acceptodds
Under review as a conference paper at ICLR 2027

Structure-Guided Capacity Allocation for Sparse-View Indoor Gaussian Splatting

Abstract

Sparse-view indoor reconstruction with 3D Gaussian Splatting remains challenging because limited cross-view supervision leaves weakly textured surfaces and occluded structures poorly constrained. Unrestricted densification can further reinforce unreliable geometry by allocating additional capacity to ambiguous regions. We propose a structure-guided capacity allocation framework that uses explicit Gaussian roles to guide candidate selection and constrain split directions. Starting from scale-consistent multi-view geometry, we initialize ordinary, line, and planar Gaussians, with planar roles further distinguished by texture complexity. For candidate selection, we introduce a type-calibrated score that combines gradient-based fitting demand with rendering contributions normalized within each structural type. The resulting scores prioritize candidates under a shared Top- growth budget. For growth, persistent structural priors restrict line and high-texture planar split displacements to one-dimensional tangents and two-dimensional tangent subspaces, respectively, while low-texture planar Gaussians are excluded from densification. Descendants inherit these structural attributes, retaining the growth constraints across successive densification stages, while geometric regularization stabilizes subsequent optimization. Experiments demonstrate improved perceptual quality and better preservation of indoor structural details in novel-view renderings compared with representative baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Sparse-View Indoor Reconstruction, Structure-Guided Densification, Capacity Allocation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BjJ8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper tackles sparse-view indoor reconstruction with 3D Gaussian Splatting by explicitly guiding both densification decisions and split directions using persistent structural roles assigned at initialization. The method initializes ordinary, line, and planar Gaussians, ranks densification candidates via a type-calibrated score that balances absolute positional gradients with within-type normalized rendering contribution, and constrains split displacements to structure-aligned subspaces. Combined with depth/normal/planarity regularization from a joint multi-view depth prior, the approach yields improved perceptual quality and better structural preservation on ScanNet, Replica, and NRGBD, with favorable training time and model size.

Critical strengths
  • Introduces persistent structural roles that are inherited across densification, decoupling split subspaces from evolving covariance orientations.

  • Proposes type-calibrated candidate selection: absolute screen-space gradient for demand, modulated by within-type percentile-normalized rendering contribution under a shared Top-K budget.

Critical weaknesses
  • The hard inheritance of structural roles risks locking in early misclassifications with no mechanism to retype or relax constraints; this could hinder recovery from initialization errors.

  • Excluding low-texture planar Gaussians from densification may under-represent legitimate small-scale structure on homogeneous planes, depending on the plane-level texture aggregation.

Questions for authors

How robust is the framework to misclassifications of plane-high vs plane-low texture? Do you support retyping or adaptive promotion of plane-low Gaussians when residuals persist on these planes?

3 comments

Q1: Can hard inheritance of structural roles lock in initialization errors?

Thank you for your thoughtful comments. We agree that incorrect structural attributes at initialization could propagate to subsequent generations through attribute inheritance. We clarify that our current framework does not support online structural reclassification or explicit relaxation of inherited constraints. Continuous optimization updates Gaussian centers, covariance parameters, and appearance attributes, while structural types and their stored geometric references remain fixed. For planar Gaussians, the actual split direction can adapt within the inherited tangent plane (Sec. 3.3, Child Construction and Attribute Inheritance). Therefore, continuous optimization alone cannot fully correct an initially misclassified structural role.

To mitigate this risk, we adopt conservative filtering and geometric consistency checks before assigning structural attributes. Specifically, unreliable line segments are filtered based on detection confidence and length (Sec. 3.1.1, Eq. (3); Appendix A.5). For planar priors, we first obtain per-view planar masks and semantic labels using LP3, and then perform cross-view plane association using PCA-based plane fitting and geometric consistency checks. These checks consider plane-normal differences, plane offsets, and spatial overlap or centroid proximity (Appendix A.4). Furthermore, the high-/low-texture classification employs robust per-view statistics, including percentile clipping and minimum valid-pixel requirements, followed by pixel-count-weighted multi-view voting to reduce sensitivity to individual-view errors (Appendix A.4, Eqs. (15)–(16)).

