructed fine meshes and supplements them with selected coarse geometry. The authors also introduce TransNest, eight synthetic scenes with separate enclosure and interior ground-truth geometry.

The reported results are promising, especially on TransNest. The main questions concern the reproducibility of the scene partitioning procedure, the attribution of improvements to training versus extraction, and the incremental value of the proposed extraction and fusion mechanisms.

Critical strengths
  • The paper addresses a meaningful reconstruction problem: recovering transparent enclosures together with the objects inside them. Separating weak interior learning from the loss of depth evidence during extraction is well motivated. The nested refinement strategy, independent depth-query integration, and coverage-based mesh supplementation form a coherent approach.
  • The evaluation spans several established benchmarks and the new TransNest dataset. The TransNest results are particularly encouraging: mean CD improves from 8.499 mm for TSGS to 3.970 mm, while F1@5mm increases from 0.611 to 0.834. StratGS achieves the lowest CD on all eight scenes, and the separate enclosure/interior metrics provide useful evidence beyond aggregate reconstruction quality (Table 2).
  • Several experimental controls are appropriate. The refinement ablations share the initial checkpoint, continuation budget, and camera sequence; the extraction/fusion ablations reuse trained representations. The supplement also explicitly matches TransNormal-2 priors between TSGS and StratGS on TransNest.
  • The pipeline is clearly organized, and the supplement clarifies important details concerning depth conventions, effective shell opacities, and TSDF integration. TransNest’s separate ground-truth geometry for enclosures and interiors is potentially useful for studying nested reconstruction failures.
Critical weaknesses

W1. The contribution of shell attenuation to improved interior learning is not fully isolated.

The paper motivates shell attenuation as a way to improve gradients reaching interior Gaussians. However, the no-attenuation ablation changes the cap during both training and depth extraction (Supplement §A.2). This also changes RADE’s cumulative-weight thresholds and selected windows, even without changing the learned geometry.

Table 5 therefore supports the combined intervention but does not distinguish better interior learning from improved extraction. Evaluating the existing capped and uncapped checkpoints under a common extraction setting, or using a small training-cap × extraction-cap comparison, would clarify the mechanism.

W2. The evidence for RADE and FUSE is weaker than the end-to-end results.

RADE searches nested depth prefixes and selects the strongest window within each prefix. A strong foreground window can therefore remain selected across all queries, preventing weaker, deeper surfaces from being recovered. The supplement acknowledges overlapping candidates, but the paper does not quantify how often the queries provide distinct, correct surface evidence.

The controlled reconstruction ablations are limited to TransLab: removing RADE increases CD from 1.55 to 1.60 mm, while replacing FUSE with direct concatenation increases it to 1.56 mm (Table 4). These results are encouraging, but a fine-only control and corresponding TransNest ablations would better establish the value of coarse supplementation and multiple queries. A shared-representation comparison with the closest extraction alternatives would further clarify the methodological contribution.

W3. Important implementation and baseline details remain underspecified.

The nested refinement depends on correctly partitioning shell, interior, and background geometry. Although the supplement specifies minimum track support and a residual-percentile threshold, the depth/normal/color consistency tolerances, front-surface association, and Gaussian assignment rules remain unclear. Errors here can freeze incorrect geometry or prevent interior initialization.

Baseline settings also need clarification. Matched normal priors are explicitly documented for TSGS and StratGS on TransNest, but the corresponding configurations on the other benchmarks are less clear. These omissions limit reproducibility and attribution of the improvements; they do not establish that the comparisons are unfair.

Questions for authors
  1. How much of the attenuation benefit remains when capped and uncapped training checkpoints are evaluated with the same extraction cap?

  2. Can you provide a focused TransNest analysis of RADE’s additional surface recovery, together with a fine-only versus concatenation versus FUSE comparison?

  3. Can you specify the partitioning and depth-validity procedures precisely, including numerical thresholds, ambiguous assignments, and any manual inputs?

