acceptodds
Withdrawn from ICLR 2027

MIRAGE: Metric Inference and Reliability-Aware Gaussian Estimation from Feed-Forward Vision

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Feed-forward 3D priors enable high-quality reconstruction from image streams but often lack reliable metric scale and suffer from long-term trajectory drift. In contrast, LiDAR–inertial estimation provides metrically accurate motion but lacks the image-space consistency needed for high-fidelity reconstruction. We propose **MIRAGE**, which anchors feed-forward visual geometry to LiDAR–IMU metric motion and regulates online pose corrections according to correspondence support and cross-modal agreement, jointly enabling accurate metric pose estimation and high-fidelity Gaussian reconstruction. Experiments on Oxford Spires, M2DGR and INS show that MIRAGE achieves accurate metric trajectory estimation, outperforming COLMAP, a widely used but computationally expensive reference for vision-only benchmarks. Motivated by the scarcity of long-horizon monocular benchmarks for large-scale indoor reconstruction, we further build a multimodal data-collection platform and use MIRAGE to establish accurate 6-DoF reference trajectories for a new vision-only benchmark. We benchmark recent monocular large-scale reconstruction methods on this dataset, providing a systematic evaluation of their performance in long-horizon indoor scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Reconstruction, SLAM, Novel View Synthesis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_huEBOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript presents MIRAGE, a visual–LiDAR–inertial mapping framework that uses mesh-Gaussian scene representation and achieves centimeter-level accuracy in sensor-rig trajectory estimation and high-quality 3D reconstruction. Built upon Voxel-SLAM, MIRAGE incorporates MASt3R and depth priors from Depth Anything into its odometry pipeline and extends the system with 3DGS-based reconstruction. The manuscript also introduces a synchronized RGB–LiDAR–IMU benchmark dataset featuring full 6-DoF ground-truth trajectories over long traversals in object-rich indoor environments.

Critical strengths
  1. A unified pipeline that integrates RGB, LiDAR, and IMU data for high-quality mesh-Gaussian reconstruction.
  2. A novel dataset featuring synchronized RGB–LiDAR–IMU data and full 6-DoF ground-truth trajectories over long traversals in object-rich indoor environments.
Critical weaknesses
  1. My primary concern is that this work appears to be an ad hoc improvement and 3DGS extension of Voxel-SLAM. While the overall goal is to achieve large-scale, high-fidelity 3D Gaussian Splatting reconstruction, the manuscript lacks a clear formulation of the underlying research problems, making the contributions appear largely engineering-driven.

  2. Despite 3D Gaussian Splatting being a central focus of the manuscript, it receives insufficient discussion in the Introduction and Related Work sections, whereas the Contributions and Method sections explicitly emphasize its treatment. This further obscures the main focus and contributions of the work.

  3. To improve localization accuracy using visual information, the front-end employs MASt3R correspondences and LiDAR depth maps completed using Depth Anything priors for camera pose estimation. This introduces substantial computational overhead, while Table 1 shows only limited improvements in trajectory accuracy over Voxel-SLAM.

  4. Although the proposed dataset is presented as one of the main contributions, its comparison with existing datasets is limited to discussions in the Introduction and Related Work sections. A more systematic comparison is missing.

Questions for authors
  1. How does the proposed method perform without loop closure? In Table 1, Fast-LIVO2 is an odometry method that inevitably accumulates trajectory drift, whereas the proposed method is a SLAM system with loop closure detection and optimization to mitigate such drift. An ablation study without loop closure would help clarify the source of the accuracy improvements.

  2. Why is Voxel-SLAM categorized as an offline method in Table 1? The original Voxel-SLAM paper reports real-time performance.

  3. How does the proposed dataset differ from existing datasets? I suggest including a comparison table in the revised manuscript to clearly highlight the distinguishing characteristics and advantages of the proposed dataset.

  4. What is the runtime performance of the proposed system?

1 comment
AuthorsOpenReview ↗

Response to R3 — huEB

Thank you for the questions on the research focus, loop closure, and dataset scope.

Research problem and contribution.

  • Contribution and motivation. To our knowledge, MIRAGE is the first loosely coupled streaming pose estimation system combining feed-forward visual geometry with LiDAR–inertial estimation for accurate metric tracking and high-quality Gaussian reconstruction. Visual constraints improve rendering alignment, while less metrically reliable updates can perturb shared navigation states. We control visual influence on metric estimates while reducing local reprojection error.
  • Design. We use fixed metric anchors, inter-anchor visual motion fusion, and gated global corrections (Sec. 3.2). The Introduction and Sec. 2.1–2.2 discuss this problem and Gaussian mapping.

