acceptodds
Under review as a conference paper at ICLR 2027

ReliaSplat: Reliable Features in Language Gaussian Splatting

Abstract

Open-vocabulary 3D scene understanding commonly builds language fields by distilling multi-view 2D language features into 3D representations. In Language Gaussian Splatting, these pre-extracted features serve as the basic semantic observations for constructing Gaussian language fields. However, their accuracy is often affected by occlusions, scale variations, region-partition errors, and noise from vision-language models. As a result, the same 3D entity may receive inconsistent or incorrect semantic observations across views. Existing methods mainly focus on how language features are represented or aggregated in 3D, while the reliability of the input features themselves receives less attention. ReliaSplat is a training-free framework for improving input language features. Cross-View Reliability-Aware Aggregation estimates observation reliability from cross-view semantic consistency and suppresses inconsistent features during aggregation. Consensus-Guided Feature Refinement uses the resulting 3D semantic consensus to correct noisy 2D features through a lightweight 3D2D3D process. Geometry-Aligned Semantic Propagation further supplements reliable semantic observations. Experiments on LERF and ScanNet show consistent improvements in open-vocabulary semantic segmentation, with an average mIoU of 64.3% on LERF. Cross-framework experiments further demonstrate that the enhanced language features can directly improve multiple Language Gaussian Splatting methods, validating the generality of the feature enhancement framework.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Open-vocabulary 3D Scene Understanding, Language Gaussian Splatting, 3D Gaussian Splatting, Language Feature Enhancement, Multi-view Feature Aggregation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UfamOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ReliaSplat, a training-free feature enhancement framework designed to tackle input feature unreliability in Language Gaussian Splatting. The authors observe that while existing works emphasize 3D representation, compression, and rendering aggregation, they largely overlook inherent errors in upstream 2D VLM features caused by occlusions, scale shifts, and region-partitioning noise. ReliaSplat addresses this at the input level via three integrated mechanisms: (1) Cross-View Reliability-Aware Aggregation, which models observation reliability via multi-view semantic consistency to penalize outlier features; (2) Consensus-Guided Feature Refinement, which utilizes the aggregated 3D consensus to correct noisy 2D features through a 3D -> 2D -> 3D projection cycle; and (3) Geometry-Aligned Semantic Propagation, which densifies semantic coverage using pseudo-camera viewpoints under geometric consistency constraints. Evaluated on LERF and ScanNet, the framework demonstrates plug-and-play generalizability by replacing input features across four downstream baselines (LangSplat, LangSplatV2, VALA, and THGS), yielding consistent mIoU gains.

Critical strengths

S1. Well-motivated perspective shifting from 3D representation engineering to input feature reliability: The vast majority of existing 3D language Gaussian splatting pipelines treat pre-extracted 2D VLM features as ground-truth semantic observations, focusing their innovations almost entirely on downstream distillation, codebook quantization, or feature field rendering. ReliaSplat directly targets the upstream bottleneck ("garbage in, garbage out"). Crucially, the authors distinguish between geometric/rendering significance and semantic consistency, correctly arguing that high rendering contribution does not imply semantic correctness. Formulating observation reliability via multi-view cross-view agreement provides an intuitive, theoretically grounded, and training-free inductive bias for 3D language fields.

S2. Genuine plug-and-play applicability supported by extensive cross-framework validation: Rather than coupling the denoising mechanism to a custom 3D architecture, the framework is designed to output cleaned, standardized 2D language features that can be fed directly into existing pipelines. The authors substantiate this generality through rigorous experiments across four major downstream architectures (LangSplat, LangSplatV2, VALA, and THGS) under their respective official training scripts and benchmark protocols on LERF and ScanNet. Achieving consistent mIoU improvements across all four frameworks confirms that the feature enhancements provide practical, architecture-agnostic utility to the community.

