acceptodds
Withdrawn from ICLR 2027

ReCall-SLAM: Long-Horizon RGB-D SLAM through Recurring Local Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Depth provides RGB-D SLAM with explicit geometry and metric scale, yet brief blur, texturelessness, weak overlap, or corrupted depth can still break an otherwise long trajectory. This exposes a weakness in local multi-view geometry. More broadly, this local capability is required not only by the frontend but throughout SLAM. RGB-D SLAM can be viewed as repeatedly solving the same local problem at increasing map scales: nearby views recover camera motion and local structure, wider observations reconcile geometry over a region, and views around a place revisited much later relate distant parts of the trajectory. Each task jointly recovers cameras and scene structure from a bounded set of spatially related RGB-D observations, then applies the estimate at the appropriate map scale. This suggests that a long-horizon system should make a strong local solver available across its hierarchy. Based on this insight, we present ReCall-SLAM, a training-free system that retains observation history and repeatedly poses these bounded reconstruction problems to one pretrained Pi3X model. Every model call natively consumes RGB, calibrated depth, and camera intrinsics and predicts cameras, dense geometry, and confidence. Bounded regional graphs and a compact global graph integrate these predictions into a persistent, revisable map while all levels advance asynchronously. Experiments demonstrate stable, coherent reconstruction on sequences extending to tens of thousands of frames; ReCall-SLAM also improves the FastCaMo-Real F-score by 20.5 percentage points and reduces Oxford Spires ATE by 58.3%. Project page: https://anonymous-submission-x.github.io/ReCallSLAM/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
SLAM, Feedforward Reconstruction Model

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

ReCall-SLAM builds a long-horizon RGB-D SLAM system around a fixed, RGB-D-conditioned Pi3X model. It applies the model to overlapping local windows for tracking, to wider sets of views for corrections within submaps, and to views from revisited places for constraints between submaps. The predictions are incorporated into local and global pose graphs without additional model training. The paper evaluates trajectories on TUM RGB-D, BundleFusion, and Oxford Spires, reconstruction on BundleFusion and FastCaMo-Real, and several long sequences qualitatively. It reports complete trajectories on its 11 selected Oxford recordings and a substantial FastCaMo-Real reconstruction gain. Ablations on TUM RGB-D show that the two higher levels improve on local processing alone.

Critical strengths
  1. The idea of using the same RGB-D geometry solver for tracking, longer-span correction, and revisits is a coherent system design. The TUM ablation gives some evidence that the higher levels contribute beyond the local solver: mean ATE falls from 5.583 cm with Level 0 alone to 2.969 cm with the full system.

  2. The difficult-sequence results are worth taking seriously. ReCall-SLAM completes all 11 selected Oxford recordings, and its FastCaMo-Real F-score is 66.82%, compared with 46.29% for PROFusion. These results support the value of the system, although they do not establish a similarly large gain on every benchmark.

Critical weaknesses
  1. The comparison does not yet establish how ReCall-SLAM fares against existing systems designed for large-scale RGB-D mapping. On the complete Oxford recordings, ORB-SLAM3 RGB-D does not reach the final frame, while the other RGB-D trajectory baselines are two RGB-D adaptations of VGGT-SLAM 2.0 constructed for this paper. A run of an established long-term system such as RTAB-Map, restricted to the same rectified images and camera-aligned depth, would be more informative. On large scenes with native dense RGB-D input, RemixFusion would also be relevant. The RGB-only results provide context, but their different input and Sim(3) alignment should remain separate from claims about RGB-D performance. Since RGB-only methods are already included, HorizonStream is a reasonable additional comparison under a matched protocol, though it is less important than a modality-matched baseline.

  2. The quantitative evidence for performance at large spatial scale comes mainly from Oxford Spires. The longer indoor sequences are shown qualitatively, so it is difficult to tell whether the Oxford result extends to a different large environment and depth sensor. A second public large-scene benchmark with quantitative trajectory and map evaluation would materially strengthen the central claim. This matters because the improvement on TUM is small—2.969 cm versus 3.165 cm for the strongest listed competitor—and ReCall-SLAM is slightly behind that competitor on BundleFusion trajectory ATE, 6.806 cm versus 6.504 cm. The TUM ablation also does not show how much each level contributes on the long sequences where the method’s advantage is most important.

Questions for authors
  1. Could you add quantitative results on a second public large-scene dataset with at least one native RGB-D baseline? Complete-sequence trajectory error, coverage, reconstruction quality, and a Level 0/1/2 ablation there would help determine whether the long-horizon gains generalize beyond Oxford.

  2. Given the existing RGB-only comparisons, could you evaluate HorizonStream and SLAMFormer-∞ on the same selected frames and trajectory protocol, reporting it in a separate RGB-only group and stating whether its optional loop closure is enabled? If that comparison is infeasible, please explain the obstacle.

