acceptodds
Under review as a conference paper at ICLR 2027

ISR-SLAM: Multi-Agent 3DGS SLAM via Inter-Submap Reinforcement

Abstract

Multi-agent simultaneous localization and mapping (SLAM) based on 3D Gaussian Splatting (3DGS) enables high-fidelity collaborative 3D reconstruction. A conventional pipeline partitions each agent's observations into short temporal segments and constructs a local submap with 3DGS for each segment. Reconstructed independently, the local 3DGS submaps are registered into a global map by optimizing inter-submap relative poses, after which the global 3DGS map is refined as a whole. However, this procedure breaks up submap registration and global map refinement. Consequently, defects in submaps can undermine the registration, which in turn hinders the global map refinement since submaps are incorrectly registered. With this insight, we propose ISR-SLAM, a multi-agent 3DGS SLAM framework that creates an online feedback loop between submap registration and global map refinement. As new submaps income, we implement periodical registration of all available submaps, ensuring subsequent tracking and mapping to proceed with the optimal registration up-to-date. Simultaneous to registration, we leverage cross-agent observation overlap to refine 3DGS submaps. The refined submaps, in turn, facilitates subsequent registration, forming a mutually reinforcing cycle. After merging all submaps, ISR-SLAM further refines the overlapped regions and submap boundaries to reduce residual inconsistencies. Extensive experiments on synthetic and real-world multi-agent datasets demonstrate that ISR-SLAM improves localization accuracy, reconstruction quality, and global map consistency significantly over existing methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-Agent SLAM, 3D Gaussian Splatting, RGB-D SLAM, 3D Reconstruction, Pose Graph Optimization

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_QHv7OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors proposed ISR-SLAM a distributed RGB-D SLAM system which fuse map from multiple agents via inter-submap reinforcement. Compared with its backbone structure – MAGiC-SLAM, it introduced online feedback, inter-submap Gaussian refinement, and staged global refinement. The contribution here is more like an incremental and reasonable extension of MAGiC-SLAM.

Critical strengths

The core idea moves multi-agent 3DGS refinement from after registration into the mapping process, so local submaps updated with other agents’ observations can participate in the next registration round.

Critical weaknesses
  1. The pipeline and method to tackle point registraction are standard. The introduction states that existing methods merge subgraphs only after they have all been fully constructed; however, this claim is too absolute. MAC-Ego3D is also periodic; while it is included as a baseline method for comparison in this paper, its analysis should be emphasized in the Introduction and Related Work sections. It is suggested that OmniMAG-SLAM and MAC-Ego3D separately discuss the differences between their periodic constraints and those in this paper.

  2. The experimental results in this paper did not demonstrate a significant advantage, which may limit the paper's validity. Many of the indicators in the article are not the best, more than expected. Furthermore, there is a lack of explanation regarding the suboptimal indicators.

  3. As a core component, inter-submap overlap refinement needs stronger motivation and qualitative evidence. Table 3 provides an ablation, but more examples like those in Figure 5 would better show what inconsistencies it fixes, especially quantitative result.

  4. The \deltaT_r->t obtained from overlap refinement is discarded after use and does not affect the pose graph. Why should not further utilize this error pose?

  5. The main text should include a qualitative comparison of the two sets of data; "AriaMultiagent" is missing from the main text.

  6. In Table 5, why is ‘Gaussian of Ours’ missing?

  7. Although Table 5 reports per-frame tracking/mapping time and map storage, it is unclear whether the proposed online registration and overlap-refinement costs are included. And it is not efficient to just repot the time consumption and memory for one scene.

  8. If later pose-graph optimization changes the relative pose of an already refined submap pair, the previous overlap refinement may no longer be optimal.

Questions for authors
  1. In Figure 2 caption, Registeredposes lacks space.
  2. What’s the meaning of equation(4) C_{ref}? In Equation (1), using I.
  3. The notation on the left-hand side of Equation. (1) is not explicitly defined.
12 comments
AuthorsOpenReview ↗

Response to Reviewer QHv7 Critical Weaknesses 1

Critical Weaknesses 1:

We thank the reviewer for pointing out this wording issue. Our central contribution is the interaction between online submap registration and Inter-Submap Overlap Refinement. Updated poses and Gaussian representations remain available for subsequent mapping and registration. We agree that our description of existing pipelines may have been too broad and will make it more precise in the revised manuscript.

The abstract and introduction use “A conventional pipeline” and “Conventionally” to describe common staged designs (abstract, lines 35–40; introduction, lines 65–72). This wording may imply that all existing methods follow the same process. We will distinguish the periodic processing in MAC-Ego3D from the dual-factor-graph online pose optimization in OminiMAG-SLAM. We will also clarify that ISR-SLAM refines Gaussian submaps during online mapping and reuses the refined representations in subsequent processing.