Critical weaknesses

W1. Confounding data augmentation via pseudo-camera pose generation undermines fair baseline comparison: In Section 4.4 and Appendix A.3, the Geometry-Aligned Semantic Propagation module synthesizes pseudo-camera poses to propagate features and supplement valid semantic observations. However, standard downstream baselines (LangSplat, THGS, VALA) operate strictly on the original training viewpoints. Synthesizing novel pseudo-view observations effectively performs multi-view semantic data augmentation. Consequently, the performance gains reported in Table 1 and Section 5.4 are confounded: it is currently impossible to isolate whether the downstream improvements stem from the core reliability modeling (reliability-aware aggregation and consensus refinement) or simply from training on a denser, augmented set of viewpoints. The paper lacks a strictly controlled baseline where downstream methods either receive the same pseudo-camera dense views, or where ReliaSplat is evaluated strictly without pseudo-camera interpolation.

W2. Metric divergence between mIoU gains and stagnant mAcc raises questions regarding the nature of semantic correction: In the cross-framework evaluation, the paper notes that ReliaSplat consistently boosts mean mIoU while keeping mean mAcc "stable overall." If the framework genuinely corrects noisy, misclassified VLM observations back to their true semantic categories, categorical accuracy (mAcc) should experience a concurrent increase. The stagnation of mAcc suggests that the 3D -> 2D -> 3D filtering and propagation cycles may primarily act as a spatial low-pass filter that smooths object boundary masks and suppresses isolated false-positive outliers (which reduces the union in IoU and inflates mIoU), rather than actively repairing true semantic misclassifications on challenging or long-tail classes.

W3. Omission of wall-clock runtime and memory profiling for a framework claiming to be lightweight and training-free: While ReliaSplat is highlighted as "training-free," training-free does not automatically imply computationally inexpensive. The pipeline requires dense multi-view feature unprojection, pairwise cross-view cosine similarity calculations across dense SAM masks, iterative 3D -> 2D rasterization rendering, and novel-view feature propagation across multiple pseudo-cameras. Depending on scene resolution and viewpoint count, these offline pre-processing steps could incur massive CPU/GPU memory footprints and substantial wall-clock overhead. Without quantitative reporting of pre-processing latency versus downstream distillation training time, the paper's claims regarding "lightweight" and practical efficiency remain unsubstantiated.

Questions for authors

Q1. Disentangling Pseudo-Camera Data Augmentation: Could the authors provide a decoupled ablation study on the downstream frameworks (e.g., on LangSplat and THGS) evaluating: (a) ReliaSplat operating strictly on original training camera poses without pseudo-camera generation, and (b) baseline methods trained with naive pseudo-camera interpolated feature maps? How much of the observed mIoU gain is uniquely attributable to reliability-aware filtering versus viewpoint densification?

Q2. Detailed Breakdown of mIoU vs. mAcc Divergence: What is the underlying cause for the improvement in mean mIoU while mean mAcc remains flat across all four downstream methods? Could you provide per-class confusion matrices or accuracy delta breakdowns on ScanNet / LERF to clarify whether the performance delta is driven by boundary sharpening/denoising or genuine semantic label corrections?

Q3. Quantitative Profiling of Computational Overhead: What is the exact wall-clock pre-processing runtime and peak GPU VRAM consumption for each stage of ReliaSplat (Aggregation, Refinement, Propagation) on standard benchmark scenes (e.g., LERF figurines and ScanNet scenes) compared to the standard feature extraction and downstream training time of LangSplat?

Q4. Robustness to Systemic Multi-View Hallucinations: How does the cross-view consistency metric behave when an underlying VLM produces a shared systematic error across a majority of viewpoints (e.g., 4 out of 5 viewpoints incorrectly classifying an ambiguous object as class A, and only 1 close-up view correctly identifying it as class B)? In such cases, does the consensus mechanism suppress the single accurate observation as an outlier?

2 comments

We thank the reviewer for the careful review and constructive suggestions. These comments have helped us strengthen our experimental analysis and improve the clarity of our method description and presentation of results. We address your questions point by point below.

Revisions and additions are highlighted in blue in the updated PDF for ease of review.