No-loop results and visual utility.

  • No-loop ablation. Table 7 defines Local as the no-loop trajectory and Global as the corrected trajectory. Mean ATE decreases from 7.784 to 4.529 cm on Oxford Spires and 7.055 to 5.195 cm on INS (Table 3(a)). With offline HBA, we obtain 2.96 cm mean ATE versus 3.31 cm for Voxel-SLAM.
  • Reconstruction benefit. Removing the visual pose/geometry branch reduces M2DGR PSNR from 23.28/24.41/24.74 to 13.90/13.04/12.89 dB (Table 3(b)), demonstrating its reconstruction contribution.

Optimization stages and runtime.

  • Voxel-SLAM classification. The offline category refers to its HBA-refined trajectory output; its frontend operates in real time.
  • MIRAGE runtime. Table 1 reports a 0.73 s/frame pose-optimization latency. On RTX 4090, our 15k-step offline refinement takes 26.77/3.47/10.48 minutes on average for Oxford Spires/M2DGR/INS, respectively. Gaussian growth is capped at 25%.

Benchmark.

  • Benchmark value. Indoor multimodal benchmarks remain scarce compared with extensive outdoor datasets. To our knowledge, ours is the first benchmark for long-horizon reconstruction in complex, object-rich indoor scenes with synchronized RGB–LiDAR–IMU data and full 6-DoF reference trajectories. We cover laboratories, dining areas, offices, and corridors, with ten sequences spanning 1.15 km, for joint evaluation of metric localization and high-fidelity reconstruction (Sec. 3.4; App. A.4).
  • Reference consistency. With identical ground-normal alignment and combined first/last-100-pose windows, our mean Z range is 1.93 cm, versus 7.22 cm for INS, which also uses a wheeled platform indoors and returns to the starting area. This demonstrates stronger endpoint vertical consistency.
Reviewer_yhQ1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

MIRAGE integrates feed-forward visual geometry with LiDAR–inertial estimation for metric camera tracking and Gaussian reconstruction. The pipeline anchors visual poses to a metric reference, selects loop corrections based on reliability, and subsequently refines the Gaussian representation while keeping camera poses fixed. LiDAR evidence guides Gaussian splitting, replacement, and insertion. The paper also introduces ten synchronized indoor sequences with reference trajectories generated by MIRAGE. Evaluation covers trajectory accuracy and rendering quality on Oxford Spires, M2DGR, and INS, alongside component ablations and assessments of vision-only methods on the proposed benchmark.

Critical strengths
  1. The system completes all 15 evaluated sequences, achieving a mean online translational ATE of 4.22 cm and the lowest reported online ATE on 12 sequences. These results support the integrated pipeline’s ability to produce accurate trajectories across varied environments, although the contributions of its individual innovations remain insufficiently established.

  2. The synchronized indoor sequences provide calibration, sensor streams, reference maps, and estimated trajectories. Long traversals through cluttered rooms and low-texture corridors offer useful settings for assessing accumulated drift and reconstruction consistency. The dataset could therefore make a practical contribution, provided the accuracy of its reference trajectories is adequately validated.

Critical weaknesses
  1. Several core ingredients have close precedents. LiDAR-VGGT [1] combines LiDAR–inertial odometry with a visual foundation model for metric-scale, globally consistent mapping. RMGS-SLAM [2] integrates LiDAR–inertial–visual Gaussian SLAM with feed-forward predictions, geometric priors, and loop closure. PLANING [3] introduces the triangle–Gaussian representation adopted by MIRAGE, while LI-GS [4] uses LiDAR-derived geometric constraints for Gaussian initialization and optimization. These precedents do not imply that MIRAGE’s exact design is redundant, but they make a clear demonstration of its additional contribution essential.

    The current experiments do not provide that demonstration. Although the reliability gate is central to the claimed novelty, Appendix A.2.3 does not specify the discrete correction-gain set or provide operational definitions and thresholds for the acceptance criteria. The stage-wise ablation compares tracking, local optimization, and global correction, but does not isolate reliability gating from standard loop correction or distinguish the effects of translation and rotation gating. Likewise, the refinement comparison changes primitive allocation and loss terms simultaneously: the photometric baseline maintains a fixed primitive count, while the proposed method can split, replace, and insert primitives. The reported 0.44 dB gain therefore cannot be attributed specifically to the LiDAR-guided repair decisions. Overall, the results demonstrate the effectiveness of the integrated pipeline more clearly than a substantive methodological advance.