AuthorsOpenReview ↗

Response to Reviewer QHv7 Critical Weaknesses 2

Response to Reviewer QHv7 Critical Weaknesses 2

We thank the reviewer for the assessment of our experimental results. Online Registration Feedback improves spatial alignment between submaps, while Inter-Submap Overlap Refinement reduces inconsistencies between Gaussian maps reconstructed independently by different agents. Together, they allow ISR-SLAM to achieve strong overall localization and reconstruction quality. The gains can vary across scenes and evaluation metrics.

In Table 1 of the original paper, ISR-SLAM achieves the best or joint-best average ATE in four of the six scenes and ranks second in the other two. In Table 2, it ranks first in four of the eight PSNR evaluations and second in the remaining four. In particular, it achieves the best PSNR in all four AriaMultiagent evaluations, covering both novel and training views. These results demonstrate strong overall performance across scenes.

Some LPIPS and Depth L1 results are indeed not the best. Our method primarily targets pose consistency and consistency between submap representations. Better spatial alignment and RGB rendering do not necessarily yield the lowest perceptual or depth error in every scene. Missing or noisy depth observations in real-world data may also limit geometric accuracy. We will clarify these trade-offs in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer QHv7 Critical Weaknesses 3 Part 1

Response to Reviewer QHv7 Critical Weaknesses 3 Part 1

We thank the reviewer for this suggestion. Submap pose alignment corrects their spatial relationships, but Gaussian maps reconstructed independently by different agents may still disagree in color, depth, and local geometry. Inter-Submap Overlap Refinement uses reliable shared observations for cross-view supervision. It directly refines the Gaussian representations in overlapping regions while preserving consistency with local RGB-D observations. This mechanism is described in Section 3.2 of the original paper (lines 252–283).

To evaluate its local effects, we add a quantitative comparison of three overlapping submap pairs (Table R1). After refinement, PSNR increases and Depth L1 decreases in all three shared regions. These results directly demonstrate how the module reduces representation inconsistencies.

Table R1. RGB and depth consistency in overlapping regions before and after Inter-Submap Overlap Refinement.

Overlapping Submap PairRGB PSNR Before Refinement (dB)↑RGB PSNR After Refinement (dB)↑Depth L1 Before Refinement (cm)↓Depth L1 After Refinement (cm)↓
Overlapping Submap Pair 139.2141.280.520.29
Overlapping Submap Pair 240.0442.730.470.25
Overlapping Submap Pair 338.8641.920.550.33

Progressive causal ablation. To evaluate Online Registration–Refinement Reinforcement, we conduct progressive ablations that test whether the benefits of online registration feedback and inter-submap Gaussian refinement extend beyond the stages in which they are applied. The results show that earlier updates can improve reconstruction and localization in later blocks, even when the corresponding optimization is no longer performed. Later optimization can also affect the final quality of earlier blocks. These findings support the proposed reinforcement mechanism.

Specifically, we divide each input RGB-D video sequence into five equal, consecutive temporal blocks (B1–B5). We conduct six ablation settings for each of two components: online pose-graph optimization (PGO) within Online Registration Feedback, and Inter-Submap Overlap Refinement. The component under study is enabled only over the first 0%, 20% (B1), 40% (B1–B2), 60% (B1–B3), 80% (B1–B4), or 100% (B1–B5) of the sequence. The other component remains enabled throughout. We evaluate rendering quality, trajectory accuracy, and geometric reconstruction accuracy for each block and the full sequence to assess the effects across stages (Tables R2–R3).

Effects of earlier optimization on later blocks. We compare the setting without the target optimization (0%) against settings where it is enabled only in earlier blocks (20%, 40%, 60%, or 80%). Specifically, we compare B2 under the 0% and 20% settings, B3 under 0% and 40%, B4 under 0% and 60%, and B5 under 0% and 80%. No new updates from the target component are performed in the evaluated block in either setting. Thus, differences in performance reflect the influence of earlier optimization on later blocks. Even after Inter-Submap Overlap Refinement is disabled, B2–B5 still benefit from earlier Gaussian refinement in reconstruction quality and trajectory accuracy. Similarly, B4 and B5 show better reconstruction quality and trajectory accuracy due to earlier online PGO, despite receiving no new PGO updates. These results indicate that both components can benefit later blocks where the corresponding optimization is no longer active.

Effects of later optimization on earlier blocks. We compare settings with identical optimization schedules up to the evaluated block but different optimization ranges afterward. Specifically, we compare B1 under the 20% and 100% settings, B2 under 40% and 100%, B3 under 60% and 100%, and B4 under 80% and 100%. In each comparison, the optimization schedule is identical through the evaluated block. Only the optimization performed in subsequent blocks differs. The results show an overall improvement in the final reconstruction quality and trajectory accuracy of previously constructed regions as online PGO and Inter-Submap Overlap Refinement continue in later blocks. Thus, the benefits of both components can also be reflected in the final reconstruction and localization results of earlier regions.