W1: Data augmentation through pseudo-camera poses confounds the results and compromises fair baseline comparisons

  1. Semantic propagation only constructs ReliaSplat's 3D semantic field. Table 6 reports its modest cost.

  2. Table 1 evaluates the performance of the 3D semantic field constructed by ReliaSplat itself and provides a fair comparison with other methods.

  3. Tables 4(C) and 13 isolate propagation's effect, showing mean mIoU gains when added to R2 on LERF and ScanNet.

  4. Table 5 shows downstream gains from replacing only real-view inputs with R1 features, without view densification.

  5. Table 9 shows sparse outputs (31.00% valid pixels at R2 versus 90.94% in source maps). Given the training costs and benefits for downstream methods, we aggregate them online into our own field without saving dense pseudo-view maps.

Q1: Disentangling Pseudo-Camera Data Augmentation

  1. For (a), Table 5 shows that refining real-view features alone improves downstream performance without view densification.

  2. Tables 4(A) and 3 show that consistency weighting and two-round feature refinement yield mean mIoU gains of 1.15 and 1.28 percentage points, respectively, without adding views, demonstrating the effectiveness of both components.

Q2: Detailed Breakdown of mIoU vs mAcc Divergence

  1. Section 5.1 of the revised manuscript explains that LERF-OVS mAcc is mean localization accuracy, which checks whether the pixel with the highest CLIP response falls inside the ground-truth bounding box; mIoU measures overlap between predicted and ground-truth masks. Thus, mIoU can improve while mAcc remains unchanged when masks improve without changing the localization decision.

  2. Figure 2 illustrates improved localization, with the localization points after refinement falling inside the ground-truth bounding boxes in all examples shown. Figure 3 illustrates improved masks, consistent with the mIoU gains reported in Table 5.

  3. LERF-OVS uses open-vocabulary text queries without a unified space of mutually exclusive class labels, so conventional multiclass confusion matrices are not directly applicable. We add this analysis on ScanNet, which has fixed class labels.

  4. Table 11 shows gains of 2.05, 1.58, and 1.72 percentage points in ScanNet 10-class mIoU, mAcc, and OA, respectively, with more corrected errors (23,377) than new errors (9,103), indicating that feature refinement improves semantic class predictions.

  5. Table 12 and Figures 7 and 8 further report per-class accuracy, confusion matrices, and their differences, including degraded classes. The existing results support the presence of semantic class correction.

Q3: Quantitative Profiling of Computational Overhead

  1. Table 6 reports detailed runtime and peak memory for each stage across the four LERF scenes.

  2. For figurines, the runtime and peak GPU memory for the four stages are:

    (1) Aggregation (including qiq_i precomputation): 355.85 s, 12.67 GiB.

    (2) R1 (cumulative from the Original feature maps): 396.85 s, 15.98 GiB.

    (3) R2 (cumulative from the Original feature maps): 589.32 s, 15.98 GiB.

    (4) Propagation: 187.04 s, 13.15 GiB.

  3. Table 7 reports the costs and peak GPU memory for feature extraction, feature refinement, and feature-field construction by downstream methods.

  4. Table 7 shows that ReliaSplat takes 776.36 s to construct its feature field, longer than VALA and THGS but approximately 81.8% and 94.6% less time than LangSplat and LangSplatV2, respectively.

Q4: Robustness to Systemic Multi-View Hallucinations

  1. ReliaSplat targets open-vocabulary segmentation by directly aggregating and refining continuous CLIP features, without relying on predefined class labels or assigning discrete class labels to observations from individual views.

  2. Cross-view CLIP agreement need not imply correctness; valid minority observations may be downweighted.

  3. The refinement teacher is obtained by alpha-compositing the consensus features of multiple Gaussians and pooling them within SAM masks, allowing it to integrate regional evidence. Thus, an incorrect consensus at a single Gaussian does not independently determine refinement for the entire region; correct evidence from other Gaussians in the region may still contribute. However, if the erroneous consensus is widespread throughout the region, the mechanism cannot guarantee recovery of the correct semantics.

  4. Figure 6 and Table 12 show gains in most scenes and most valid classes.

Thank you for considering our responses. We would be happy to clarify any points that remain unclear.

