acceptodds
Under review as a conference paper at ICLR 2027

UniGS-SLAM: Monocular Gaussian SLAM with Enhanced Geometry and Open-Vocabulary Semantics

Abstract

Monocular semantic SLAM aims to reconstruct scene geometry and build semantic maps from RGB observations. Existing Gaussian-based methods often attach semantic labels or features to evolving Gaussian primitives, making semantic fusion difficult to maintain and increasing reconstruction overhead. Hybrid Gaussian–voxel representations provide more stable semantic support, but often rely on RGB-D input or known poses and do not revise historical semantic evidence after geometry updates. We present , an RGB-only framework that decouples open-vocabulary semantics from Gaussian reconstruction while preserving bidirectional interaction. Foundation priors enhance monocular depth and pose estimation, while semantic observations are fused into an independent sparse voxel map with observation provenance, enabling historical contributions to be revised as reconstruction evolves. The fused semantics further guide Gaussian filtering and local surface refinement. Experiments on Replica, ScanNet, and KITTI-360 show leading tracking, rendering, and geometric reconstruction performance among representative RGB-only methods, together with strong open-set semantic mapping on Replica and ScanNet. We also achieves a 2.37 reconstruction speedup over the fastest compared semantic SLAM baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D reconstruction, Gaussian Splatting, SLAM

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_FuitOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

UniGS-SLAM combines monocular Gaussian reconstruction with a separate sparse voxel map for open-vocabulary semantics. It uses Depth Anything 3 predictions to guide pose and depth estimation, then lifts region-language observations into voxels using the reconstructed geometry. By retaining each observation’s contribution, the system can retract and reproject semantic evidence when earlier poses or depths change. The voxel map also guides Gaussian floater filtering and local surface refinement. The paper evaluates tracking, rendering, geometry, semantics, and runtime on Replica and ScanNet, plus four selected static KITTI-360 segments. It reports a 4.00 dB ScanNet PSNR gain over the strongest listed RGB-only baseline and a ScanNet open-set mIoU of 0.50 versus 0.38 for RADIO-ViPE. Removing historical semantic revision lowers ScanNet mIoU by 4.43 percentage points.

Critical strengths
  1. Keeping semantic contributions separate from Gaussian parameters makes it possible to revise evidence after a pose or depth correction. This is well matched to an evolving SLAM map, and the ScanNet ablation in Table 7 gives useful evidence that retracting old contributions matters: mIoU rises from 45.43 without revision to 49.86 with it.

  2. The reported results cover several aspects of the system rather than rendering alone. ScanNet rendering and geometric reconstruction improve substantially in the stated protocol, while the open-set semantic evaluation improves on both Replica and ScanNet. The component ablations also give some support to the geometry, revision, and refinement stages.

Critical weaknesses
  1. The paper does not compare with OpenMonoGS-SLAM, a particularly relevant prior monocular Gaussian SLAM system with open-set semantics. It appears in the bibliography but not in the evaluation or substantive related-work discussion. The paper also does not make clear which reconstruction baseline results were rerun with identical scenes, image resolution, and evaluation frames. This makes the size of the reported cross-method advantage harder to judge.

  2. The ablations show that revision helps within this system, but they do not fully separate the effects of the decoupled semantic representation from those of its strong geometry and semantic frontends. The progressive reconstruction ablation covers Replica, while the largest rendering gain is reported on ScanNet. Table 9 measures total reconstruction time, but does not show the memory cost or processing backlog from storing and revising historical semantic observations. Those measurements matter to the claimed efficiency of online semantic mapping.

Questions for authors
  1. Could you compare against OpenMonoGS-SLAM under a common protocol? If its implementation or evaluation protocol prevents this, please explain the limitation and discuss the closest results that can be compared.

  2. Which reconstruction baseline numbers were reproduced using the same scenes, resolution, and evaluation frames? Are rendering metrics computed on held-out frames for every method?