More importantly, we conduct additional perturbation experiments to evaluate the sensitivity of subsequent structural growth to imperfect initialization. Please refer to Rebuttal Table R3 in our response to Reviewer eyrC, Q4, which reports sensitivity to 20% high-/low-texture label flipping and 10°10° line-direction perturbations on Replica.

Q2: Can excluding low-texture planes from densification miss legitimate fine-scale structure?

We thank the reviewer for raising this important concern. We agree that low texture does not necessarily imply the absence of fine-scale geometric structures. The policy is a capacity-allocation choice under sparse supervision: unrestricted growth on homogeneous surfaces may fit ambiguous photometric evidence without reliably recovering additional structure.

Plane-low regions receive an initial Gaussian representation whose parameters remain trainable. Excluding these Gaussians from cloning and splitting restricts additional capacity, while their existing geometry and appearance can still adjust. This flexibility cannot replace additional primitives when the initial support is insufficient.

Texture classification combines robust per-view gradient statistics, percentile clipping, valid-pixel requirements, and pixel-count-weighted multi-view voting (Appendix A.4). In addition, detected lines take precedence over planar sampling (Sec. 3.1.1), so recognized narrow structures can receive line Gaussians independently of the surrounding plane's texture role.

These measures do not guarantee preservation of every local feature. Plane-level voting may overlook small same-color details, and line initialization cannot protect undetected structures. We therefore acknowledge possible under-representation of local geometry on low-texture planes. The perturbation test measures aggregate sensitivity, rather than directly validating preservation of those fine structures.

Q3: How robust are texture roles, and is residual-triggered promotion supported?

We thank the reviewer for raising this important concern. We agree that low texture does not necessarily imply the absence of fine-scale geometric structures, and incorrectly classifying a high-texture plane as Plane-low could restrict its capacity to represent local details. Residual-triggered retyping or promotion is not supported in the current method. A Plane-low Gaussian remains outside the densification candidate pool even when residuals persist. Its continuous parameters can change, but its structural role and growth eligibility do not. Under the tested 20% texture-label flipping, PSNR decreases from 33.50 to 33.41 dB, while SSIM remains 0.958 (Rebuttal Table R3 in our response to Reviewer eyrC, Q4). This indicates limited aggregate sensitivity in the tested setting, rather than correction of misclassified roles or guaranteed preservation of local details.

Thank you for the careful rebuttal for structural roles, excluding low-texture plane, and its robustness. My concerns have been resolved. I maintain my score.

Thank you for your careful review and for confirming that your concerns have been resolved. We sincerely appreciate the time and attention you devoted to reviewing our work and the additional evidence.

Best regards,

The Authors

Reviewer_o4ptOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a structure-guided capacity allocation framework for sparse-view indoor Gaussian splatting. Starting from multi-view depth and structural priors, it initializes ordinary, line, and texture-dependent planar Gaussians. A type-calibrated score combines absolute screen-space gradients with rendering contributions normalized within each structural type to select densification candidates under a shared Top-K budget. Persistent line and plane attributes constrain split displacements and are inherited by descendants, while geometric regularization stabilizes subsequent optimization. Experiments on ScanNet, Replica, and Neural RGB-D report improved novel-view rendering quality, supported by common-depth-prior comparisons, component ablations, and evaluations with varying numbers of input views.