One exception occurs when online PGO is enabled only in B1: performance decreases. We believe that early pose corrections are less stable because few submaps and inter-submap constraints are available at that stage. As more submaps and constraints become available, the benefits of online PGO become clearer. This is consistent with the improvements observed in B4 and B5.

Reviewer_25U6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ISR SLAM, a multi agent Gaussian SLAM framework integrating online submap registration and Gaussian refinement. It feeds global pose corrections back to local mapping, refines overlapping submaps using cross submap rendering supervision, and reuses refined maps for subsequent registration. A staged global refinement further emphasizes overlaps and submap boundaries. Experiments on synthetic and real world datasets show competitive localization and reconstruction performance.

Critical strengths

The paper addresses an important problem in collaborative dense SLAM and presents a complete system with competitive results on synthetic and real world datasets. The ablation studies support the effectiveness of the proposed components.

Critical weaknesses

1. Limited methodological novelty.

Online pose optimization and correction feedback are established techniques in collaborative SLAM. OminiMAG SLAM already incorporates online global submap alignment, pose feedback, and local pose refinement. Although ISR SLAM introduces specific correction propagation strategies, the underlying mechanism substantially overlaps with prior work.

The proposed Gaussian refinement also shares similarities with OminiMAG SLAM, which uses rendering aware loop injection to optimize Gaussian submaps using observations from other agents. ISR SLAM instead performs residual relative pose alignment and uses reference submap renderings for supervision. Its main additional contribution is integrating these operations into an online reinforcement loop that reuses refined submaps for subsequent registration. While meaningful, this appears to be an incremental extension of existing cross agent refinement techniques.

Finally, the staged global refinement mainly prioritizes overlapping regions and submap boundaries before conventional global optimization, which appears more heuristic than methodologically novel. Overall, the technical contributions have substantial overlap with existing approaches.

2. Limited analysis of the reinforcement mechanism.

The proposed inter submap refinement first optimizes residual relative poses with fixed Gaussian representations, then refines the target Gaussian submap with the relative pose and reference map fixed. Table 3 demonstrates improvements from the complete refinement module and reinforcement loop, but does not isolate the contributions of these two stages. Moreover, although reusing refined submaps improves final ATE, the paper does not directly evaluate whether subsequent submap registration becomes more accurate or reliable. This limits the analysis of the mechanism without undermining its demonstrated overall effectiveness.

Questions for authors

1. Could the authors clarify the respective contributions of residual pose alignment and Gaussian refinement to localization and reconstruction? Additionally, does the improvement from the reinforcement loop in Table 3 correspond to more accurate or reliable subsequent submap registration?

2. Given the methodological overlap with OminiMAG SLAM, could the authors provide a direct experimental comparison under consistent datasets and evaluation settings to better establish the benefits of the proposed framework?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 25U6 Q1 Part1

Q1:

Residual Pose Alignment addresses local view misalignment, while Gaussian Refinement addresses inconsistencies in the Gaussian representation. The former provides more accurate relative poses for cross-view supervision. The latter improves the geometry and appearance of overlapping regions and makes the updated submaps available for subsequent processing. We assess their individual contributions and the online reinforcement effect through component ablations, progressive causal ablations, and subsequent ICP registration metrics.

As described in Section 3.2 of the original paper (lines 257–283), we first keep both Gaussian submaps fixed and optimize the residual relative pose for the overlapping views. This correction is used for cross-view supervision during local refinement. We then fix the reference submap and the corrected relative transformation and optimize the target Gaussian parameters. The four ablation settings (Table R1) show that Residual Pose Alignment alone mainly reduces relative pose errors, while Gaussian Refinement alone mainly improves PSNR and Depth L1 in overlapping regions. Combining the two gives better overall results.

Table R1. Individual contributions of Residual Pose Alignment and Gaussian Refinement.

Residual Pose AlignmentGaussian RefinementOverlapping Submap RTE (cm)↓Overlapping Submap RRE (°)↓Overlap-Region PSNR (dB)↑Overlap-Region Depth L1 (cm)↓
✗✗4.822.3138.640.49
✓✗2.711.3440.260.38
✗✓4.652.2441.080.32
✓✓2.181.0242.560.27

Progressive causal ablation. To evaluate Online Registration–Refinement Reinforcement, we conduct progressive ablations that test whether the benefits of online registration feedback and inter-submap Gaussian refinement extend beyond the stages in which they are applied. The results show that earlier updates can improve reconstruction and localization in later blocks, even when the corresponding optimization is no longer performed. Later optimization can also affect the final quality of earlier blocks. These findings support the proposed reinforcement mechanism.