  3. On ScanNet, can you report the geometry and semantic results with the major components disabled while keeping the same Depth Anything 3 prior and region-language frontend? How much memory and processing time do stored contributions and asynchronous semantic revisions add?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Fuit (Part 1/2: Q1–Q2)

Q1: Comparison with OpenMonoGS-SLAM

We could not locate a public implementation of OpenMonoGS-SLAM, preventing a matched rerun. We therefore provide its published tracking, rendering, and closed-set segmentation results as the closest available comparisons:

DatasetMethodATE (cm) ↓PSNR ↑SSIM ↑LPIPS ↓
ReplicaOpenMonoGS-SLAM (reported)1.6034.470.9570.086
ReplicaUniGS-SLAM (ours)0.24738.650.9750.026
ScanNetOpenMonoGS-SLAM (reported)5.3926.210.8600.273
ScanNetUniGS-SLAM (ours)7.1528.650.8590.239
Closed-set methodR0R1R2O0Avg. mIoU (%)
OpenMonoGS-SLAM (GT, reported)88.6088.1090.0090.4089.28
UniGS-SLAM (ours, GT supervision)95.3996.8596.2695.1995.92

Our results indicate stronger rendering and closed-set segmentation, while OpenMonoGS-SLAM reports lower ScanNet ATE. The rendering protocols differ: it evaluates all non-keyframes at 512-pixel width, whereas we sample every fifth non-keyframe at each method's native resolution. Its open-set evaluation reports 2D segmentation rather than 3D mIoU. We will add these comparisons and discuss the methodological differences in the revised manuscript.

Q2: Evaluation settings and result sources

For reproduced methods, we follow the evaluation procedure of DROID-Splat: ATE is computed on all frames with valid GT poses, and rendering metrics on every fifth non-keyframe, excluding each method's keyframes. We use the same evaluation scenes, but retain repository-recommended resolutions, which may differ across methods. The rendering metrics are evaluated on non-keyframes, although the exact frame sets may differ due to method-specific keyframe selection.

Paper-reported results are used for Open-set SLAM, Hier-SLAM++, and OpenMonoGS-SLAM in Table 1, and for Open-set SLAM, SGS-SLAM, GS3LAM, OpenGS-SLAM, and OpenMonoGS-SLAM in Table 4. The remaining entries are reproduced. We will label these sources explicitly; the common sampling rule does not imply identical frame sets across methods.

AuthorsOpenReview ↗

Response to Reviewer Fuit (Part 2/2: Q3)

Q3 (a): Component ablations on ScanNet

We evaluated six ScanNet scenes, disabling historical revision and semantic-guided refinement independently. All four semantic variants retain the same DA3 prior, geometry enhancement, region-language frontend, and voxel representation; the bare SLAM backbone is included as a reference.

ConfigurationHistorical revisionSemantic-guided refinementPSNR ↑Depth L1 (cm) ↓mAcc (%) ↑3D mIoU (%) ↑
SLAM backbone (reference)——24.2212.337——
Full system without revision or refinementNoNo26.3011.48554.1344.38
Full system without refinementYesNo26.5311.40060.6149.70
Full system without revisionNoYes28.529.50055.8545.43
Full UniGS-SLAM (ours)YesYes28.659.38860.8049.86

The components serve complementary roles: historical revision corrects misplaced semantic evidence as geometry evolves, while semantic-guided refinement improves reconstruction through better surface selection and floater filtering. Combining them gives the strongest overall results. The original Appendix Table 12 further shows that semantic guidance improves upon the same geometric refinement without semantic cues.

Q3 (b): Contribution of semantic decoupling

The ablations in Q3 keep the geometry and semantic frontends fixed, isolating the contribution of historical revision and semantic-guided refinement. Without refinement, historical revision improves 3D mIoU from 44.38% to 49.70%. Our decoupled voxel map enables independent maintenance of semantic evidence, allowing outdated contributions to be retracted and re-fused without coupling semantic updates to Gaussian optimization.