Critical strengths
  1. The evaluation includes meaningful controls for the depth prior. The authors evaluate D²GS, TWINGS, and VGA-GS variants using the same cached DA3 predictions. The proposed method still outperforms VGA-GS+DA3 across all three datasets, indicating that replacing the depth source alone does not explain the complete framework’s gains. This strengthens the empirical evidence beyond comparisons using different depth predictors.
  2. The framework demonstrates a useful quality–capacity trade-off. On ScanNet, adding the two capacity-allocation components to structure-aware initialization and geometric regularization improves PSNR from 21.68 to 21.92 dB and LPIPS from 0.310 to 0.287, while reducing the Gaussian count from 620.2K to 172.5K. Replica component ablations and experiments with different input-view counts provide additional evidence that the overall design is useful beyond a single setting.
Critical weaknesses
  1. The incremental contribution over closely related structural densification methods is not sufficiently established. IndoorGS already combines structural Gaussian types with type-specific density control and tangent-space densification, but is absent from the quantitative evaluation. Although the paper explains its differences in persistent directional priors and shared candidate ranking, comparisons with other sparse-view methods do not establish the significance of these particular changes. Given the substantial conceptual overlap, a controlled comparison with IndoorGS or its growth strategy is important to assessing the methodological contribution.
  2. The ablations support the complete configuration more strongly than the specific mechanisms emphasized as novel. Table 6 appropriately compares against gradient-only ranking under a fixed Top-K budget, but jointly introduces contribution weighting and within-type calibration. It therefore does not establish whether type-specific calibration is preferable to a simpler global normalization. Likewise, SSG combines persistent directional constraints, attribute inheritance, and suppression of low-texture planar densification without separating their effects. Consequently, it remains unclear how much of the benefit requires the proposed calibration and inherited subspaces, rather than simpler capacity control using the same structural priors. The relatively small gains over gradient-only ranking further make this distinction consequential.
Questions for authors
  1. Clarification on related-method comparison: Why is IndoorGS absent from the quantitative evaluation? Are there methodological or implementation barriers to comparing its structural density-control strategy under the same sparse-view protocol?
  2. Targeted validation of type calibration: With the contribution-modulation parameter and Top-K budget fixed, how does within-type percentile normalization compare with a single global percentile normalization? This comparison would help determine whether the type-specific calibration itself contributes beyond rendering-contribution weighting.
  3. Clarification and targeted validation of structural growth: What growth eligibility rules and split directions are used when SSG is disabled in Tables 3 and 7? In particular, can the benefit of persistent structural directions be isolated while retaining the same low-texture densification policy and growth budget?
5 comments

Q1: Is the incremental contribution over IndoorGS sufficiently established?

We thank the reviewer for raising this concern. We agree that IndoorGS already uses structural Gaussian types and tangent-space densification. Our contribution lies in a different formulation of capacity allocation and persistent structural growth, which addresses two complementary questions: which Gaussians should receive additional capacity, and within which structural subspaces should they grow?

(1) Comparison with IndoorGS. The original submission did not include IndoorGS because its official implementation was unavailable, rather than because of protocol incompatibility or method inapplicability. We have now independently reimplemented IndoorGS and evaluated it using the same input views, train/test splits, image resolutions, and evaluation metrics as our method. We also evaluate a variant using the shared DA3 predictions. The results are explicitly labelled as our reimplementation; methodological details are discussed in our responses to Reviewer eyrC, Q2–Q3. We have uploaded our code to the anonymous repository.

Rebuttal Table R5. Quantitative comparison with our IndoorGS reimplementation on ScanNet, Replica, and NRGBD.

MethodScanNetReplicaNRGBD
IndoorGS (Re-impl.)20.979/0.732/0.38431.560/0.942/0.16430.608/0.943/0.160
IndoorGS(Re-impl.)+DA320.978/0.721/0.39731.129/0.936/0.17429.831/0.933/0.174
Ours21.920/0.786/0.28733.500/0.958/0.07832.180/0.955/0.074

(2) Mechanism differences. IndoorGS adopts geometry-dependent growth policies, including tangent-space densification for line Gaussians and priority-based treatment of different geometric cues. In contrast, our TACS instead ranks all eligible types under one Top-K budget using gradients modulated by type-calibrated rendering contribution. SSG uses stored 1D line tangents and 2D planar subspaces, independently of covariance orientation. It also permits eligible high-texture planes to refine while excluding low-texture planes, whereas IndoorGS suppresses plane-like densification.