  2. Gaussian-LIC2 is a relevant baseline, but direct competitors such as RMGS-SLAM [2] and GS-LIVO [5] are missing from the quantitative comparisons. Both address LiDAR–inertial–visual Gaussian mapping and are therefore important references for assessing the claimed combination of localization accuracy and reconstruction quality.

    The rendering comparisons also differ materially in their optimization and evaluation protocols. Appendix A.5 states that MIRAGE’s results include offline Gaussian refinement, whereas Gaussian-LIC2 is evaluated using its online maps. On the new benchmark, MIRAGE’s “Stream PSNR” is computed on held-out images that differ from the shared test set used for the visual baselines. These comparisons therefore do not establish an advantage under matched conditions.

    Finally, the benchmark adopts MIRAGE’s estimated trajectories as references without an independent quantitative assessment of their accuracy on the newly collected sequences. Strong performance on existing datasets does not establish reference accuracy for these particular recordings. This is especially important because the benchmark provides purported full 6-DoF reference trajectories, while the reported trajectory validation primarily concerns translational ATE. Together, these limitations leave both the comparative advantage and a central dataset contribution insufficiently supported, warranting rejection in the current form.

[1] Lijie Wang et al. LiDAR-VGGT: Cross-Modal Coarse-to-Fine Fusion for Globally Consistent and Metric-Scale Dense Mapping. arXiv:2511.01186, 2025.

[2] Dongen Li et al. RMGS-SLAM: Real-time Multi-sensor Gaussian Splatting SLAM. arXiv:2604.12942, 2026.

[3] Changjian Jiang et al. PLANING: A Loosely Coupled Triangle-Gaussian Framework for Streaming 3D Reconstruction. arXiv:2601.22046, 2026.

[4] Changjian Jiang et al. LI-GS: Gaussian Splatting with LiDAR Incorporated for Accurate Large-Scale Reconstruction. arXiv:2409.12899, 2024.

[5] Sheng Hong et al. GS-LIVO: Real-Time LiDAR, Inertial, and Visual Multi-sensor Fused Odometry with Gaussian Mapping. arXiv:2501.08672, 2025.

Questions for authors
  1. What correction-gain candidates, scoring functions, support requirements, and acceptance thresholds are used in Appendix A.2.3? Are these settings fixed across datasets? Can you isolate the gate’s contribution by comparing ungated loop correction, translation-only gating, and the complete gate with the same frontend and loop proposals?

  2. What evidence supports the claimed advantage over closely related systems such as RMGS-SLAM [2] and GS-LIVO [5]? If direct comparisons are unavailable, what specific obstacles prevent them? Does the rendering advantage persist with comparable refinement budgets and identical test images?

  3. Does geometry-aware refinement outperform a photometric baseline that also supports adaptive densification, under comparable primitive counts and optimization budgets? This comparison would help determine whether the gains are specifically attributable to the proposed geometric repair criteria.

  4. What independent evidence validates the translational and rotational accuracy of the reference trajectories for the ten newly collected sequences? Which trajectory variant is used to produce these references, and how does its estimated uncertainty compare with the performance differences between the evaluated methods?

[1] Lijie Wang et al. LiDAR-VGGT: Cross-Modal Coarse-to-Fine Fusion for Globally Consistent and Metric-Scale Dense Mapping. arXiv:2511.01186, 2025.

[2] Dongen Li et al. RMGS-SLAM: Real-time Multi-sensor Gaussian Splatting SLAM. arXiv:2604.12942, 2026.

[3] Changjian Jiang et al. PLANING: A Loosely Coupled Triangle-Gaussian Framework for Streaming 3D Reconstruction. arXiv:2601.22046, 2026.

[4] Changjian Jiang et al. LI-GS: Gaussian Splatting with LiDAR Incorporated for Accurate Large-Scale Reconstruction. arXiv:2409.12899, 2024.

[5] Sheng Hong et al. GS-LIVO: Real-Time LiDAR, Inertial, and Visual Multi-sensor Fused Odometry with Gaussian Mapping. arXiv:2501.08672, 2025.

1 comment
AuthorsOpenReview ↗

Response to R2 — yhQ1

Thank you for the detailed questions.

Why loose coupling?

  • Motivation. Visual alignment improves rendering, while less metrically reliable visual updates can perturb shared navigation states. Gaussian-LIC2 fails on Observatory 02/Church 03 and reaches 3.08 m ATE on Keble 02, versus our 3.60 cm (Table 1).
  • Contribution. To our knowledge, MIRAGE is the first loosely coupled streaming pose estimation system combining feed-forward visual geometry with LiDAR–inertial estimation for accurate metric tracking and high-quality Gaussian reconstruction. We control visual influence through fixed metric anchors, inter-anchor visual motion (Eq. (2)), and gated corrections.
  • Novelty. RMGS-SLAM tightly couples estimation. PLANING lacks LiDAR–IMU metric anchoring. LiDAR-VGGT and LI-GS perform post-processing without online pose estimation. Unlike these algorithm-focused works, we also designed and built a wheeled data-collection platform and used it with MIRAGE to establish, to our knowledge, the first benchmark for long-horizon reconstruction in complex, object-rich indoor scenes with synchronized RGB–LiDAR–IMU data and full 6-DoF reference trajectories.