  4. Which code versions, normal priors, and training/extraction settings were used for TSGS and TSPE-GS on each benchmark? Can the extraction alternatives be compared on a shared Gaussian representation?

4 comments

We sincerely thank the reviewer for the time and care invested in this detailed and constructive review, and we are especially grateful for the recognition of the coherent design, the TransNest evaluation, and our controlled comparisons.

W1. Training vs. extraction attenuation

Following your suggestion, we added a training-cap × extraction-cap ablation on TransNest with three random seeds (Table 1), testing the two effects separately.

  • Training effect: With the extraction cap fixed at 0.2 or 1.0, training attenuation lowers mean CD by 6.29% and 5.75%, respectively. At extraction cap 1.0, it also lowers Inner error by 14.56%.
  • Extraction effect: With models trained at cap 0.2 fixed, changing only the extraction cap from 1.0 to 0.2 lowers Inner error by 20.06%, showing that extraction itself also affects interior coverage.

With the extraction cap fixed at 0.2 or 1.0, training attenuation lowers mean CD by 6.29% and 5.75%, respectively, so its benefit is not merely a side effect of extraction. We keep the extraction cap at 0.2 in the full method because it reduces Inner error from 5.627 to 4.498 mm, at the cost of a slightly higher CD (4.022 vs. 3.590 mm).

Table 1. Additional Training-cap × extraction-cap ablation on TransNest with three random seeds.

Training capExtraction capCD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
0.20.24.022±0.0520.835±0.0022.599±0.0114.498±0.025
0.21.03.590±0.0170.858±0.0012.351±0.0385.627±0.040
1.00.24.292±0.0190.821±0.0032.441±0.0224.151±0.012
1.01.03.809±0.0380.847±0.0022.424±0.0336.586±0.205

W2.1. Do later queries provide new surface evidence?

RADE retains the strongest candidate within each search range. For a ray with NN contributions sorted by depth, let di,wid_i,w_i denote depth and rendering weight, hh the window width, and τ\tau the cumulative-weight threshold. The query selects a contiguous window W=[l,r]W=[l,r]:

W∗(τ)=arg⁡max⁡W=[l,r],;1≤l≤r≤N\dr−dl≤h,;∑j<rwj≤τ;∑i=lrwi.W^*(\tau)=\underset{\substack{W=[l,r],\\;1\le l\le r\le N\\\d_r-d_l\le h,\\;\sum_{j<r}w_j\le\tau}}{\arg\max}\\;\sum_{i=l}^{r}w_i.

A small τ\tau searches the near part of the ray. A larger τ\tau also checks farther regions. If a farther surface has stronger evidence, the query selects a new depth. Otherwise, it returns the same depth.

Across the random views, we checked whether predictions hit GT surface locations along the rays (error ≤5 mm). Counting each GT location only once, the first query covers 23.96%. Adding the second raises coverage to 37.91%, and adding the third raises it to 41.73%. Later queries therefore find more real surface locations.

W2.2. RADE and FUSE ablations

We add three further controls on TransNest with fixed trained models and the same evaluation protocol: a fine-only setting that keeps only the fine meshes from the RADE queries with no coarse supplement, direct concatenation of the fine and coarse meshes, and the separately disabled RADE/FUSE settings (Table 2).

Table 2. RADE/FUSE ablations on TransNest.

RADEFusionCD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
OffSingle TSDF28.810.23358.5812.22
OffFUSE26.370.28153.649.32
OnSingle TSDF13.860.49824.326.08
OnFine only (no coarse)4.800.8053.684.92
OnDirect concatenation4.900.7954.035.11
OnFUSE3.970.8342.624.58

Table 3. Fixed-model RADE/FUSE extraction ablations on αSurf, TransLab, and ClearPose.

DatasetRADEFUSECD ↓F1 ↑
αSurfOffOff1.530.466
αSurfOffOn0.950.510
αSurfOnOff0.770.598
αSurfOnOn0.520.723
TransLabOffOff2.120.909
TransLabOffOn1.600.910
TransLabOnOff1.560.925
TransLabOnOn1.550.959
ClearPoseOffOff4.730.569
ClearPoseOffOn4.690.569
ClearPoseOnOff3.810.606
ClearPoseOnOn3.270.623

αSurf uses CD (×100) and [email protected]; TransLab and ClearPose use CD in mm (ClearPose GT→pred.) and F1@5mm.