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

This is a system paper that presents RECALL-SLAM, an RGB-D SLAM system that uses a pretrained feed-forward geometry model (Pi3X) as a 3D estimator across three hierarchical levels: local reconstruction (Level 0), wide-span correction (Level 1), and submap-level revisit detection and global correction (Level 2).

The central idea is that the off-the-shelf Pi3X can serve all scales of the hierarchy above, and provide relative reconstruction as constraints for a following graph optimization. Experiments on several datasets demonstrate competitive trajectory accuracy and reconstruction quality, with notable success in reconstructing several failure cases of existing methods.

While the system design is overall reasonable and the results appear competitive, the work offers limited insight beyond system integration. It also suffers from writing issues and insufficient validation. Therefore, I believe it requires a major revision.

Critical strengths
  1. Effective integration of feed-forward geometry model and graph optimization. I appreciate the engineering effort required to integrate Pi3X consistently at the local, context, and submap levels and to couple its predictions with local and global graph optimization. Although the overall design is straightforward, implementation takes effort. The proposed RECALL-SLAM system appears effective across the tested indoor and outdoor datasets, and the ablation supports the usefulness of the individual hierarchy levels.

  2. Strong long-sequence results. As shown in the figures and the video, RECALL-SLAM produces competitive reconstruction results. It also handles several failure cases of existing methods.

Critical weaknesses
  1. Limited novelty and insight. Local and global bundle adjustment or graph optimization are standard in conventional SLAM, and prior works such as MASt3R-SLAM and VGGT-SLAM already use feed-forward models in a similar spirit. This paper essentially replaces the model with a stronger one, Pi3X, and exploits its RGB-D and calibration conditioning. Therefore, it lacks insight beyond system integration.

  2. The writing can be improved, e.g., the method section is not self-contained and unclear in details. In particular, L234–238, Section 3.3.2, and Section 3.3.3 are described at an abstract level and hard to follow. To name a few:

  • The difference between the terms a "window" and a "piece"?
  • Level 1 wide-span correction is heuristic and may require ad hoc tuning. It uses sparse views to correct the continuous Level 0 trajectory. While sparse views may lack sufficient overlap, and the selection of these views is not clearly adaptive to different motion speeds or camera frame rates.
  • Whether neighboring submaps contain overlapping frames, how many of them, and their impact;
  • Joint revisit estimation. Does it mean the alignment between submaps is performed by Pi3X? Or can it be solved with a conventional solver such as orthogonal Procrustes? How the relation Z_{ab} is obtained, and does it represent an SE(3) transformation?
  • How poses are represented (homogeneous matrix?) and why they are optimized in log space; at least cite relevant works as reference.
  • The necessity of optimizing Q in Eq. (4) is unclear. It appears to be introduced to absorb the gauge freedom of Pi3X. However, it is not clear why not simply fix one frame’s Q to T_i^{-1}, i.e., select one frame and convernt to identity after Pi3X's prediction.
  • In equations (4) and (5), what exactly does \rho represent, e.g., a robust kernel?
  • Does PGO refer to pose graph optimization? How it is formulated exactly?
  1. The experimental validation is not yet convincing, particularly regarding model dependence and efficiency. The performance of RECALL-SLAM relies heavily on Pi3X, whereas baselines such as VGGT-SLAM 2.0 are originally designed on VGGT rather than Pi3X; this makes it difficult to separate the contribution of the proposed hierarchy from the advantage of using a stronger and RGB-D-conditioned feed-forward model. In addition, the paper lacks a comparison on efficiency. I appreciate that the appendix reports a detailed runtime. But it is about 3 FPS and 50 GB of GPU memory, which makes it hard to deploy on consumer GPUs.
Questions for authors

Please respond to my major concerns in the weakness section. Other minor questions:

  1. The optimization ignores depth and pointmaps? The proposed graph optimization appears to involve only poses, while depth and point maps are either ignored or kept fixed. This may lead to inconsistency between the optimized trajectory and the reconstructed scene structure. Are the trajectory and reconstruction aligned and evaluated separately? In addition, continuously optimizing all cameras over a long trajectory may be redundant.

  2. Motivation not well supported. The paper motivates the method by the need to handle blurred, textureless, and geometrically weak observations. However, the system appears to reject or drop inconsistent predictions rather than explicitly recovering from these failure cases. The limitations section also reveals that blur and weak overlap can still produce incorrect geometry. The errors can impact different hierarchies, but are not solved by the hierarchy design. Therefore, the claimed motivation is not fully supported. Please clarify.

Other suggestion:

Provide detailed discussion on failure cases: e.g., why existing methods fail, and why RECALL-SLAM underperforms on BundleFusion. Such an analysis would provide valuable evidence and insight for attributing success and failure. For example, it would be informative to determine whether VGGT-SLAM 2.0 with Pi3X fails because of false positive retrieval or because the optimization failed to converge. Conversely, whether RECALL-SLAM succeeds because its CLS and patch verification, or because its multi-level corrections.