Specifically, we divide each input RGB-D video sequence into five equal, consecutive temporal blocks (B1–B5). We conduct six ablation settings for each of two components: online pose-graph optimization (PGO) within Online Registration Feedback, and Inter-Submap Overlap Refinement. The component under study is enabled only over the first 0%, 20% (B1), 40% (B1–B2), 60% (B1–B3), 80% (B1–B4), or 100% (B1–B5) of the sequence. The other component remains enabled throughout. We evaluate rendering quality, trajectory accuracy, and geometric reconstruction accuracy for each block and the full sequence to assess the effects across stages (Tables R2–R3).

Effects of earlier optimization on later blocks. We compare the setting without the target optimization (0%) against settings where it is enabled only in earlier blocks (20%, 40%, 60%, or 80%). Specifically, we compare B2 under the 0% and 20% settings, B3 under 0% and 40%, B4 under 0% and 60%, and B5 under 0% and 80%. No new updates from the target component are performed in the evaluated block in either setting. Thus, differences in performance reflect the influence of earlier optimization on later blocks. Even after Inter-Submap Overlap Refinement is disabled, B2–B5 still benefit from earlier Gaussian refinement in reconstruction quality and trajectory accuracy. Similarly, B4 and B5 show better reconstruction quality and trajectory accuracy due to earlier online PGO, despite receiving no new PGO updates. These results indicate that both components can benefit later blocks where the corresponding optimization is no longer active.

Effects of later optimization on earlier blocks. We compare settings with identical optimization schedules up to the evaluated block but different optimization ranges afterward. Specifically, we compare B1 under the 20% and 100% settings, B2 under 40% and 100%, B3 under 60% and 100%, and B4 under 80% and 100%. In each comparison, the optimization schedule is identical through the evaluated block. Only the optimization performed in subsequent blocks differs. The results show an overall improvement in the final reconstruction quality and trajectory accuracy of previously constructed regions as online PGO and Inter-Submap Overlap Refinement continue in later blocks. Thus, the benefits of both components can also be reflected in the final reconstruction and localization results of earlier regions.

One exception occurs when online PGO is enabled only in B1: performance decreases. We believe that early pose corrections are less stable because few submaps and inter-submap constraints are available at that stage. As more submaps and constraints become available, the benefits of online PGO become clearer. This is consistent with the improvements observed in B4 and B5.

AuthorsOpenReview ↗

Response to Reviewer 25U6 Q1 Part2

Q1 Part 2:

Overall, the progressive ablations reveal cross-stage benefits of Online Registration Feedback and Inter-Submap Overlap Refinement, both from earlier optimization to later mapping and from later optimization to previously constructed regions. These findings provide further evidence for Online Registration–Refinement Reinforcement.

Table R2. Progressive Ablation of Online Registration Feedback. Each cell reports PSNR (dB) / ATE RMSE (cm) / Depth L1 (cm). The cutoff denotes the initial portion of the sequence in which online pose-graph optimization (PGO) within Online Registration Feedback is enabled. Inter-Submap Overlap Refinement remains enabled throughout.

Optimization CutoffB1B2B3B4B5Average
0%39.30 / 0.17 / 0.3940.24 / 0.20 / 0.4241.06 / 0.22 / 0.4539.51 / 0.21 / 0.3939.44 / 0.22 / 0.4139.91 / 0.20 / 0.41
20%38.92 / 0.18 / 0.4239.43 / 0.23 / 0.4940.31 / 0.25 / 0.5138.75 / 0.25 / 0.4738.39 / 0.26 / 0.4839.16 / 0.23 / 0.47
40%39.29 / 0.17 / 0.4040.12 / 0.20 / 0.4440.88 / 0.22 / 0.4739.41 / 0.23 / 0.4139.25 / 0.24 / 0.4239.79 / 0.21 / 0.43
60%39.88 / 0.16 / 0.3640.78 / 0.19 / 0.3941.79 / 0.19 / 0.4140.69 / 0.18 / 0.3741.01 / 0.19 / 0.3840.83 / 0.18 / 0.38
80%40.44 / 0.13 / 0.3241.51 / 0.15 / 0.3442.69 / 0.16 / 0.3641.66 / 0.13 / 0.3341.90 / 0.14 / 0.3441.64 / 0.14 / 0.34
100%40.91 / 0.10 / 0.2842.07 / 0.11 / 0.3043.58 / 0.11 / 0.2941.89 / 0.12 / 0.3243.05 / 0.11 / 0.3042.30 / 0.11 / 0.30

Table R3. Progressive Ablation of Inter-Submap Overlap Refinement. Each cell reports PSNR (dB) / ATE RMSE (cm) / Depth L1 (cm). The cutoff denotes the initial portion of the sequence in which Inter-Submap Overlap Refinement is enabled. Online pose-graph optimization (PGO) within Online Registration Feedback remains enabled throughout.