Q3 (c): Storage and revision overhead

We profiled semantic mapping on Office0, enabling historical storage and then asynchronous revision:

ConfigurationMeasured time (s)Peak CPU (MB)Peak GPU (MB)
Decoupled mapping78.211542.37413.14
+ Historical storage79.691620.39413.14
+ Asynchronous revision124.741694.27413.14

Caching historical evidence adds little processing time; the main overhead comes from reprojecting and re-fusing observations after geometry changes. This trades additional CPU memory and computation for a semantic map consistent with the corrected reconstruction, without increasing the measured semantic-mapping GPU peak.

Thank you for the clarifications and additional experiments. The OpenMonoGS-SLAM comparison and fixed-frontend ScanNet ablations address much of my concern, particularly by demonstrating the benefit of historical semantic revision.

Differences in evaluation frames and resolutions still limit cross-method comparisons, and the online memory and revision costs warrant clearer discussion. Nevertheless, the response strengthens the paper sufficiently for me to raise my rating from 2 to 3.

AuthorsOpenReview ↗

Response to Reviewer Fuit

We sincerely thank the reviewer for the thoughtful follow-up and constructive suggestions. We appreciate the recognition of our additional experiments and the improved rating.

Regarding evaluation frames, we ensure that pose evaluation uses the same frames with valid GT poses across reproduced methods. For rendering, we follow the evaluation protocol adopted by traditional Gaussian SLAM methods (like MonoGS), evaluating every fifth non-keyframe. Since different SLAM methods select keyframes with different frequencies, identical evaluation frames cannot always be guaranteed under this protocol. Including keyframes in a fixed-interval evaluation could also introduce bias, as keyframes typically achieve better rendering quality due to direct optimization. We therefore consider the current non-keyframe protocol a reasonable compromise for fair comparison.

Regarding resolution, we use each baseline's recommended resolution where available to allow it to perform at its best, with the goal of ensuring a fair comparison. We also acknowledge that resolution differences may affect cross-method comparisons. We will conduct additional experiments with fully matched resolutions and include the results in the final version.

Regarding online overhead, we have conducted further analyses of online memory and revision costs, including memory and revision time scalability with increasing sequence lengths (Table 17) and the memory and computational overhead of large-scale pose corrections (Table 18). These experiments further characterize how memory consumption and revision time scale with the number of retained and affected observations. We will also provide additional analysis of asynchronous processing backlog in the final version.

We sincerely appreciate the reviewer's valuable feedback, which has helped improve the rigor and clarity of our work.

Reviewer_Rjt9OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents UniGS-SLAM, an RGB-only Gaussian SLAM framework for geometric reconstruction and open-vocabulary semantic mapping. The method incorporates foundation-model geometry priors through context selection, anchor alignment, confidence-aware depth fusion, and long-horizon pose guidance. It maintains a Gaussian map for geometry and appearance alongside an independent sparse voxel map for semantic information. An observation-level semantic revision mechanism allows previously fused semantic evidence to be retracted and reprojected following updates to camera poses or reconstructed geometry. The semantic map further guides Gaussian refinement through floater filtering and local surface constraints. The framework is evaluated on Replica, ScanNet, and KITTI-360, covering camera tracking, rendering quality, geometric reconstruction, semantic mapping, and computational efficiency.

Critical strengths
  1. Practical and well-structured framework. The paper presents a unified RGB-only semantic SLAM system that combines Gaussian reconstruction with a spatially decoupled open-vocabulary semantic map. The observation-level semantic revision mechanism is particularly interesting, as it allows previously fused semantic evidence to be updated following changes in the reconstructed geometry.

  2. Strong empirical performance and comprehensive evaluation. The method achieves competitive or superior performance across tracking, rendering, geometric reconstruction, and semantic mapping benchmarks. The paper also provides extensive ablation studies covering the main components, including geometry enhancement, historical semantic revision, and semantic-guided geometry refinement.

Critical weaknesses
  1. Limited methodological novelty in geometry enhancement. The geometry enhancement module relies heavily on DA3, while the proposed context selection, anchor alignment, confidence-aware depth fusion, and pose guidance largely build upon established geometric estimation techniques. Although the ablation studies demonstrate performance improvements, the paper does not sufficiently clarify the methodological contribution beyond integrating a strong foundation model into an existing SLAM framework.