Q2: Does Table 6 isolate type calibration from contribution weighting?

Thank you for pointing out this important issue. We agree that the original Table 6 jointly introduces contribution modulation and type-specific calibration, making it difficult to attribute their respective effects. It compares gradient-only ranking (λ = 1) with the full contribution-modulated score (λ = 0.5), including calibration. Its PSNR gains of 0.13 dB on ScanNet and 0.061 dB on Replica therefore support the combined ranking design.

To isolate the effect of contribution calibration, we compare three strategies under the same (λ=0.5), Top-K budget, initialization, and training configuration: (1) Raw Contribution, which directly uses the original rendering contribution without Q20–Q80 normalization or clipping; (2) Global Calibration, which applies shared Q20–Q80 normalization across all Gaussian types; and (3) Type-specific Calibration, which performs normalization independently within each structural type. The results are reported in the Rebuttal Table R6.

Rebuttal Table R6. Contribution-calibration ablation on Replica.

Contribution treatmentPSNR ↑SSIM ↑LPIPS ↓Depth RMSE ↓Gaussians
Raw contribution33.460.9570.0770.072204.8K
Global calibration33.480.9580.0780.068207.7K
Type-specific Calibration33.500.9580.0780.067208.1K

The proposed type-specific Calibration improves PSNR by 0.02 dB and depth RMSE from 0.068 to 0.067, with identical SSIM and LPIPS at the reported precision. Raw contributions have slightly better LPIPS. We therefore claim a small numerical PSNR/depth benefit, rather than uniform superiority or attributing the full pipeline improvement to type calibration.

Q3: Does the SSG ablation separate its constituent mechanisms?

The ablations in original Tables 3 and 7 are component comparisons: disabling SSG allows line and high-texture planar Gaussians to clone as well as split, and removes their structural splitting constraints. They therefore change both permitted growth operations and the splitting rule.

The new Rebuttal Table R7 controls for this confound. Both variants share initialization, the TACS selection rule, the per-step Top-K budget, and the metadata-inheritance policy. In both, ordinary Gaussians can clone or split; line/high-texture planar Gaussians are split-only; and low-texture planes are excluded before Top-K selection. The controlled change is vanilla versus structural splitting.

Rebuttal Table R7. Eligibility-matched splitting comparison on Replica.

Splitting rulePSNR ↑SSIM ↑LPIPS ↓Depth RMSE ↓Gaussians
Vanilla splitting33.140.9560.0780.078267.5K
SSG33.500.9580.0780.067208.1K

As shown in Rebuttal Table R7, SSG improves PSNR by 0.36 dB, reduces depth RMSE by 14.1%, and uses 22.2% fewer Gaussians, with LPIPS unchanged. These gains go beyond changing cloning eligibility or excluding low-texture planes. These results demonstrate that persistent structure-constrained splitting improves reconstruction quality and geometric accuracy while reducing Gaussian count under matched growth eligibility and capacity budgets. We will clarify the distinction between the original SSG ablation and this additional controlled experiment in the revised manuscript.

Q4: Do calibration and inherited subspaces improve reconstruction beyond simpler capacity control.

We thank the reviewer for raising this important concern. The original Table 7 shows that TACS does not simply suppress capacity: adding it increases the Gaussian count while improving appearance quality. SSG alone produces the fewest Gaussians and lowest depth RMSE, but worse LPIPS. The full configuration uses an intermediate count and achieves the best PSNR and SSIM in this ablation.

This establishes a quality–size tradeoff, rather than a general advantage from minimizing primitive count. The more direct mechanism control is Rebuttal Table R7: under the same structural initialization, selection rule, eligibility, and per-step budget, SSG improves both rendering quality and depth accuracy while ending with fewer Gaussians. The final count remains a training outcome; this is not a fixed-final-count comparison.