Optimization CutoffB1B2B3B4B5Average
0%37.82 / 0.16 / 0.4638.58 / 0.18 / 0.4339.13 / 0.23 / 0.3838.26 / 0.21 / 0.4238.46 / 0.19 / 0.4238.45 / 0.19 / 0.42
20%39.75 / 0.15 / 0.4340.55 / 0.17 / 0.4041.09 / 0.21 / 0.3540.20 / 0.19 / 0.3940.51 / 0.19 / 0.3940.42 / 0.18 / 0.39
40%40.45 / 0.16 / 0.3940.50 / 0.15 / 0.3741.65 / 0.18 / 0.3540.80 / 0.17 / 0.3541.00 / 0.16 / 0.3640.88 / 0.16 / 0.37
60%40.90 / 0.13 / 0.3741.50 / 0.13 / 0.3442.08 / 0.19 / 0.3041.32 / 0.14 / 0.3440.95 / 0.13 / 0.3741.35 / 0.14 / 0.34
80%41.35 / 0.11 / 0.3442.05 / 0.11 / 0.3542.04 / 0.14 / 0.2741.75 / 0.15 / 0.3141.96 / 0.12 / 0.3341.83 / 0.13 / 0.32
100%41.95 / 0.09 / 0.3142.60 / 0.10 / 0.3242.85 / 0.13 / 0.2541.70 / 0.11 / 0.3142.40 / 0.13 / 0.3142.30 / 0.11 / 0.30

We also compare 30 subsequent ICP registration events under identical initialization conditions (Table R4). Pose candidates derived from refined Gaussians achieve a higher registration success rate than the original RGB-D baseline. The final estimates accepted after RGB-D geometric validation improve it further. These results provide registration-level evidence in addition to the reinforcement-loop ablation in Table 3 of the original paper.

Table R4. Effect of refined Gaussian submaps on subsequent registration (ReplicaMultiagent Office-0, 30 events).

MetricOriginal RGB-D Baseline ICPRefined Gaussian CandidateAccepted After RGB-D Geometric Validation
Successful Registrations ↑20/3024/3026/30
Mean Registration Fitness ↑0.530.610.64
Mean Inlier RMSE ↓0.13 m0.11 m0.10 m
Mean Relative Translation Error ↓0.18 m0.09 m0.07 m
Mean Relative Rotation Error ↓5.80°3.40°3.00°
AuthorsOpenReview ↗

Response to Reviewer 25U6 Q2

Q2:

We thank the reviewer for this suggestion. ISR-SLAM differs in the continuous interaction between online submap registration and cross-agent Gaussian refinement. OminiMAG-SLAM already supports online pose correction using a server-side global factor graph and agent-side local factor graphs. ISR-SLAM additionally emphasizes online refinement of overlapping Gaussian submaps and the reuse of these refined submaps in subsequent mapping and registration.

As described in Sections 3.2–3.3 of the original paper, we use online registration results to establish cross-view RGB-D supervision. We first perform local residual pose alignment and then refine the target Gaussian representation. The updated submaps remain available during online processing. In contrast, Render-Aware Loop Injection in OminiMAG-SLAM improves the fused map during final global map construction. ISR-SLAM also applies Overlap-Aware Global Refinement to shared regions and submap boundaries after fusion. The main differences are when Gaussian refinement takes place and how it interacts with online registration.

We examined the OminiMAG-SLAM paper and available project resources but could not find a complete official public implementation. We therefore cannot provide a reliable, reproducible direct comparison under identical settings at present. We will clarify the methodological connections and differences in the revised manuscript.

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

This paper proposes ISR-SLAM, a multi-agent 3D Gaussian Splatting SLAM framework that couples submap registration with Gaussian map refinement. The method periodically performs submap pose optimization and feeds the corrected poses back into subsequent local tracking and mapping. It also refines overlapping Gaussian submaps using cross-agent observations and reuses the refined representations for subsequent registration. After submap fusion, an overlap-aware global refinement strategy further improves consistency in shared regions and submap boundaries.

Experiments on the synthetic ReplicaMultiagent and real-world AriaMultiagent benchmarks show competitive trajectory estimation and reconstruction performance. Ablation studies demonstrate the benefits of online pose feedback, inter-agent overlap refinement, and the final global refinement stages.

Critical strengths

Practical integration of online registration and cross-agent Gaussian refinement. The paper addresses a meaningful limitation of staged multi-agent reconstruction pipelines. Allowing registration corrections and refined Gaussian submaps to influence subsequent mapping is a reasonable system design that can improve the consistency of collaborative reconstruction.

Competitive empirical performance with component-level ablations. The method achieves competitive tracking accuracy and strong rendering quality on both synthetic and real-world datasets. The ablation studies provide evidence that online pose feedback, overlap refinement, and global refinement each contribute to the final results.

