acceptodds
Desk rejected at ICLR 2027

-SLAM: Omega-Native Loop Closure for Uncalibrated Dynamic SLAM

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

We propose Ω-SLAM, a training-free dense RGB SLAM system built on VGGT-Ω for uncalibrated monocular videos. Extending feed-forward reconstruction to long sequences requires loop constraints that remain reliable under perceptual aliasing and scene motion. We address this challenge by using VGGT-Ω’s own representations for both retrieval and geometric verification, eliminating a separate place-recognition network. Pre-attention patch embeddings retrieve candidate frames, while compact scene registers provide submap context for reranking. Jointly predicted cameras and depths then support a depth-conditioned rigid reprojection test on mutual patch correspondences, checking geometric consistency beyond epipolar agreement. Verified loops are integrated as quality-weighted robust factors in an SL(4) graph, with multi-anchor calibration bridges stabilizing submap alignment and geometry-based observation weights reducing the influence of moving content. Retrieval state uses 96 KiB for 16 frames, while a fixed set of active submaps bounds resident dense-map memory. Experiments reduce mean absolute trajectory error relative to VGGT-SLAM 2.0 by 36.5% on 7-Scenes and 47.2% across Bonn and TUM dynamic sequences, demonstrating improved trajectory accuracy without additional training or an external retrieval network.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
SLAM, Loop Closure, Feed-forward Reconstruction, Uncalibrated Cameras, Dynamic Scenes

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bvuCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents omega-SLAM, a training-free dense SLAM system built upon VGGT-omega. It reuses the internal representations of VGGT-omega for loop retrieval and geometric verification, and integrates multi-anchor submap alignment and dynamic observation weighting into an SL(4) factor graph. The method reports improved trajectory accuracy over VGGT-SLAM 2.0 on both static and dynamic benchmarks.

Critical strengths
  • The paper is clearly written and the overall pipeline is easy to follow.

  • The proposed system achieves promising results on several benchmarks without additional training or an external place-recognition network.

Critical weaknesses
  • Limited methodological novelty. The overall framework closely follows VGGT-SLAM and VGGT-SLAM 2.0, including submap reconstruction, alignment, loop closure, and SL(4) factor graph optimization. The main differences are the use of the stronger VGGT-omega backbone and several incremental modifications to retrieval, geometric verification, and observation weighting. These components largely rely on existing techniques, and I find the technical contributions rather limited. The overall pipeline is also conceptually similar to VGGT-Long. Overall, I do not see significant contribution beyond existing VGGT-like-based SLAMs.

  • Insufficient comparison with recent dynamic SLAM methods. Although the paper emphasizes robustness in dynamic environments, the evaluation omits relevant recent methods such as DROID-W and MegaSam. The dynamic handling itself mainly relies on geometric consistency-based weighting rather than a substantially new dynamic-scene modeling approach. Without comparisons to recent dymaic SLAMs, the claimed advantages for dynamic SLAM cannot be supported.

  • In ablation, it's shown the most contributed components are all the necessary ones from existing VGGT-like-based SLAMs (e.g. the used loop closure), while no sufficient original contibutions introduced. This weakened the value of this work.

Overall, I'm tending giving a typical weak reject below borderline reject. But as the granularity of this ICLR does not support it, I'll recommend reject.

Questions for authors

See the weaknesses.

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

This paper presents Ω-SLAM, a training-free dense monocular SLAM system built on VGGT-Ω for uncalibrated videos. It reuses pre-attention frame descriptors and scene registers for loop retrieval, verifies candidates using patch correspondences and depth-conditioned rigid reprojection, and integrates accepted loops into a quality-weighted SL(4) factor graph. Multi-anchor alignment and rigid-static observation weighting further improve submap consistency. Experiments demonstrate the effectiveness of the proposed system in both static and dynamic scenes.

Critical strengths
  1. I appreciate the authors’ loop-closure design, which addresses the important problem of unifying place recognition and loop closure using the reconstruction model’s own representations, without a separate place-recognition network.
  2. The method achieves strong trajectory accuracy on 7-Scenes, Bonn, and TUM-Dynamic, showing improvements over VGGT-SLAM 2.0 in both static and dynamic environments.
  3. The paper is clearly written and well presented.
Critical weaknesses
  1. Limited qualitative evidence for loop closure and scalability. Figure 2 does not clearly demonstrate how the proposed method differs from VGGT-SLAM in loop-closure behavior or correction quality. Furthermore, the visualizations focus on local indoor scenes. Results from larger environments and longer trajectories would better demonstrate global consistency and long-horizon mapping capabilities.
  2. Limited reconstruction evaluation. Among SLAM systems, Table 11 compares reconstruction quality only against VGGT-SLAM. Comparisons with additional SLAM baselines would better establish the method’s dense mapping performance.
  3. Insufficient coverage of recent methods. Recent relevant approaches such as LingBot-Map and SLAM-Former are absent from the main comparisons and insufficiently discussed throughout the paper. A discussion of their differences and, where feasible, quantitative or qualitative comparisons would better position the contribution.
Questions for authors
  1. Could the authors provide additional visualizations demonstrating loop-closure correction and mapping in larger-scale environments? These would help substantiate the method’s benefits for global consistency and long-horizon mapping.
  2. Could the authors extend the reconstruction evaluation to include more SLAM baselines, particularly VGGT-SLAM 2.0?
  3. I would be glad to see a more detailed discussion of related methods, particularly LingBot-Map and SLAM-Former, and experimental comparisons if possible. These would help clarify the contribution relative to recent advances.
Reviewer_jEK4OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Ω-SLAM, a training-free RGB SLAM system built around VGGT-Ω. Without retraining VGGT-Ω, Ω-SLAM introduces a mechanism to add loop closure to the system, thereby reducing mean ATE on benchmarks such as 7-Scenes, Bonn, and TUM dynamic.

Critical strengths
  • Ω-SLAM has good and well-motivated design. Specifically, it is able to avoid retraining VGGT-Ω by smartly reusing and condensing its learned representations. Examples include the topology tokens based on the scene registers after attention, and mean-pooled pre-attention image patch embeddings.
  • For the most part, the writing is clear and well-motivated.
Critical weaknesses
  • In Tab. 3, the precision and recall values for loop verification precision seem relatively low across the board. In particular, this suggests that Ω-SLAM may be susceptible to false positive loop closures, which can be catastrophic.
  • For the section on geometric verification gate, the actual acceptance condition for the top candidate image pair does not seem to be written out. The two types of tests (epipolar constraint, reprojection error) are well described, but I was not able to figure out how these two separate metrics / tests were used together.
Questions for authors
  • Clarification request: Can the authors confirm or discuss further whether false positives are a major concern for Ω-SLAM’s performance? Is its typical failure mode different than previous VGGT-based works without loop closure, and we are trading one kind of problem for another? If false positives are a major concern, why does ATE remain relatively low in evaluation?
  • How does Ω-SLAM perform on SLAM benchmarks with low textures / repetitive textures that pose a challenge for accurate loop closure? Evaluating on a benchmark specializing in this, such as GV-Bench, MulRan, or alternatives, could help further strengthen evaluation.
  • Clarification request: what is the explicit acceptance condition for the geometric verification gate, based on the two types of tests mentioned?

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

Related papers

Discussion (0)

Sign in to comment.