Rebuttal Table R6 shows that global normalization is competitive, so we do not claim type calibration is indispensable. The stronger controlled evidence concerns the structural splitting operator, rather than a large benefit from normalization alone.

Q5: Why was IndoorGS absent, and are there barriers to a fair comparison?

We thank the reviewer for pointing out the missing comparison with IndoorGS. Its original omission was primarily due to unavailable official code at the time of our experiments. There is no fundamental barrier to evaluating IndoorGS under our sparse-view protocol, although its original evaluation uses a conventional multi-view setting.

We have now added our reimplementation and its shared-DA3 variant in Rebuttal Table R5. We match input views, train/test splits, resolutions, and metrics. This matches the data and evaluation protocol without claiming identical method-specific preprocessing or optimization settings.

The comparison shows an advantage over the evaluated reimplementation, including after depth-prior replacement. Please refer to our response to Reviewer eyrC, Q2–Q3 for more details.

Q6: With λ and K fixed, how does within-type normalization compare with global normalization?

Thank you for this suggestion. We added the controlled comparison in Rebuttal Table R6. With λ = 0.5, K, and the remaining settings fixed, only the grouping used to compute contribution percentiles changes.

Global calibration achieves 33.48 dB PSNR and 0.068 depth RMSE; within-type calibration achieves 33.50 dB and 0.067. Both obtain 0.958 SSIM and 0.078 LPIPS. Thus, the observed advantage is small and numerical. Within-type calibration accounts for relative contribution within each structural role, while global calibration remains a competitive simpler alternative. Please refer to Q2 and Rebuttal Table R6.

Q7: What eligibility rules and split directions are used when SSG is disabled?

For the original SSG-disabled configuration in Tables 3 and 7:

  • Ordinary, line, and Plane-high Gaussians can use standard cloning or splitting, subject to the selection and size criteria.
  • Splitting samples offsets from the current Gaussian covariance, without projection onto a stored line tangent or planar subspace.
  • Plane-low Gaussians remain excluded before Top-K selection and consume no growth budget.

Full SSG makes eligible line and Plane-high Gaussians split-only and applies their 1D tangent / 2D planar-subspace constraints. Ordinary Gaussians retain cloning/splitting. The planar reference normal persists, while the split direction can adapt within its tangent plane. K is unchanged in the paired TACS-enabled comparison.

The original ablation therefore changes both growth operations and structural splitting. Rebuttal Table R7 supplies the separate control in which both variants retain full SSG's eligibility policy.

Q8: Can structural directions be evaluated with the same low-texture policy and growth budget?

We thank the reviewer for this valuable suggestion. To test the contribution of persistent structural directions while retaining the same low-texture densification policy and growth budget, we conduct an additional controlled comparison between vanilla 3DGS splitting and our structure-guided splitting strategy. We added the controlled comparison in Rebuttal Table R7, keeping initialization, the TACS selection rule, the per-step Top-K budget, growth eligibility, and the metadata-inheritance policy fixed. In both variants, ordinary Gaussians can clone or split, line and Plane-high Gaussians are split-only, and Plane-low Gaussians are excluded before Top-K selection.

The controlled factor is the splitting operator: vanilla 3DGS splitting versus SSG using persistent structural references, under matched growth eligibility and per-step budgets. As defined in Sec. 3.3 and Eq. (10), SSG uses Pi=t~it~i⊤P_i=\tilde t_i\tilde t_i^\top for lines and Pi=I−nini⊤P_i=I-n_i n_i^\top for Plane-high Gaussians, where t~i\tilde t_i and nin_i are the stored line tangent and target plane normal. The split direction is di=vmax⁡(PiΣiPi)d_i=v_{\max}(P_i\Sigma_iP_i). Thus, the structural reference fixes the admissible subspace, while the current covariance determines the direction within it. The vanilla baseline retains the structural metadata but does not use these constraints when splitting.