“Single TSDF” integrates all RADE candidates into one TSDF and thus keeps a single mesh, whereas “Fine only” keeps the separate fine meshes without any coarse supplement. Relative to fine-only, FUSE reduces CD by 17.29% and increases F1 by 3.60%. Relative to direct concatenation, FUSE reduces CD by 18.98% and increases F1 by 4.91%.

After RADE, FUSE further reduces CD on all three benchmarks: 0.77→0.52 mm (32.47%) on αSurf, 1.56→1.55 mm (0.64%) on TransLab, and 3.81→3.27 mm (14.17%) on ClearPose. The gain is small on TransLab, where nested structures are rare, but consistent in direction.

W3.1. Automatic inputs and seed selection

TransNest seed selection and grouping use no manual boxes, masks, or GT meshes. We select SfM points to initialize interior regions that Stage I has not explained well. Selection has three checks:

  1. Find unexplained image regions: local rendering error is in the image's top 30%, while the point color agrees with the photograph (mean RGB difference <0.045).
  2. Exclude wall points: use the frontmost Stage-I surface as a wall reference and retain candidates beyond its neighborhood.
  3. Confirm across views: appearance and depth checks each need at least three supporting views, with spatial support from different sides (>90°). Reject a point if at least three views place it in front of the shell.

As illustrated in Figure 3 of the paper, TransNest seed selection starts from the Stage-I Gaussians and SfM point cloud. It first locates high-residual, color-consistent regions, removes candidates near the front shell surface, and then confirms interior candidates using cross-view depth, appearance, and directional support. The process is automatic and uses no scene-specific manual input.

W3.2. Gaussian groups and training permissions

We divide the Stage-I Gaussians into four groups using multi-view residuals, color agreement, and depth/normal support:

  • Shell Group Gs\mathcal{G}_s: reliable shell Gaussians. Fix position, shape, and opacity, and train color only.
  • Retained Group Gc\mathcal{G}_c: retained original Gaussians, including the background. Keep position and shape fixed; color and opacity remain trainable. No densification or pruning.
  • Replacement Group Gr\mathcal{G}_r: old Gaussians identified as erroneous interior representations by multi-view evidence. Remove them before Stage II.
  • New Interior Group Gn\mathcal{G}_n: new Gaussians initialized from selected SfM points, with all parameters trainable and densification/pruning enabled.

As illustrated in Figure 3 of the paper, we use multi-view residuals, color agreement, and depth/normal support to assign the four Gaussian groups and determine which groups are retained, replaced, or optimized. The corresponding interior-initialization and Stage-I Gaussian controls are reported in Table 4.

Table 4. Interior initialization and Stage-I Gaussian controls on TransNest.

SettingCD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
No interior reinitialization4.830.8023.147.37
Retain spurious Stage-I interior Gaussians4.600.8123.046.73
Unfreeze remaining Stage-I Gaussians5.000.7543.084.54
Full method (Ours)3.970.8342.624.58

Relative to the full method, the three controls increase CD by 21.66%, 15.87%, and 25.94%, while reducing F1 by 3.84%, 2.64%, and 9.59%, respectively.

Q4. Baselines and shared fields

Both TSGS and TSPE-GS use publicly available latest code. For TSGS, the αSurf, TransLab, and ClearPose results follow the authors' code, use 30k training steps, and use single-depth TSDF extraction. αSurf uses no external normal supervision, while TransLab and ClearPose follow the normals in their released configurations. TSPE-GS uses its official open-source implementation. The αSurf row is the reported result, while the DTU and TransLab rows are our reproductions.

Following your suggestion, we fixed the Gaussian representation trained by TSGS on TransNest Scene 05 and changed only the geometry extraction backend. We compare TSGS extraction, TSPE-GS extraction, and RADE+FUSE extraction. The results are below.