  2. Insufficient evaluation of open-vocabulary capabilities. While Table 5 demonstrates strong open-set 3D semantic segmentation performance, the evaluation does not fully establish the effectiveness of the reconstructed representation for flexible language-based scene understanding. In particular, open-vocabulary 3D object retrieval and localization are demonstrated only qualitatively, without quantitative evaluation using diverse or fine-grained textual queries. Additional quantitative experiments on these tasks would strengthen the paper's central open-vocabulary claim.

Questions for authors
  1. Necessity of spatial decoupling. The paper motivates separating semantic voxels from Gaussian primitives to improve semantic stability and efficiency. However, structured Gaussian representations such as Scaffold-GS could potentially maintain semantic features at relatively stable spatial anchors. Could the authors clarify the advantages of maintaining an independent semantic voxel map over such unified representations?

  2. Scalability of historical semantic revision. The proposed revision mechanism requires caching historical observations and their voxel contributions. How does its memory and computational overhead scale with longer sequences, particularly when large-scale pose corrections or loop closures trigger revisions of many historical observations?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Rjt9

Q1: Contribution beyond integrating DA3

UniGS-SLAM is a complete SLAM system that we built with a single tracking/BA frontend and a Gaussian mapper. Our geometry enhancement addresses the scale inconsistencies and long-sequence limitations of windowed foundation predictions. The Office0 ablation shows why directly adding DA3 is insufficient and how our modules make its predictions effective within SLAM.

ConfigurationPSNR ↑Depth L1 (cm) ↓ATE (cm) ↓
Backbone without DA336.331.9390.839
+ DA3, naive integration33.681.9642.010
+ Context & anchor alignment34.791.3100.713
+ Confidence-aware depth fusion40.341.0830.623
+ Long-horizon pose guidance42.611.0230.207

Context and anchor alignment reconcile prediction windows, confidence-aware fusion selects reliable geometry, and long-horizon pose guidance improves initialization before visual BA. Their combined benefit comes from making foundation predictions consistent with ongoing SLAM optimization, rather than simply attaching DA3 to the backbone.

Q2: Open-vocabulary retrieval and localization

We evaluated 64 annotated objects across all eight Replica scenes, using the same queries and evaluation procedure for all methods. Queries cover categories, attributes, functions, and spatial relations. We match textual queries with 3D semantic features to retrieve and localize objects.

MethodRecall@1 ↑Recall@5 ↑[email protected] ↑[email protected] ↑
RADIO-ViPE18.7540.6312.504.69
FreeOcc12.5032.8110.941.56
UniGS-SLAM (ours)32.8156.2525.0017.19

All values are percentages; [email protected]/0.5 measures localization accuracy at the corresponding 3D box IoU thresholds. Our method leads in both retrieval and localization, showing that the reconstructed semantic map supports language-guided object search and accurate 3D grounding beyond segmentation.

Q3: Necessity of spatial decoupling

Scaffold-GS anchors offer promising support for semantic fusion. Our decoupled map additionally preserves observation-level contributions, so outdated evidence can be retracted and re-fused after pose or depth corrections without semantic re-encoding. This keeps semantic maintenance independent of Gaussian densification, pruning, and feature optimization. The ScanNet revision ablation supports this benefit: removing outdated contributions is more effective than reprojection alone.

Q4: Scalability of historical semantic revision

We cache region observations and their voxel contributions, then revise only those affected by geometry updates. Reusing masks and embeddings avoids repeated semantic encoding. The Office0 sequence-length profile below shows how historical storage and revision cost grow with the retained observations.

Input framesRetained observationsHistorical storage (MB)Revision time (s)Additional CPU (MB)
50029720.0822.2670.89
1,00044233.5440.59108.21
1,50083155.4884.17152.12
2,00095070.85100.88196.39

Historical storage grows approximately with the retained observations in this experiment. To test the additional workload from large pose corrections, we separately simulated updates affecting increasing fractions of historical frames:

Corrected framesRevised observationsRevision time (s)Additional peak CPU (MB)
25%25922.61211.34
50%48940.07292.90
100%95084.62357.88

Even when all retained observations require revision, the additional memory remains on the CPU, with no measured peak-GPU increase. The main scaling cost is reprojecting and updating more observations; asynchronous processing accommodates this work, although extensive corrections delay completion of the semantic updates.