On Replica, SSG improves PSNR from 33.14 to 33.50 dB, reduces depth RMSE from 0.078 to 0.067, and decreases the Gaussian count from 267.5K to 208.1K, with LPIPS unchanged at 0.078. These correspond to +0.36 dB, approximately 14.1% lower depth RMSE, and 22.2% fewer Gaussians. The gain therefore goes beyond the fixed low-texture exclusion and per-step budget policies. This experiment evaluates persistent structure-constrained splitting as a whole.

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

This paper tackles the problem of sparse-view gaussian splatting by examining the current densification rules and tying them more closely to the "type" of gaussian. Gaussians are categorized into different types based on the region of the scene they represent, and then are densified according to special rules based on the type (e.g. "line" gaussians are densified along the line only). The densification process itself is also enhanced by using not just the gradients but also the rendering contribution as a signal. The technique is evaluated on indoor scenes and shows improvement in PSNR/SSIM/LPIPs over prior sparse-view work.

Critical strengths
  1. The results appear strong on the 3 benchmarks tested against the compared baselines
  2. The problem being addressed is relevant and important
Critical weaknesses
  1. I am unsure why this method was evaluated only on Indoor scenes, but also was not compared against any prior work mentioned in the Related Works section which is tailored to this setting. All of the baseline comparisons are related to sparse-view techniques which while relevant might underperform compared to indoor-specific methods.
  2. IndoorGS (Ruan et al. 2025) is described as including "type-specific initialization and density control". I would like to see a more clear comparison of how the proposed method is different from these techniques and would like to see the proposed method compared against this work.
  3. The technique relies heavily on prior models for depth estimation, line detection, and plane detection. I would like to see some evaluation of how robust the technique is to noise/inaccuracies in these inputs.
Questions for authors
  1. Why was the method evaluated on indoor scenes only? It seems the techniques would apply to many outdoor scenes as well (for e.g. street scenes have many planes/lines/etc...)
  2. Does the training time include all the data pre-processing and estimation of depth/lines/etc..?
5 comments

Q1: Why was evaluation initially limited to indoor scenes?

Thank you for raising this question. Our initial evaluation focused on sparse-view indoor reconstruction, where large weakly textured surfaces, occlusions, and thin structures make reliable densification particularly challenging. These conditions directly motivate our structure-guided capacity allocation: preserving structural details while limiting unnecessary growth in poorly constrained regions. The underlying growth rules are defined by line and plane geometry and are therefore not inherently restricted to indoor scenes.

To evaluate applicability beyond this initial scope, we added experiments on three outdoor ETH3D scenes: Courtyard, Terrace, and Electro. Our method achieves the highest PSNR (20.93 dB) and the second-best SSIM (0.807) and LPIPS (0.248) among the compared methods. We have also added qualitative comparisons to the Rebuttal Revision (Fig. 3), which provides complementary visual evidence of the reconstruction quality in outdoor scenes. Together, these results support applicability beyond indoor reconstruction to the evaluated structured outdoor environments.

** Rebuttal Table R1. Quantitative comparison on ETH3D.**

MethodPSNR↑SSIM↑LPIPS↓
DropGaussian20.440.8010.260
D$^2$GS20.510.7970.264
TWINGS20.750.8150.245
StereoGS20.770.7970.303
VGA-GS20.290.7620.303
IndoorGS (Re-impl.)20.590.7830.309
Ours20.930.8070.248

Q2: Are indoor-specific methods adequately represented in the comparisons?

The original comparison already includes P²NeRFP^²NeRF and VGA-GS, which target few-shot or sparse-view indoor reconstruction. Nevertheless, we agree that IndoorGS is a particularly relevant comparison because it also incorporates structural cues and type-dependent density control. IndoorGS was omitted primarily because official code was unavailable. Its original evaluation also follows a conventional multi-view setting.