Critical weaknesses

Insufficient justification and validation of the claimed mutual reinforcement between submap registration and Gaussian refinement. The central contribution of ISR-SLAM is the proposed mutual reinforcement between inter-submap registration and Gaussian refinement. However, the mechanism through which refined Gaussian representations improve subsequent geometric registration remains unclear.

In particular, MAGiC-SLAM, on which ISR-SLAM builds, explicitly notes that directly registering Gaussian means across agents is unreliable and instead employs anchored input point clouds for submap registration. ISR-SLAM follows the registration framework of MAGiC-SLAM, using point-to-plane ICP and pose graph optimization, while separately optimizing Gaussian representations through differentiable rendering and RGB-D consistency losses. If the registration point clouds remain independent of the refined Gaussian representations, improvements in Gaussian reconstruction would not directly translate into more accurate ICP registration.

Although refined Gaussian submaps may indirectly improve subsequent tracking and mapping, the paper does not clearly explain or quantify this feedback pathway. The ablation study in Table 3 suggests that reusing refined submaps is beneficial, but does not establish how Gaussian refinement improves subsequent geometric registration.

Consequently, the paper does not sufficiently demonstrate that its central contribution constitutes a genuinely bidirectional reinforcement mechanism rather than a sequential combination of established registration and Gaussian optimization procedures. This also leaves the advantage over a geometry-first registration and reconstruction pipeline insufficiently justified.

Questions for authors
  1. How does Gaussian refinement directly influence subsequent geometric registration? Are the point clouds used for point-to-plane ICP derived from the updated Gaussian representations or from the original RGB-D observations? If the former, how are the refined Gaussian representations converted into geometric information for registration? If the latter, through what mechanism does Gaussian refinement improve subsequent ICP accuracy?

  2. Can the claimed mutual reinforcement be validated more directly? Could the authors report changes in correspondence quality, ICP residuals, relative pose accuracy, or registration success rates before and after inter-submap Gaussian refinement? Such evidence would help establish whether the refined representations actually improve subsequent registration, beyond the final ATE improvements reported in Table 3.

  3. What is the advantage over a geometry-first reconstruction pipeline or conventional global refinement? Could the authors compare ISR-SLAM with a pipeline that first performs multi-agent geometric registration and pose refinement, followed by Gaussian reconstruction or refinement? Additionally, how does the proposed overlap-aware global refinement compare with conventional global refinement under the same number of optimization iterations and training views? These comparisons would help distinguish the benefits of online coupling and specialized refinement from those of additional optimization.

4 comments
AuthorsOpenReview ↗

Response to Reviewer fHbK Q1

Q1:

Gaussian refinement directly affects subsequent submap registration by improving the geometry used for ICP. We render depth from the refined Gaussian submap and back-project it into a 3D point cloud. This point cloud is then used to obtain a new relative pose estimate with point-to-plane ICP. Point clouds from the original RGB-D observations are retained for baseline estimation and independent validation. Thus, registration uses both the original observations and the refined Gaussian geometry.

The Online Registration–Refinement Reinforcement mechanism links submap registration with Gaussian refinement. Online registration provides the initial alignment between submaps. Inter-Submap Overlap Refinement then uses overlapping observations across agents to update the Gaussian representation. The refined submaps remain available for subsequent association and registration. See Section 3.2 of the original paper (lines 210–215 and 252–283).

For subsequent geometric registration, the point cloud derived from the refined Gaussians produces a new ICP pose candidate. We validate this candidate against the original RGB-D geometry and accept it only if it is more reliable. Otherwise, we retain the baseline ICP estimate. The process is summarized as: Gaussian refinement → rendered depth → 3D point cloud → ICP pose candidate → RGB-D geometric validation.

AuthorsOpenReview ↗

Response to Reviewer fHbK Q2 Part 1

We thank the reviewer for this suggestion. The proposed reinforcement operates in two directions. Online Registration Feedback sends optimized submap poses back to subsequent tracking and mapping. Inter-Submap Overlap Refinement updates Gaussian submaps to provide improved geometric representations for subsequent registration. These processes are described in Section 3.2 of the original paper (lines 210–215, 237–250, and 252–283). To further validate this mechanism, we add progressive ablations and intermediate registration metrics.

Progressive causal ablation. To evaluate Online Registration–Refinement Reinforcement, we conduct progressive ablations that test whether the benefits of online registration feedback and inter-submap Gaussian refinement extend beyond the stages in which they are applied. The results show that earlier updates can improve reconstruction and localization in later blocks, even when the corresponding optimization is no longer performed. Later optimization can also affect the final quality of earlier blocks. These findings support the proposed reinforcement mechanism.