Dear Reviewer,

As the rebuttal discussion period is approaching its deadline, we would appreciate it if you could review our response and share any further comments or concerns.

Thank you for your time and consideration.

Reviewer_nx6KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ReliaSplat, a training-free framework designed to improve the quality and reliability of pre-extracted 2D language features used in Language Gaussian Splatting。 While prior LGS works primarily focus on how 2D features are represented, distilled, or aggregated in 3D, they generally treat the input 2D vision-language features as ground-truth observations—ignoring corruptions caused by occlusion, scale variation, region-partition errors, and vision-language model (VLM) noise.

Critical strengths

Identifying input 2D observation noise as a primary bottleneck—and decoupling observation refinement from the downstream 3D representation. The cross-framework generalization experiments convincingly demonstrate that ReliaSplat is not merely another standalone heuristic aggregation method, but a versatile pre-processing/refinement module that improves both training-based and training-free frameworks without modifying their core architectures.

The training-free 3D→2D→3D3\text{D} \to 2\text{D} \to 3\text{D} feedback loop is clean and well-motivated.

Critical weaknesses
  • While the third module (Geometry-Aligned Semantic Propagation) introduces substantial procedural complexity and numerous hyperparameters (Beta(0.2,0.2)\text{Beta}(0.2, 0.2) sampling, camera pairing thresholds, τz=0.02\tau_z = 0.02, τf=0.9\tau_f = 0.9, λ=1\lambda = 1), its empirical benefit is nearly negligible. In Table 4(C), adding propagation on top of R2-refined features improves LERF mean mIoU by only +0.13% (from 64.12% to 64.25%). Similarly, in Appendix Tables A.1 and A.2, propagation improves ScanNet mIoU by only 0.15% to 0.18%. Furthermore, the cross-framework experiments in Table 5 do not even use this module (they only use R1 refined real-view features). The paper's complexity-to-gain ratio is significantly weakened by heavily featuring a third component that contributes almost nothing over the first two stages.

  • Because ReliaSplat is pitched as a training-free framework and pre-processing enhancer, wall-clock runtime and memory overhead are critical metrics. Each refinement round requires rasterizing 512-dimensional features across all training views, average-pooling over regions, and re-aggregating across Gaussians. However, the paper provides no wall-clock timing comparisons showing how long R1, R2, and Propagation take relative to baseline feature extraction and aggregation.

Questions for authors
  • What is the exact wall-clock time and GPU memory overhead required to run Cross-View Reliability-Aware Aggregation, one round of Consensus-Guided Feature Refinement (R1), two rounds (R2), and Geometry-Aligned Semantic Propagation on a standard LERF scene? How does this compare to the runtime of the downstream methods in Table 5 (e.g., VALA, Occam's LGS, and LangSplat)?

  • In Section 3 and 4.2, features fivf_{iv} are indexed by Gaussian ii and view vv, whereas Section 4.3 introduces region jj (Xjv,ZjvX_{jv}, Z_{jv}). Could you clarify how Gaussians are mapped to 2D regions jj during the initial aggregation step (e.g., SAM mask assignment following LangSplat), and whether all pixels inside region jj share the exact same initial feature XjvX_{jv}?

-In Eq. (6), the similarity pivot τ=0.99\tau = 0.99 and minimum similarity ρmin⁡=0.90\rho_{\min} = 0.90 define a very narrow cosine similarity band [0.90,0.99][0.90, 0.99] before saturating at λmax⁡=0.5\lambda_{\max} = 0.5. How sensitive is the refinement process to τ\tau, ρmin⁡\rho_{\min}, and λmax⁡\lambda_{\max} across different VLM backbones or feature dimensions?

2 comments

We thank the reviewer for the careful review and constructive suggestions. These comments have helped us strengthen our experimental analysis and improve the clarity of our method description and presentation of results. We address your questions point by point below.

Revisions and additions are highlighted in blue in the updated PDF for ease of review.

Q1: What are the exact wall-clock time and GPU memory costs of the modules on a standard LERF scene?

How do these costs compare with the runtimes of the downstream methods in Table 5?