To address this gap, we independently reimplemented IndoorGS following its paper and supplementary material. We evaluated this implementation using the same input views, image resolutions, and evaluation metrics as our method. We also evaluated a variant using the shared DA3 depth predictions.

As shown in Rebuttal Table R2, our method improves PSNR over our IndoorGS reimplementation by 0.94/1.94/1.57dB on ScanNet/Replica/NRGBD, with higher SSIM and lower LPIPS throughout. The shared-DA3 variant does not close the gap, suggesting that the depth source alone does not explain the observed advantage.

Rebuttal Table R2. Quantitative comparison with our IndoorGS reimplementation.

MethodScanNetReplicaNRGBD
IndoorGS (Re-impl.)20.979/0.732/0.38431.560/0.942/0.16430.608/0.943/0.160
IndoorGS (Re-impl.)+DA320.978/0.721/0.39731.129/0.936/0.17429.831/0.933/0.174
Ours21.920/0.786/0.28733.500/0.958/0.07832.180/0.955/0.074

Q3: How does the method differ from IndoorGS, and how does it compare quantitatively?

Thank you for raising this important question. We agree that IndoorGS is a closely related work, as both methods exploit structural cues to initialize Gaussian primitives and employ type-dependent density control. Our main differences concern how growth capacity is allocated and how structural Gaussians expand. We have uploaded our code to the anonymous repository.

  1. Shared-budget selection. IndoorGS uses cue-dependent growth policies, including prioritizing line-derived over SfM-derived Gaussians. Our TACS combines absolute screen-space gradients with rendering contributions calibrated within each type. All eligible candidates compete under one Top-K budget, explicitly allocating growth capacity across types without type-specific quotas.

  2. Persistent geometric references. IndoorGS densifies line-derived Gaussians within a 2D tangent space determined by their learned covariance. Our SSG instead uses stored line tangents to constrain line growth to 1D, and target plane normals to define 2D admissible subspaces for high-texture planar growth. These geometric references persist across generations independently of evolving covariance orientations, keeping the growth constraints anchored to the initialized structure.

  3. Texture-dependent planar refinement. IndoorGS disables densification for plane-like Gaussians. We distinguish low- and high-texture planes: low-texture planes remain excluded from growth, while eligible high-texture planes can split within their tangent planes. This permits refinement of textured planar details while limiting redundant growth on weakly constrained homogeneous surfaces.

Q4. How robust is the method to inaccurate depth, line, and plane priors?

Thank you for raising this question. To evaluate the sensitivity of our framework to prior inaccuracies, we conduct additional experiments covering depth-prior variations and structural-prior perturbations.

Depth priors. The original Table 4 compares DA3 and MVSAnywhere, changing only the depth predictor used for initialization and geometric regularization while keeping the remaining settings fixed. On ScanNet, the predictors produce similar results: PSNR differs by only 0.06 dB, while MVSAnywhere achieves slightly better SSIM and LPIPS. On Replica, DA3 performs better, with a 0.40 dB PSNR advantage. These results demonstrate compatibility with both tested predictors, although reconstruction quality remains sensitive to the depth source.

Structural priors. Line filtering, geometric consistency checks, and multi-view aggregation reduce unreliable assignments before training. We additionally test two separate perturbations on Replica: 20% high-/low-texture planar-label flips and 10° line-direction perturbations.

Rebuttal Table R3. Sensitivity to structural-prior perturbations on Replica.

SettingPSNR↑SSIM↑LPIPS↓Gaussians
No perturbation33.500.9580.0780208.1K
Flipping (20%)33.410.9580.0754215.0K
Perturbation (10°)33.440.9570.0810205.1K

PSNR decreases by only 0.09 and 0.06 dB, respectively, and SSIM changes by at most 0.001. These results support tolerance to the tested structural errors. The planar test specifically concerns texture-role misclassification and does not establish robustness to arbitrary plane-detection or normal-estimation errors.