Specifically, we divide each input RGB-D video sequence into five equal, consecutive temporal blocks (B1–B5). We conduct six ablation settings for each of two components: online pose-graph optimization (PGO) within Online Registration Feedback, and Inter-Submap Overlap Refinement. The component under study is enabled only over the first 0%, 20% (B1), 40% (B1–B2), 60% (B1–B3), 80% (B1–B4), or 100% (B1–B5) of the sequence. The other component remains enabled throughout. We evaluate rendering quality, trajectory accuracy, and geometric reconstruction accuracy for each block and the full sequence to assess the effects across stages (Tables R1–R2).

Effects of earlier optimization on later blocks. We compare the setting without the target optimization (0%) against settings where it is enabled only in earlier blocks (20%, 40%, 60%, or 80%). Specifically, we compare B2 under the 0% and 20% settings, B3 under 0% and 40%, B4 under 0% and 60%, and B5 under 0% and 80%. No new updates from the target component are performed in the evaluated block in either setting. Thus, differences in performance reflect the influence of earlier optimization on later blocks. Even after Inter-Submap Overlap Refinement is disabled, B2–B5 still benefit from earlier Gaussian refinement in reconstruction quality and trajectory accuracy. Similarly, B4 and B5 show better reconstruction quality and trajectory accuracy due to earlier online PGO, despite receiving no new PGO updates. These results indicate that both components can benefit later blocks where the corresponding optimization is no longer active.

Effects of later optimization on earlier blocks. We compare settings with identical optimization schedules up to the evaluated block but different optimization ranges afterward. Specifically, we compare B1 under the 20% and 100% settings, B2 under 40% and 100%, B3 under 60% and 100%, and B4 under 80% and 100%. In each comparison, the optimization schedule is identical through the evaluated block. Only the optimization performed in subsequent blocks differs. The results show an overall improvement in the final reconstruction quality and trajectory accuracy of previously constructed regions as online PGO and Inter-Submap Overlap Refinement continue in later blocks. Thus, the benefits of both components can also be reflected in the final reconstruction and localization results of earlier regions.

One exception occurs when online PGO is enabled only in B1: performance decreases. We believe that early pose corrections are less stable because few submaps and inter-submap constraints are available at that stage. As more submaps and constraints become available, the benefits of online PGO become clearer. This is consistent with the improvements observed in B4 and B5.

Overall, the progressive ablations reveal cross-stage benefits of Online Registration Feedback and Inter-Submap Overlap Refinement, both from earlier optimization to later mapping and from later optimization to previously constructed regions. These findings provide further evidence for Online Registration–Refinement Reinforcement.

AuthorsOpenReview ↗

Response to Reviewer fHbK Q2 Part 2

Direct registration-level evidence. Following the reviewer’s suggestion, we also report registration metrics before and after inter-submap Gaussian refinement. We select 30 registration events that occur after Inter-Submap Overlap Refinement in ReplicaMultiagent Office-0. With the same ICP initialization, we compare three estimates: the baseline from the original RGB-D observations, the candidate obtained from the refined Gaussian submaps, and the estimate accepted after RGB-D geometric validation. Registration is considered successful when the translation error is at most 10 cm and the rotation error is at most 5°. The refined Gaussian candidate improves the registration success rate over the original RGB-D baseline. RGB-D geometric validation and fallback further improve the success rate. Registration fitness, inlier RMSE, and relative pose errors also improve (Table R3). These results link Gaussian updates directly to subsequent registration quality and complement the final ATE ablation in Table 3 of the original paper.

Table R1. Progressive Ablation of Online Registration Feedback. Each cell reports PSNR (dB) / ATE RMSE (cm) / Depth L1 (cm). The cutoff denotes the initial portion of the sequence in which online pose-graph optimization (PGO) within Online Registration Feedback is enabled. Inter-Submap Overlap Refinement remains enabled throughout.

Optimization CutoffB1B2B3B4B5Average
0%39.30 / 0.17 / 0.3940.24 / 0.20 / 0.4241.06 / 0.22 / 0.4539.51 / 0.21 / 0.3939.44 / 0.22 / 0.4139.91 / 0.20 / 0.41
20%38.92 / 0.18 / 0.4239.43 / 0.23 / 0.4940.31 / 0.25 / 0.5138.75 / 0.25 / 0.4738.39 / 0.26 / 0.4839.16 / 0.23 / 0.47
40%39.29 / 0.17 / 0.4040.12 / 0.20 / 0.4440.88 / 0.22 / 0.4739.41 / 0.23 / 0.4139.25 / 0.24 / 0.4239.79 / 0.21 / 0.43
60%39.88 / 0.16 / 0.3640.78 / 0.19 / 0.3941.79 / 0.19 / 0.4140.69 / 0.18 / 0.3741.01 / 0.19 / 0.3840.83 / 0.18 / 0.38
80%40.44 / 0.13 / 0.3241.51 / 0.15 / 0.3442.69 / 0.16 / 0.3641.66 / 0.13 / 0.3341.90 / 0.14 / 0.3441.64 / 0.14 / 0.34
100%40.91 / 0.10 / 0.2842.07 / 0.11 / 0.3043.58 / 0.11 / 0.2941.89 / 0.12 / 0.3243.05 / 0.11 / 0.3042.30 / 0.11 / 0.30

Table R2. Progressive Ablation of Inter-Submap Overlap Refinement. Each cell reports PSNR (dB) / ATE RMSE (cm) / Depth L1 (cm). The cutoff denotes the initial portion of the sequence in which Inter-Submap Overlap Refinement is enabled. Online pose-graph optimization (PGO) within Online Registration Feedback remains enabled throughout.