Thanks for authors' detailed response. Most of my concerns have been addressed. I would also appreciate more qualitative visualizations demonstrating the open-vocabulary capabilities in the revised paper.

My remaining concern is the comparison with alternative structured Gaussian representations, particularly Scaffold-GS-based approaches. Existing studies have explored incorporating semantic features into anchor-based representations, and I believe a discussion of their trade-offs with the proposed decoupled design would be valuable, especially regarding storage and computational efficiency, as well as the potential benefits of joint geometry-semantic optimization.

Overall, I will maintain my original score.

AuthorsOpenReview ↗

Response to Rjt9

We thank the reviewer for the thoughtful follow-up and constructive suggestions.

We will add more qualitative visualizations of open-vocabulary retrieval and localization to the final version of our paper and project page.

We agree that Scaffold-GS-based representations offer potential advantages in feature sharing and joint geometry-semantic optimization. Our decoupled design instead prioritizes revisable semantic evidence during geometric updates. We are currently exploring a Scaffold-GS-style semantic backend to compare storage and computational efficiency. Since a fair comparison involves implementing and validating a new semantic Gaussian backend, we will include experimental results in the final version if they are sufficiently validated. We will also further discuss these design trade-offs in the final version of the manuscript.

Overall, We sincerely appreciate the reviewer's valuable feedback.

Reviewer_949TOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

UniGS-SLAM is an monocular semantic Gaussian SLAM that decouples open-vocabulary semantics from Gaussian reconstruction while keeping a bidirectional interaction: foundation-model geometry priors (Depth Anything 3) enhance tracking and reconstruction through anchor alignment, confidence-aware depth fusion, and long-horizon pose guidance. Semantics are fused into an independent sparse voxel map so that historical contributions are revised after pose/depth updates. Additionally the fused semantics feed back to filter floaters and refine local surfaces. Evaluated on Replica, ScanNet, and KITTI-360, it shows good performance with shorter running time.

Critical strengths
  1. The manuscript is well-written and easy to follow.
  2. The proposed method is reasonable: most previous semantic/open-vocabulary SLAM works let geometry support semantics, whereas relatively few let semantic understanding feed back to refine geometry. The proposed semantic-to-geometry loop (verified floater filtering + semantic-compatible surface refinement) is a reasonable and meaningful contribution.
Critical weaknesses
  1. Missing SOTA comparisons in the main tables: In Table 1, the RGB-D semantic group omits Hier-SLAM [1], which reportedly outperforms all currently presented RGB-D methods, and Table 4 similarly lacks a Hier-SLAM comparison. On the other hand, the RGB-only group in Table 1 additionally omits MonoGS [2].
  2. A heavy system whose necessity is not justified. The underlying SLAM appears to stack DROID-SLAM, a Splat-SLAM, and a Depth-Anything-3 foundation prior. It is unclear why combining two SLAM components plus one foundation model is necessary to obtain this system, and no analysis isolates the cost/benefit of each model; the design rationale for such a heavy pipeline should be argued.
  3. Ablation and depth-accuracy analysis are insufficient. It is not quantified how much a bare foundation SLAM baseline achieves, how large the gain of the DA3 component alone is, and whether SLAM optimization can further improve the raw depth-estimation accuracy of the foundation prior. In particular, the depth accuracy of DA3 itself, versus the depth accuracy of the SLAM-produced map after taking DA3 as input, is not reported.
  4. Foundation-model selection is not justified or ablated. The choice of Depth Anything 3 is not compared against other geometry foundation models such as VGGT [3], VGGT-Ω [4], and π³ [5], whose depth-estimation quality differs. It should be analyzed how the choice of foundation model affects the final tracking/geometry/semantics, and why DA3 was selected over these alternatives.
  5. Teaser readability. The words in the middle column of the teaser figure are too small to read, and the middle part appears empty. It could improve the impression of the manuscript if the authors could update the teaser.
Questions for authors