Table 5. Geometry extraction on the same Gaussian representation.

Extraction methodCD (mm) ↓F1 ↑
TSGS extraction7.340.601
TSPE-GS extraction10.940.407
RADE+FUSE (Ours)7.100.622

On the same Gaussian representation, RADE+FUSE achieves the lowest CD and highest F1. This shows that the extraction method directly affects the final reconstruction, and our extraction gives the best result.

Thanks the authors for the response, it resolves my initial concerns, and I would like to raise my score towards acceptance of this work.

Thank you for taking the time to consider our response and for raising your score to 3. We’re glad our response addressed your concerns.

Reviewer_LgJnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper reconstructs transparent scenes, including transparent enclosures and the objects inside them, from multi-view images using Gaussian splatting. It freezes the shell and re-learns the interior during training, and extracts several depths per ray whose meshes are merged by a coverage rule. It reports the lowest mean errors on four transparent benchmarks, including a new synthetic nested dataset developed by the authors.

Critical strengths
  • Strong results on complex transparent scenes. The paper addresses the challenging task of reconstructing scenes containing both transparent and opaque objects. Its reconstructions appear more complete than those of the compared methods (Figure 3), and it achieves the lowest mean errors among the evaluated methods on all four transparent-scene benchmarks (Tables 1 and 2).
Critical weaknesses
  • The paper presents its two-stage strategy, which freezes the outer shell and re-optimizes the interior, as its first contribution. However, reconstructing the outer surface first and then recovering the interior is already the core idea of NU-NeRF (SIGGRAPH Asia 2024); a similar strategy is also used in Opti-NeuS (CVPR 2026). The main distinction therefore lies in the GS implementation. However, the paper provides little evidence that these specific designs account for the improvement. Removing the entire refinement increases the inner reconstruction error from 4.58 to 26.12 mm, whereas retaining the two-stage refinement but disabling the shell contribution cap increases it to only 7.35 mm. The cap thus accounts for only a small share of the gain, and the other GS-specific choices are not ablated separately. Hence, the novelty of this first contribution remains unclear. Clear gains over NU-NeRF and Opti-NeuS may offset this, but the paper does not compare with either of them on nested scenes.

  • The second claimed contribution is multi-surface extraction using RADE and FUSE. However, the evidence that these components drive the reported improvements is weak. (1) They are ablated only on TransLab (Table 4), where the gains are marginal (1.55 mm with RADE vs. 1.60 mm without; 1.56 mm without FUSE). (2) There is no ablation where they should matter most. Multi-surface extraction is motivated by surfaces seen through other surfaces, as in TransNest and parts of αSurf, yet neither benchmark includes an ablation of RADE or FUSE. (3) Other changes may explain the improvements. Compared with TSGS (ACM MM 2025), the training incorporates multi-view losses from PGSR (TVCG 2025) and a stronger normal prior, and neither is ablated. On DTU, the method uses neither RADE nor FUSE yet still improves over TSGS. Hence, the contribution of RADE and FUSE to the reported improvements remains unconvincing to me.

Post-rebuttal update: I read the rebuttal and the additional experiments. They resolve my concerns about the multi-view losses and the normal prior. However, my concerns about the novelty of the two-stage strategy (W1) and the additional benefit of RADE/FUSE over existing extraction approaches (W2) remain (see my official comment for details). Therefore, I keep my rating at 2.

Questions for authors

See above weaknesses.

(W1) Comparison with NU-NeRF on TransNest, using the metrics of Table 2. (W2) Ablation of RADE and FUSE on TransNest, using the metrics of Table 2. (W2) Ablation of the PGSR multi-view losses on TransNest, and TSGS with TransNormal-2 on αSurf, TransLab, and ClearPose.

5 comments

We sincerely thank the reviewer for the time and care invested in this thorough and constructive review, and we are especially grateful for the recognition of the completeness of our reconstructions on transparent scenes. These comments have directly helped us sharpen both the evidence and the presentation, and we address each of them in turn below.