  1. Table 6 reports detailed runtime and peak memory for each stage across the four LERF scenes.

  2. For figurines, the runtime and peak GPU memory for the four stages are:

    (1) Aggregation (including qiq_i precomputation): 355.85 s, 12.67 GiB.

    (2) R1 (cumulative from the Original feature maps): 396.85 s, 15.98 GiB.

    (3) R2 (cumulative from the Original feature maps): 589.32 s, 15.98 GiB.

    (4) Propagation: 187.04 s, 13.15 GiB.

  3. Table 7 reports the costs and peak GPU memory for feature extraction, feature refinement, and feature-field construction by downstream methods.

  4. Table 7 shows that ReliaSplat takes 776.36 s to construct its feature field, longer than VALA and THGS but approximately 81.8% and 94.6% less time than LangSplat and LangSplatV2, respectively.

Q2: How are Gaussians mapped to 2D regions jj during initial aggregation (e.g., is SAM-mask assignment as in LangSplat used)? Also, do all pixels within the same region jj share exactly the same initial feature XjvX_{jv}?

  1. The initial feature maps are extracted as follows: SAM generates 2D region masks, and CLIP extracts a semantic feature for each mask, denoted by XjvX_{jv}.

  2. For Gaussian ii in view vv, its 3D center is projected onto the image plane, and the SAM region index jj at that projected location is read. Its initial observation feature is therefore fiv=Xjvf_{iv}=X_{jv}.

  3. All pixels in the same SAM mask share exactly the same initial feature XjvX_{jv}. Consequently, different Gaussians projected into the same region also share the initial observation feature for that view.

Q3: How sensitive is the refinement process to τ\tau, ρmin⁡\rho_{\min}, and λmax⁡\lambda_{\max} across different VLM backbones or feature dimensions?

  1. Open-vocabulary image-text retrieval typically matches text and images semantically using cosine similarity in a shared feature space. We L2-normalize the 2D and teacher features to remove differences in magnitude, adjust refinement strength based on cosine similarity, and bound the maximum refinement strength; the formulation does not require a fixed feature dimensionality.

  2. Figure 5 compares ρmin⁡∈{0.70,0.80,0.90}\rho_{\min}\in\{0.70,0.80,0.90\} while fixing τ=0.99\tau=0.99 and λmax⁡=0.5\lambda_{\max}=0.5. Mean R3 mIoU across the four LERF scenes increases consistently as ρmin⁡\rho_{\min} rises, indicating that performance is relatively insensitive to this parameter within the tested range.

  3. LERF-OVS and ScanNet use the same globally fixed thresholds without per-scene tuning. Figures 5 and 6 show improvements in most of the four LERF-OVS scenes and ten ScanNet scenes, indicating that the method is robust to scene variation.

W1: Role and cost of the third module

  1. Semantic propagation is used only to construct ReliaSplat's own 3D semantic field and does not participate in input feature refinement. For ReliaSplat, the additional cost is modest; Table 6 reports runtime and peak GPU memory for each scene.

  2. Table 1 evaluates the performance of the 3D semantic field constructed by ReliaSplat itself and provides a fair comparison with other methods.

  3. Table 5 evaluates ReliaSplat's improvement of the original features by replacing them with R1-refined real-view features to exclude the effect of view densification on the results.

  4. Table 9 shows that only 31.00% of pixels in the R2 propagated feature maps are valid, compared with 90.94% in the source feature maps, so propagation remains sparse. Considering the costs and benefits of storing dense feature maps for these sparse observations and using them for downstream training, we aggregate propagated observations online only into ReliaSplat's own feature field, without generating dense pseudo-view feature maps.

Thank you for considering our responses. We would be happy to clarify any points that remain unclear.

Dear Reviewer,

As the rebuttal discussion period is approaching its deadline, we would appreciate it if you could review our response and share any further comments or concerns.

Thank you for your time and consideration.