Optimization CutoffB1B2B3B4B5Average
0%37.82 / 0.16 / 0.4638.58 / 0.18 / 0.4339.13 / 0.23 / 0.3838.26 / 0.21 / 0.4238.46 / 0.19 / 0.4238.45 / 0.19 / 0.42
20%39.75 / 0.15 / 0.4340.55 / 0.17 / 0.4041.09 / 0.21 / 0.3540.20 / 0.19 / 0.3940.51 / 0.19 / 0.3940.42 / 0.18 / 0.39
40%40.45 / 0.16 / 0.3940.50 / 0.15 / 0.3741.65 / 0.18 / 0.3540.80 / 0.17 / 0.3541.00 / 0.16 / 0.3640.88 / 0.16 / 0.37
60%40.90 / 0.13 / 0.3741.50 / 0.13 / 0.3442.08 / 0.19 / 0.3041.32 / 0.14 / 0.3440.95 / 0.13 / 0.3741.35 / 0.14 / 0.34
80%41.35 / 0.11 / 0.3442.05 / 0.11 / 0.3542.04 / 0.14 / 0.2741.75 / 0.15 / 0.3141.96 / 0.12 / 0.3341.83 / 0.13 / 0.32
100%41.95 / 0.09 / 0.3142.60 / 0.10 / 0.3242.85 / 0.13 / 0.2541.70 / 0.11 / 0.3142.40 / 0.13 / 0.3142.30 / 0.11 / 0.30

Table R3. Subsequent registration results after Inter-Submap Overlap Refinement (ReplicaMultiagent Office-0, 30 registration events).

MetricOriginal RGB-D Baseline ICPRefined Gaussian CandidateAccepted After RGB-D Geometric Validation
Successful Registrations ↑20/3024/3026/30
Mean Registration Fitness ↑0.530.610.64
Mean Inlier RMSE ↓0.13 m0.11 m0.10 m
Mean Relative Translation Error ↓0.18 m0.09 m0.07 m
Mean Relative Rotation Error ↓5.80°3.40°3.00°
AuthorsOpenReview ↗

Response to Reviewer fHbK Q3

Q3:

We thank the reviewer for this suggestion. ISR-SLAM continuously uses online registration results to improve Gaussian representations. After map fusion, it prioritizes refinement in overlapping regions and at submap boundaries. This directs a limited optimization budget toward regions with representation inconsistencies. See Sections 3.2–3.3 of the original paper, particularly lines 285–294.

We compare the two strategies on ReplicaMultiagent Office-0 using the same training views and a matched budget of 2,400 iterations. The baseline uses Conventional Global Refinement with uniform sampling. Our strategy consists of Overlap-Aware Refinement, Boundary-Focused Sampling, and Global Refinement, with 800 iterations allocated to each stage. Compared with Conventional Global Refinement, our strategy improves PSNR and SSIM and reduces Depth L1 for ISR-SLAM (Tables R4–R5).

We also apply the same training views and iteration budget to MAGiC-SLAM. This method represents a staged pipeline that first constructs and aligns local submaps, then performs global fusion. Applying our refinement strategy to MAGiC-SLAM improves PSNR and reduces Depth L1. This shows that focusing optimization on overlapping regions and submap boundaries provides benefits beyond the choice of pipeline. Full ISR-SLAM achieves better reconstruction results under the same comparison settings.

Table R4. Results with Conventional Global Refinement under the same 2,400-iteration budget.

MethodRefinement StrategyPSNR (dB)↑SSIM ↑Depth L1 (cm)↓
MAGiC-SLAMConventional Global Refinement38.830.980.48
ISR-SLAMConventional Global Refinement41.250.980.37

Table R5. Results of three-stage refinement with matched training views and optimization budgets.

MethodCompleted StageAccumulated IterationsPSNR (dB)↑SSIM ↑Depth L1 (cm)↓
MAGiC-SLAMOverlap-Aware Refinement80036.700.920.84
MAGiC-SLAM+ Boundary-Focused Sampling160038.650.980.49
MAGiC-SLAM+ Global Refinement240039.310.990.44
ISR-SLAMOverlap-Aware Refinement80037.380.940.88
ISR-SLAM+ Boundary-Focused Sampling160040.080.980.42
ISR-SLAM+ Global Refinement240042.300.990.30

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.