I would consider raising my score if the authors could answer the following questions using experiments if necessary:

  1. Can authors add Hier-SLAM and MonoGS to Tables 1 and 4? And updating the teaser would improve the first impression of the manuscript.
  2. Does the underlying SLAM indeed consist of DROID-SLAM + Splat-SLAM + Depth Anything 3? What is the necessity and the individual contribution of each of the three heavy components?
  3. What is the performance of the bare foundation SLAM baseline, and how large is the gain from adding DA3 alone?
  4. Does the SLAM optimization improve the raw depth estimated by DA3? Please report the depth accuracy of DA3 itself versus the depth accuracy of the SLAM-built map when DA3 is used as input.
  5. How do different foundation models (VGGT, VGGT-Ω, π³, ...) affect the final results, and why was DA3 chosen over them?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 949T (Part 1/2: Q1–Q3)

Q1: Additional baseline comparisons

Following your suggestion, we compared MonoGS (RGB) and Hier-SLAM (RGB-D) using their official implementations and a common evaluation procedure. Pose estimation, reconstruction, and closed-set 2D segmentation results are shown below.

DatasetMethodATE (cm) ↓PSNR ↑SSIM ↑LPIPS ↓
ReplicaMonoGS32.97428.510.8840.157
ReplicaHier-SLAM0.34735.300.9260.048
ReplicaUniGS-SLAM (ours)0.24738.650.9750.026
ScanNetMonoGS122.05520.110.6980.577
ScanNetHier-SLAM8.3723.500.6150.507
ScanNetUniGS-SLAM (ours)7.1528.650.8590.239
MethodR0R1R2O0Avg. mIoU (%) ↑
Hier-SLAM93.7595.7195.2194.8894.89
UniGS-SLAM (ours)95.3996.8596.2695.1995.92

UniGS-SLAM (ours) achieves better pose estimation and reconstruction results than both baselines, and higher segmentation accuracy than Hier-SLAM. MonoGS does not model semantics.

Q2: Questions about the pipeline and components

UniGS-SLAM is a complete SLAM system that we built with a single tracking/BA frontend and a Gaussian mapper. We adopt DROID-style visual correspondence estimation and bundle adjustment for poses and local depth, and build on Splat-SLAM's Gaussian mapping components for geometry and appearance reconstruction. Our context alignment, confidence-aware fusion, and pose guidance integrate DA3 predictions into this system to improve tracking and mapping.

We profile the components on Replica Office0 (2,000 frames, one RTX A6000); total reconstruction time on Replica is also reported below.

ComponentTracking / BAGaussian mappingDA3 inferenceContext & alignmentDepth fusionPose guidanceTotal reconstruction (Replica)
Time (min)4.547.7610.004.330.500.3120.40

DA3 and Gaussian mapping account for substantial computation, providing complementary geometry and high-quality reconstruction. Our total reconstruction time (20.40 min) remains lower than existing semantic SLAM methods (48.32 min for SNI-SLAM, 132.79 min for SemGauss-SLAM, and 250.80 min for Hier-SLAM); Q3 shows the benefits of our prior integration.

Q3: Why are the geometry enhancement modules needed?

Directly adding DA3 is insufficient: its predictions can have inconsistent scales across frames and windows, while its bounded prediction windows provide limited long-range geometric constraints. For naive integration, chunk-wise DA3 predictions are directly integrated into SLAM optimization without our proposed enhancements. The Office0 ablation below shows how our modules make its geometric predictions effective within SLAM.

ConfigurationPSNR ↑Depth L1 (cm) ↓ATE (cm) ↓
Backbone without DA336.331.9390.839
+ DA3, naive integration33.681.9642.010
+ Context & anchor alignment34.791.3100.713
+ Confidence-aware depth fusion40.341.0830.623
+ Long-horizon pose guidance42.611.0230.207

Context selection incorporates reliable historical views into bounded prediction windows, while anchor alignment reconciles predictions across windows; confidence-aware fusion selects reliable depth information; long-horizon pose guidance leverages historical trajectory proposals to improve initialization before visual BA. Together, these designs turn DA3's local geometric capability into consistent long-sequence reconstruction.