Reviewer_zHT7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents ReCall-SLAM, a training-free RGB-D SLAM system built on the premise that local tracking, regional correction and loop closure all solve the same bounded multi-view reconstruction problem. It poses each of these to a single pretrained, depth-conditioned Pi3X model at three levels: local windows, context and submap correction, and joint revisit reconstruction. The predictions are integrated through intra- and inter-submap pose graphs that run asynchronously. Validation covers TUM RGB-D, BundleFusion, FastCaMo-Real and Oxford Spires, with ATE and F-score, a hierarchy ablation and runtime analysis. The main claims are a F-score gain on FastCaMo-Real and lower Oxford ATE than all baselines.

Critical strengths
  1. The gains are large exactly where long horizons matter.
    • On the 11 complete Oxford Spires routes, ATE is 76 cm against 297 to 510 cm for the VGGT-SLAM 2.0 variants.
    • On FastCaMo-Real, under heavy motion blur, F@5 rises from 46.3 (PROFusion) to 66.8.

Revisits are handled by jointly reconstructing both visits in one Pi3X call rather than registering two separately reconstructed chunks. That is an interesting design and a plausible explanation for the outdoor gains.

  1. The reporting is unusually candid.
    • Runtime and memory are reported (3.47 frames/s, about 51 GB of GPU memory).
    • The model-choice study (Table 6) admits that RGB-only Pi3 has a lower median ATE than the Pi3X model used.
    • The paper explains how ORB-SLAM3's incomplete runs are handled, and per-sequence tables are given.
Critical weaknesses
  1. The indoor claims are thinner than stated, and the ablation does not test the long-horizon claim.

    • TUM. The "lowest TUM ATE" (2.969 vs 3.165 cm for VGGT-SLAM 2.0-Pi3X) comes from a single sequence. In Table 8, fr2/desk is 4.690 vs 7.401. Without fr2/desk the means are 2.812 (ours) vs 2.780, and head-to-head the method wins 7 of 12 sequences.
    • BundleFusion. VGGT-SLAM 2.0-Pi3X is better on both ATE and F-score (84.82 vs 80.63).
    • Ablation. The hierarchy ablation (Table 3) uses TUM only. These sequences are short, so revisits matter little. Adding Level 1 on top of Level 0 + 2 changes ATE from 3.004 to 2.969 cm. The central claim is that one primitive used across the hierarchy helps over long horizons, so it needs the ablation on Oxford and BundleFusion.
  2. The method is under-specified, and the "streaming SLAM" framing does not match the cost.

    • Missing definitions. Several key pieces are described only in words, with no definitions:
      • the admission tests.
      • how the information Ω is "derived from this evidence".
      • the "additional depth constraints" in Eq. (4).
      • gravity estimated from "measured horizontal surfaces", which is questionable on outdoor slopes.
    • Hyperparameter tuning. Table 4 lists many hand-set thresholds (0.80 new-space criterion, 96-frame separation, 0.10 m grid, and others). The paper does not say which data they were chosen on, and the tuning may have been done on the evaluation benchmarks themselves.
    • Cost. The profiled run takes 3.47 frames/s on an RTX PRO 6000 with a 381 s backlog. That is not online SLAM, and it belongs in the main text.
    • Oxford claim. The abstract's "58.3%" Oxford reduction applies only to the 14 segments where ORB-SLAM3 finishes, and the abstract should say so.
    • Baseline adaptation. The authors' own VGGT-SLAM 2.0-Pi3X adaptation is much worse on Oxford than VGGT-SLAM 2.0-D (510 vs 297 cm), despite having the stronger local model. That points to an adaptation problem and makes the large margin harder to attribute.
Questions for authors

Answering the following queries could change my assessment.

  1. How were the Table 4 hyperparameters chosen, and were any of the evaluation sequences used to choose them?
  2. Can you run the hierarchy ablation (Level 0 / +1 / +2 / full) on the Oxford complete routes and on BundleFusion?
  3. Why is VGGT-SLAM 2.0-Pi3X so much worse than VGGT-SLAM 2.0-D on Oxford? Were its retrieval and loop-closure thresholds adapted to sparse LiDAR depth?
  4. ORB-SLAM3 terminates early on 5 of 7 BundleFusion sequences and on all 11 Oxford routes, which is unusual for RGB-D. Was multi-map (Atlas) relocalization enabled?
  5. Does the training data of Pi3 or Pi3X include TUM, BundleFusion or Oxford Spires? This matters for the "training-free" framing.
  6. Can you give precise formulas for the admission criteria, for Ω, and for the depth constraint term?
  7. Add a classic dense RGB-D system with loop closure (for example BundleFusion itself or BAD SLAM) on the BundleFusion sequences.

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

Related papers

Discussion (0)

Sign in to comment.