Gate evidence and configuration.

  • Evidence. Table 7 shows improvements on all eight Oxford Spires sequences and a 56.9% ATE reduction on INS sq1. Across 92 events with identical current trajectories, proposals, and anchors, gating achieves 58 wins, 29 losses, five ties against full correction in event-time anchor ATE; GT is used post hoc. The 14 accepted corrections comprise 12 translation-only and two involving rotation; eight use partial translation gains.
  • Settings. Candidates, scores, support requirements, and thresholds are identical across all datasets and sequences, with no scene-specific tuning. Translation/rotation gains α/β use {0, 0.25, 0.5, 0.75, 1}, testing translation-only first. Selection and validation subsets both require revisit and local support. Visual acceptance balances reprojection gains against local regression. Geometric benefit is the relative reduction in point-to-plane residuals over fixed LiDAR correspondences, compared with the currently accepted trajectory. Geometric fallback requires positive net gains on both subsets and conservative benefit exceeding degradation risk; rotation must add geometric benefit over the same translation-only candidate (App. A.2.3).

Geometry-aware refinement. Both variants share initialization, 15k steps, poses, exposure, renderer, and test views. Our repairs and geometric regularization yield a 2.12 dB mean PSNR gain over streaming, including 0.44 dB beyond fixed-count photometric refinement, with better SSIM/LPIPS on all three sequences; Gaussian growth is capped at 25% (Table 3(b)).

Baselines and evaluation.

  • Availability. We found no public RMGS-SLAM implementation; GS-LIVO reconstruction failed on long Oxford Spires sequences in our tests.
  • Gaussian-LIC2. Its official repository lists offline post-optimization as TODO. Refining its maps with the photometric/depth losses specified in the Gaussian-LIC2 paper yields mean changes of +0.106 dB PSNR, +0.00195 SSIM, −0.00233 LPIPS across 12 scenes. MIRAGE retains a 3.26 dB mean PSNR advantage, with higher PSNR on 10/12 scenes.
  • Protocol. Table 4 ranks common-frame results using shared images and common downstream 3DGS; Stream PSNR separately evaluates the full pipeline's held-out images.

Benchmark references and utility. With identical ground-normal alignment and combined first/last-100-pose windows, our mean Z range is 1.93 cm, versus 7.22 cm for INS, which also uses a wheeled platform indoors and returns to the starting area. This demonstrates stronger endpoint vertical consistency. Our poses support 24.06–30.89 dB Stream PSNR, demonstrating reconstruction utility and multi-view consistency.

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

This paper presents MIRAGE, a multimodal framework for accurate metric camera pose estimation and Gaussian reconstruction. The method combines feed-forward visual correspondence priors with LiDAR–IMU motion estimation. It introduces a reliability-aware correction mechanism to selectively incorporate visual information while preserving metric trajectory consistency. The estimated poses are then used for Gaussian mapping, followed by geometry-aware refinement with LiDAR guidance. The authors also introduce a long-horizon indoor reconstruction benchmark with MIRAGE-generated reference trajectories.

Critical strengths
  • Strong trajectory estimation results. The method is evaluated on Oxford Spires, M2DGR, and INS, achieving an average online ATE of 4.22 cm compared with 7.49 cm for EllipseLIO. The results demonstrate the effectiveness of the overall system for accurate metric pose estimation.

  • Good reconstruction quality. The proposed method achieves better rendering quality than Gaussian-LIC2 on most evaluated sequences. The geometry-aware refinement also improves over photometric-only refinement under the same optimization budget.

  • A potentially useful benchmark. The newly collected dataset contains 10 long-horizon indoor sequences with synchronized RGB, LiDAR, and IMU measurements. It could provide a useful testbed for evaluating long-range monocular reconstruction.

Critical weaknesses

Critical Weaknesses

  1. The contribution of the proposed reliability-aware trajectory correction is not sufficiently demonstrated. This is one of the central methodological contributions of the paper, but the ablation in Table 3(a) only compares different stages of the overall system. Most of the trajectory improvement comes from local optimization, while global correction provides relatively limited additional gains and even slightly degrades performance on M2DGR. Without a direct comparison against standard loop correction without reliability gating, it is difficult to determine whether the proposed mechanism is actually necessary. This also makes the methodological novelty less convincing, as much of the framework builds upon existing visual matching, LiDAR–inertial optimization, and Gaussian mapping techniques.