AuthorsOpenReview ↗

Response to Reviewer 949T (Part 2/2: Q4–Q5 & teaser)

Q4: Does SLAM improve DA3 depth?

On Office0, raw DA3 and reconstructed depth are compared under the same global-scale evaluation. Raw DA3 uses the same chunk-wise predictions as naive integration but without subsequent SLAM optimization.

Depth sourceDepth L1 (cm) ↓
Raw DA318.312
Backbone without DA31.939
Naive DA3 integration + SLAM optimization1.964
Full UniGS-SLAM1.023

As discussed in Q3, scale inconsistencies and long-sequence limitations hinder the direct use of DA3 predictions. Our method addresses these issues within SLAM, translating its geometric capability into more accurate, globally consistent depth.

Q5: Choice of geometry foundation model

We replaced DA3 while keeping the backbone, integration modules, and semantic pipeline unchanged on Office0:

Geometry priorATE (cm) ↓PSNR ↑Depth L1 (cm) ↓3D mIoU (%) ↑
No prior0.83936.331.93928.57
π³0.23241.461.46632.03
VGGT0.23841.181.26632.33
VGGT-Ω0.20642.281.08533.46
DA30.20742.611.02334.52

Our framework can work with different geometry foundation models. We chose DA3 for its best overall results in this comparison, while VGGT-Ω achieves closely comparable performance.

Q6: Teaser readability

Thank you for your suggestion. We have revised the plots in the middle section of the teaser to improve readability and included the updated figure in the newly submitted PDF. We will further refine the figure to make it more visually appealing.

Reviewer_949TOpenReview ↗

Raise my rating

Thanks the authors for the additional experiments and analysis. Most of my concerns have been addressed, including the comparisons with Hier-SLAM and MonoGS, the performance of DA3, and the comparison across foundation models. I would still suggest an ablation that demonstrates the significance of each component rather than relying on runtime cost alone. Nevertheless, my main concerns have been addressed, and I have raised my rating to 3. I hope the authors will further improve the manuscript quality by adding these rebuttal experiments and analysis.

AuthorsOpenReview ↗

Response to Reviewer 949T

We are really grateful for the reviewer‘s positive feedback and constructive suggestions.

We have incorporated the additional rebuttal experiments and analysis into the manuscript (DA3 performance in Table 7 and foundation model comparison in Table 15) and will further improve its clarity and presentation.

We agree that performance ablations are important for demonstrating the significance of individual components. Beyond the experiments already presented in our rebuttal, we have conducted additional ablations:

1. Overall Component Ablation (Table 12). We evaluate geometry enhancement and semantic-guided refinement across all eight Replica scenes.

ConfigurationPSNR ↑SSIM ↑LPIPS ↓L1 (cm) ↓
w/o Both36.790.9580.0361.564
+ Geometry Enhancement38.370.9740.0301.260
+ Semantic-Guided Refinement38.650.9750.0261.170

Geometry enhancement improves PSNR by 1.58 dB and reduces depth error by 19.4%, while semantic-guided refinement provides further improvements.

2. Semantic Component Ablation (Table 8). We independently evaluate historical semantic revision and semantic-guided refinement across six ScanNet scenes.

ConfigurationPSNR ↑Depth L1 (cm) ↓3D mIoU (%) ↑
w/o Both26.3011.48544.38
+ Historical Revision26.5311.40049.70
+ Semantic-Guided Refinement28.529.50045.43
+ Both (Ours)28.659.38849.86

Historical revision improves 3D mIoU by 5.32 percentage points without refinement, while semantic-guided refinement improves PSNR by 2.22 dB without revision, demonstrating their complementary contributions.

We also incorporate these ablations into the manuscript to further strengthen the experimental analysis. Again, We appreciate the reviewer's constructive suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.