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

This paper presents ReliaSplat, a training-free framework for improving language features used in Language Gaussian Splatting. Given reconstructed Gaussian geometry and pre-extracted multi-view language features, it estimates observation reliability using cross-view cosine agreement and combines this with rendering significance for aggregation. The resulting 3D semantic consensus is rendered back into the input views to adaptively refine their region-level features. A further propagation stage transfers real-view features to intermediate viewpoints under depth and semantic consistency constraints. Experiments evaluate open-vocabulary segmentation on LERF and ScanNet, together with component ablations and feature-replacement experiments across four downstream frameworks.

Critical strengths
  1. A practical feature-enhancement interface with evidence of cross-framework utility. The method improves the observations supplied to downstream systems rather than requiring a new language-field architecture. In Table 5, replacing only the input features improves mean mIoU across LangSplat, LangSplatV2, VALA, and THGS. The gains are particularly meaningful for LangSplat, from 49.10 to 53.86, and LangSplatV2, from 61.02 to 63.16. Although some individual scenes regress, the paired experiments provide useful evidence that the refined features benefit multiple methods rather than only the proposed aggregation rule.
  2. A coherent formulation supported by component-level evaluation. The distinction between rendering significance and semantic agreement is well motivated. Tables 3–4 evaluate their combination, teacher construction, refinement rounds, and propagation filters. On LERF, significance-only aggregation obtains 61.69 mIoU, reliability-aware aggregation reaches 62.84, and one refinement round reaches 63.90. The ScanNet experiments and supplementary per-scene results provide additional evidence beyond image-space evaluation on LERF. These studies support the usefulness of the overall enhancement pipeline.
Critical weaknesses
  1. The distinctive contribution and the reliability interpretation are insufficiently established. ReLaGS already uses the same self-excluded mean-cosine statistic for object-level outlier rejection, while VALA explicitly addresses noisy multi-view features through robust aggregation. The contribution therefore lies mainly in Gaussian-level weighting and reusable 3D-to-2D feature correction, not semantic reliability estimation broadly. However, the experiments do not isolate adaptive correction from fixed-strength consensus smoothing or alternative robust-aggregation teachers. Moreover, agreement need not imply correctness: correlated errors may form the consensus, while a correct but uncommon observation may be suppressed. Direct validation of reliability scores and correction outcomes would strengthen this central claim.
  2. The significance and practical cost of the remaining improvements are unclear. Some additions contribute only small gains: self-exclusion changes LERF mIoU from 62.83 to 62.84, while propagation after R2 changes it from 64.12 to 64.25. ScanNet propagation gains are similarly small. These results warrant a quality–cost analysis, but end-to-end runtime and peak memory are not reported. Experimental comparability also needs clarification: Tables 1–2 use published baseline results, whereas the locally evaluated baselines differ substantially—for example, VALA obtains 61.7 in Table 1 but 55.35 in Table 5. The paired improvements remain useful, but broader superiority claims require a reconciled protocol and clearer parameter-validation procedures.
Questions for authors
  1. Contribution isolation: What distinguishes the proposed reliability estimator from ReLaGS’s Eq. (7), beyond its Gaussian-level application and soft weighting? Have you compared adaptive correction with fixed-strength refinement using the same teacher, or with a VALA/ReLaGS-based robust teacher under matched settings?
  2. Reliability validation: How well do the consistency scores predict actual semantic errors? Can you show whether refinement corrects erroneous observations without erasing valid view-dependent or part-specific information, particularly when incorrect features agree across several views? Please also clarify how the correction thresholds were selected and whether validation scenes were separate from evaluation scenes.
  3. Protocol and efficiency: What explains the baseline differences between Tables 1 and 5? Please specify the geometry checkpoints, feature extraction, annotations, and metric definitions used in each setting, and report runtime and peak memory for aggregation, refinement, and propagation separately.
2 comments

We thank the reviewer for the careful review and constructive suggestions. These comments have helped us strengthen our experimental analysis and improve the clarity of our method description and presentation of results. We address your questions point by point below.

Revisions and additions are highlighted in blue in the updated PDF for ease of review.

Q1: Contribution isolation

  1. ReliaSplat and Eq. (7) of ReLaGS use the same self-excluded mean cosine similarity (Eq. (1)); we do not claim that this metric is new. We have clarified this in Section 2.2 of the main text.