W1.1. Gaussian-specific novelty

Our design targets occlusion competition in Gaussian rasterization. For a ray, front-to-back compositing is C^(r)=∑iTiαici,Ti=∏j<i(1−αj).\hat{\mathbf C}(\mathbf r)=\sum_i T_i\alpha_i\mathbf c_i,\qquad T_i=\prod_{j<i}(1-\alpha_j). When a shell lies in front of the interior with high accumulated opacity, it explains the pixel color, drives the interior transmittance TiT_i toward zero, and triggers early termination. Interior Gaussians then contribute little and receive weak RGB gradients. Freezing the shell therefore does not remove the blockage. We cap the shell contribution and replace conflicting old points with SfM-seeded interior Gaussians that are optimized separately. The outer-then-inner order follows NU-NeRF and Opti-NeuS. These mechanisms are the Gaussian-specific distinction.

W1.2. Effect of the training design

We ablate shell attenuation separately during training and extraction on TransNest (Table 1). At fixed extraction attenuation, removing training attenuation raises mean CD from 3.97 to 4.34 mm (+9.32%) and lowers F1 from 0.834 to 0.819. Extraction attenuation acts in the opposite direction: removing it lowers CD from 3.97 to 3.58 mm but raises inner error from 4.58 to 5.63 mm (+22.93%). Removing both raises inner error to 6.31 mm (+37.77%). Training attenuation mainly improves CD, whereas extraction attenuation protects the interior.

Table 1. Shell attenuation during training and extraction on TransNest.

TrainingExtractionCD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
OnOff3.580.8562.375.63
OffOn4.340.8192.484.23
OffOff3.780.8462.396.31
OnOn (full method)3.970.8342.624.58

Separate ablations also support SfM reinitialization, replacing spurious Stage-I interior Gaussians, and constraining updates to retained Gaussians (Table 2). Relative to the full method, the three controls increase CD by 21.66%, 15.87%, and 25.94%, while reducing F1 by 3.84%, 2.64%, and 9.59%, respectively.

Table 2. Interior initialization and Stage-I Gaussian controls on TransNest.

SettingCD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
No interior reinitialization4.830.8023.147.37
Retain spurious Stage-I interior Gaussians4.600.8123.046.73
Unfreeze remaining Stage-I Gaussians5.000.7543.084.54
Full method (Ours)3.970.8342.624.58

“Retain spurious Stage-I interior Gaussians” uses the same filtered SfM points to initialize new interior Gaussians while retaining the Stage-I interior Gaussians that the conflict rule would normally remove. “Unfreeze remaining Stage-I Gaussians” keeps the SfM points and removal set fixed, but unfreezes all parameters of the other retained Stage-I Gaussians. Both worsen overall geometry.

W1.3. Comparison with NU-NeRF and Opti-NeuS

Thank you for pointing us to NU-NeRF and Opti-NeuS. This helped us clarify the related work and the method differences. We have added the Opti-NeuS citation and clarified the distinction in Sec. 2, Related Work, paragraph 3, marked in blue. Opti-NeuS learns refractive indices and traces refracted rays. We control Gaussian shell contributions and initialize the interior from multi-view SfM evidence. Since no public Opti-NeuS implementation was available, we could not reproduce it. As a reproducible nested-scene comparison, we ran the official NU-NeRF code on TransNest. Our method reduces mean CD by 84.00% compared with NU-NeRF (Table 3). NU-NeRF takes about 2504.5 minutes, whereas ours takes about 30.1 minutes, or 1.2% of the NU-NeRF runtime.

Table 3. Mean reconstruction results on TransNest.

MethodCD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
NU-NeRF24.820.22443.1122.87
Ours (StratGS)3.970.8342.624.58

W2.1. Ablation of RADE and FUSE

We additionally evaluated RADE/FUSE on TransNest (Table 4). Relative to disabling both modules, enabling RADE alone reduces CD by 51.89% and increases F1 by 113.73%. Enabling FUSE alone reduces CD by 8.47% and increases F1 by 20.60%. Enabling both reduces CD by 86.22% and increases F1 by 257.94%.