  2. The reliability of the proposed benchmark's reference trajectories is not independently validated. The authors use trajectories estimated by MIRAGE itself as reference poses for evaluating other reconstruction methods. Although MIRAGE achieves strong trajectory accuracy on existing benchmarks, this does not necessarily establish the accuracy of its estimates on the newly collected sequences. Since providing accurate reference trajectories is a major contribution of the paper, the lack of independent validation raises concerns about the reliability of the resulting benchmark and the conclusions drawn from it.

Questions for authors
  • How much does reliability-aware gating actually contribute? Could the authors compare the proposed correction mechanism against accepting all loop corrections, fixed-weight correction, and correction based only on geometric or visual consistency? It would also be helpful to report how frequently proposed corrections are accepted or rejected.

  • How much of the final improvement comes from visual pose refinement rather than LiDAR–IMU estimation? Could the authors evaluate the same Gaussian mapping pipeline using LiDAR–IMU poses, locally optimized poses, and globally corrected poses, while keeping all reconstruction settings fixed?

  • What is the individual contribution of each Gaussian repair operation? Could the authors separately ablate splitting, replacement, and insertion, and report their effects on reconstruction quality and Gaussian count?

  • How reliable are the reference trajectories in the new benchmark? Is there any independent validation of their accuracy, such as surveyed reference measurements or comparisons against another independently optimized trajectory? Also, what is the overall computational cost when streaming reconstruction and offline refinement are both included?

1 comment
AuthorsOpenReview ↗

Response to R1 — TkzH

Thank you for recognizing MIRAGE's accuracy, reconstruction quality, and dataset utility. To our knowledge, MIRAGE is the first loosely coupled streaming pose estimation system combining feed-forward visual geometry with LiDAR–inertial estimation for accurate metric tracking and high-quality Gaussian reconstruction.

Reliability-aware correction.

  • Trajectory gains. Table 7 compares no-loop Local and corrected Global trajectories. All eight Oxford Spires sequences improve, including 69.3%/77.4% ATE reductions on Blenheim01/02; INS sq1 improves by 56.9%. Across 15 sequences, the largest ATE increase is 2.53 mm. No loop corrections were accepted on M2DGR.
  • Direct comparison. Across 92 events with identical current trajectories, proposals, and anchors, gating yields 58 wins, 29 losses, and five ties against full correction in event-time anchor ATE. GT is used only post hoc. On O1, gating preserves 3.38 cm, versus 7.00 cm with full correction.
  • Acceptance statistics. We accept 14 corrections: 12 translation-only and two involving rotation; eight use partial translation gains. The other 78 lack sufficient evidence through sequence end and leave the trajectory unchanged.

Visual refinement and Gaussian repair.

  • Visual contribution. Removing the visual pose/geometry branch while retaining RGB mapping supervision reduces M2DGR PSNR from 23.28/24.41/24.74 to 13.90/13.04/12.89 dB (Table 3(b)).
  • Refinement contribution. Splitting repairs oversized primitives, replacement corrects depth disagreement, and insertion fills coverage gaps. With geometric regularization, these improve streaming-map PSNR by 2.12 dB on average, including 0.44 dB beyond photometric-only refinement (Table 3(b); App. A.3.2). SSIM and LPIPS also improve on all three sequences.

Computation.

  • Pose optimization and refinement. Table 1 reports a 0.73 s/frame pose-optimization latency. On RTX 4090, our 15k-step offline refinement takes 26.77/3.47/10.48 minutes on average for Oxford Spires/M2DGR/INS, respectively. Gaussian growth is capped at 25%.

Benchmark value and reference consistency.

  • Benchmark value. Indoor multimodal benchmarks remain scarce compared with extensive outdoor datasets. To our knowledge, ours is the first benchmark for long-horizon reconstruction in complex, object-rich indoor scenes with synchronized RGB–LiDAR–IMU data and full 6-DoF reference trajectories.
  • Reference consistency. With identical ground-normal alignment and combined first/last-100-pose windows, our mean Z range is 1.93 cm, versus 7.22 cm for INS, which also uses a wheeled platform indoors and returns to the starting area. This demonstrates stronger endpoint vertical consistency.
  • Reconstruction utility. Our poses support 24.06–30.89 dB Stream PSNR across ten scenes, demonstrating multi-view consistency and reconstruction utility.

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

Related papers

Discussion (0)

Sign in to comment.