  2. The methodological distinction lies in its use and weighting: ReLaGS uses this statistic for object-level rejection of anomalous observations; ReliaSplat combines semantic consistency cc with RGB rendering significance ss into a Gaussian-level soft weight w=s cw=s\,c, and further uses it for iterative 2D feature refinement guided by region-level teachers.

  3. Eq. (3) is equivalent to Eq. (1), reducing consistency computation from O(Ni2D)O(N_i^2D) to O(NiD)O(N_iD) (Appendix A.2). In Table 6, qiq_i-efficient uses this formula and reduces mean preparation time across four scenes by 14.9%, with unchanged peak GPU memory and consistency metric.

  4. Figure 4 compares fixed-weight and adaptive fusion from the same R0 features and RGB Gaussian checkpoint. No fixed weight is optimal across scenes; adaptive fusion outperforms all fixed settings at R3 on all four scenes.

Q2: Reliability validation

  1. The predictive ability of the consistency score has so far been validated only indirectly: Table 4(A) compares w=cw=c with w=s cw=s\,c; the ScanNet 10-class results in Table 11 show gains of 2.05, 1.58, and 1.72 pp in mIoU, mAcc, and OA, respectively, with more corrected errors (23,377) than new errors (9,103).

  2. The refinement teacher is obtained by alpha-compositing the consensus features of multiple Gaussians and pooling them within SAM masks, allowing it to integrate regional evidence. Thus, an incorrect consensus at a single Gaussian does not independently determine refinement for the entire region; correct evidence from other Gaussians in the region may still contribute. However, if the erroneous consensus is widespread throughout the region, the mechanism cannot guarantee recovery of the correct semantics.

  3. Pseudo-labels are SAM-mask-level CLIP features; annotations do not support dedicated tests of view-dependent or part-level semantics or shared errors. Incorrect majority consensus may weaken correct minority observations. Figures 2-3 show localization and mask improvements in test-view examples; Figure 6 and Table 12 show gains in most scenes and most valid classes.

  4. Figure 4 shows that adaptive fusion outperforms fixed settings at R3 on all four scenes. Assuming greater stability of multi-view Gaussian consensus, we preset ρmin⁡=0.9\rho_{\min}=0.9, below which refinement has maximum strength. This threshold was not selected using query annotations; Figure 5 reports its sensitivity.

  5. LERF-OVS provides only a few evaluation frames with text queries and mask annotations for open-vocabulary retrieval. Training and feature refinement do not access query GT; annotations are used only for evaluation and the sensitivity analysis in Figure 5, not parameter selection. The same parameters are applied directly to LERF-OVS and ScanNet without dataset- or scene-specific tuning.

Q3: Protocol and efficiency

  1. The differences may arise from upstream settings: pretrained Gaussian models depend on the initial point cloud, training configuration, random seeds, and random initialization; supervision pseudo-features depend on SAM-generated masks, and SAM versions and parameters can affect region partitioning and feature extraction.

  2. Section 5.1 of the revised manuscript adds details on geometry checkpoints, feature extraction, and evaluation settings; Appendix A.1 specifies dataset annotations and metric definitions.

  3. Table 6 reports detailed runtime and peak memory for each stage across the four LERF scenes.

W2: Role and cost of the third module

  1. Propagation only constructs ReliaSplat's 3D semantic field, without refining input features. Its additional cost is modest; see Table 6 for runtime and peak GPU memory.

  2. Table 1 evaluates the performance of the 3D semantic field constructed by ReliaSplat itself and provides a fair comparison with other methods.

  3. Table 9 shows 31.00% valid pixels in R2 propagation outputs versus 90.94% in source maps. Given downstream training costs and benefits, we aggregate them online only into our own field without saving dense pseudo-view maps.

Thank you for considering our responses. We would be happy to clarify any points that remain unclear.

Dear Reviewer,

As the rebuttal discussion period is approaching its deadline, we would appreciate it if you could review our response and share any further comments or concerns.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.