Q5: Does the training time include preprocessing and depth/line/plane estimation.

Thank you for pointing out this clarification. Our reported time in Table 2 measures Gaussian optimization and excludes depth, line, and plane prior extraction, as well as model loading. We now report these additional components separately in Rebuttal Table R4. Measurements were conducted on a single NVIDIA RTX 3090 across eight Replica scenes, each with 20 training views. Acquisition times are measured per scene after model warm-up and include image loading, inference, filtering, and cache writing. Model loading is measured separately as a one-time cost per extractor.

Rebuttal Table R4. Preprocessing costs on Replica. (Acquisition times are reported as mean ± standard deviation across scenes).

PriorModel loading (s)Acquisition (s/scene)
DA317.492.35 ± 0.11
ScaleLSD2.561.24±0.10
LP310.2219.53±1.06
Total30.2723.12

The measured prior-acquisition stages require 23.12 s per scene after warm-up, with an additional 30.27 s for loading all three models. Adding acquisition to the 4.08 min optimization time on Replica gives an estimated subtotal of 4.47 min per scene with preloaded models. Charging all reported model-loading costs to one scene increases this subtotal to 4.97 min.

Thank you for your response. The addition of IndoorGS helps address my concerns about indoor specific comparison, but there are two points I am unsure about:

  • I am surprised that adding DA3 hurts the performance of IndoorGS. How was this integrated into the code?
  • It seems on the outdoor set, TWINGS outperforms on 2/3 metrics and these results are only on 3 scenes which is a limited set to evaluate from

For the depth priors, I would've liked to see an evaluation of depth with noise added rather than simply testing multiple different depth models to get a better understanding of how the technique handles noisy/inaccurate depth.

The inclusion of the model loading/forward pass also changes the ranking of timings as well and has a significant effect. I will maintain my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Reviewers,

We sincerely thank all reviewers for their careful reading and constructive feedback. Your comments have helped us clarify the methodological contributions, strengthen the experimental evidence, and improve the presentation of our work. We have uploaded a revised manuscript, with revised and newly added text highlighted in blue, and provide detailed responses to each reviewer in the corresponding discussion thread.

Our code is available in the anonymous repository. The table and figure numbers below refer to the revised manuscript.

The main additions are:

  • Independent implementation and evaluation of IndoorGS. As official code was unavailable at the time of our experiments, we independently reimplemented IndoorGS based on its paper and supplementary material. We added comparisons on ScanNet, Replica, and NRGBD under the same sparse-view data and evaluation protocols, including a variant using the shared DA3 depth predictions (Table 1). We also included IndoorGS in the ETH3D evaluation (Table 4), experiments with varying numbers of input views (Table 13), and qualitative rendering and mesh comparisons.

  • Evaluation beyond indoor scenes. We added quantitative and qualitative results on three outdoor ETH3D scenes to examine applicability to structured outdoor environments (Table 4 and Fig. 3).

  • We added comparisons of raw contributions, global percentile calibration, and type-specific calibration under the same scoring formulation, contribution-modulation parameter, and growth budget (Table 11). The analysis distinguishes the small numerical benefit of type-specific calibration from the gains of the complete framework.

  • We compared vanilla splitting with SSG while retaining the same growth eligibility, low-texture exclusion policy, selection rule, and per-step budget (Table 12 and Fig. 7). This evaluates the structural splitting operator beyond changes in growth eligibility.

  • Sensitivity to structural-prior errors. We added separate experiments with 20% high-/low-texture planar-label flipping and 10° line-direction perturbations on Replica, together with a discussion of the scope of these robustness results (Table 5).

  • Preprocessing cost analysis. We reported depth, line, and planar-prior acquisition times and model-loading costs separately from Gaussian optimization time, providing a more complete account of computational overhead (Table 10).

We appreciate the reviewers’ suggestions and hope that these additions and clarifications address the main concerns.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.