Table 4. RADE/FUSE ablations on TransNest.

RADEFUSECD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
OffOff28.810.23358.5812.22
OffOn26.370.28153.649.32
OnOff13.860.49824.326.08
OnOn3.970.8342.624.58

We also added RADE/FUSE ablations on αSurf, TransLab, and ClearPose (Table 5). All metrics improve on all three benchmarks: relative to disabling both modules, enabling both reduces CD by 66.01% and increases F1 by 55.15% on αSurf, reduces CD by 26.89% and increases F1 by 5.50% on TransLab, and reduces CD by 30.87% and increases F1 by 9.49% on ClearPose.

Table 5. RADE/FUSE ablations on αSurf, TransLab, and ClearPose.

RADEFUSEαSurf CD (×100) ↓αSurf [email protected] ↑TransLab CD (mm) ↓TransLab F1@5mm ↑ClearPose GT→pred. (mm) ↓ClearPose F1@5mm ↑
OffOff1.530.4662.120.9094.730.569
OffOn0.950.5101.600.9104.690.569
OnOff0.770.5981.560.9253.810.606
OnOn0.520.7231.550.9593.270.623

W2.2. Ablation of the PGSR multi-view loss

We remove the Stage-I multi-view geometry term and retrain on TransNest (Table 6). The term has only a marginal effect: removing it moves CD from 3.97 to 4.07 mm (0.10 mm, 2.5%) and leaves F1 at 0.834. It therefore accounts for about 2% of our 4.53 mm CD margin over TSGS under the same TN2 prior (8.50 → 3.97), so the improvement is not caused by this loss. TSGS already includes the same multi-view losses, which further cancels the term in the comparison. The stronger normal prior is addressed separately in W2.3.

Table 6. Stage-I multi-view geometry loss ablation on TransNest.

SettingCD (mm) ↓F1@5mm ↑Outer (mm) ↓Inner (mm) ↓
Without multi-view term4.070.8342.634.50
With multi-view term (Ours)3.970.8342.624.58

W2.3. Ablation of the normal prior

We replace TSGS's normal input with the stronger TN2 prior and compare it with ours (Table 7). On TransNest, both methods use the same TN2 prior, yet ours achieves lower CD, so the gain goes beyond the normal prior.

Table 7. Normal-prior comparison with TSGS (bidirectional CD ↓, F1 ↑).

DatasetMethodCD ↓F1 ↑
TransNestTSGS + TN28.500.611
Ours (StratGS)3.970.834
αSurfTSGS original prior0.900.774
TSGS + TN21.030.794
Ours (StratGS)0.520.723
ClearPoseTSGS original prior13.440.366
TSGS + TN212.610.378
Ours (StratGS)9.710.623
DTUTSGS + TN20.48–
Ours (StratGS)0.47–
TransLabTSGS original prior1.690.959
TSGS + TN21.660.959
Ours (StratGS)1.550.959

On αSurf, TN2 improves TSGS F1 by 0.020 but increases its CD by 0.13, so it is not an overall gain on that benchmark. DTU evaluates base reconstruction on opaque scenes, so RADE/FUSE are disabled, and under the TN2 configuration the two methods are close.

Taken together, W2.2 and W2.3 rule out the multi-view loss and the normal prior as the source of the gain, and the ablations in Tables 4–5 show that changing only the extraction stage, with the trained model and normal prior fixed, improves results on both the transparent and the nested benchmarks. The improvement is therefore attributable to RADE and FUSE.

Thank you again for asking what the two-stage design contributes beyond the Gaussian backbone. Our rebuttal addresses this through controlled training comparisons and separate ablations of RADE and FUSE. As the discussion period draws to a close, we’d welcome any remaining questions about these comparisons. If our response has addressed your concerns, we hope you’ll revisit your assessment in light of the additional evidence and consider raising your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.