acceptodds
Withdrawn from ICLR 2027

SLAM: TURNING FEED-FORWARD GEOMETRY INTO PERSISTENT FACTORS FOR SLAM

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

18 views

Abstract

Feed-forward 3D models provide strong multi-view geometric priors, while online simultaneous localization and mapping (SLAM) relies mainly on local measurements and can accumulate drift over long sequences. Existing attempts to combine the two typically treat feed-forward predictions as an external geometric state that is aligned or fused with the online estimate after the fact, which keeps broader multi-view evidence outside the optimizer that refines the SLAM state. We present , which instead converts feed-forward geometry directly into optimization-native target-weight measurements attached to a persistent dense factor graph. A high-frequency stream maintains local tracking constraints and graph connectivity, while a low-frequency stream uses wider multi-view context to selectively refresh existing measurements after a state-consistency check. Both streams constrain the same poses, inverse depths, and optional camera intrinsics through a single dense bundle adjustment. Experiments on multiple benchmarks demonstrate consistently strong trajectory estimation and improved dense reconstruction in both calibrated and uncalibrated settings. Notably, the uncalibrated configuration reduces the average ATE RMSE from  m for the strongest feed-forward baseline to  m on the Replica dataset.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Simultaneous Localization and Mapping

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_M2cmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

F²SLAM integrates high-frequency DROID-style tracking and low-frequency Flow3r predictions within a shared dense bundle-adjustment framework. The high-frequency branch updates tracking correspondences, while the low-frequency branch selectively replaces them after consistency checks and provides geometry for initialization. Experiments demonstrate competitive trajectory estimation and dense reconstruction.

Critical strengths
  1. Practical integration. Expressing both branches’ outputs as correspondence targets and weights provides a practical way to incorporate multi-view information into an existing dense BA pipeline.

  2. Competitive results. The system achieves strong trajectory estimation in the uncalibrated setting, particularly on TUM and Replica, and improves reconstruction completeness on 7-Scenes.

Critical weaknesses
  1. The independent contribution of measurement refresh is insufficiently established. On Replica, basic tracking already achieves a strong ATE of 0.0027 m, leaving relatively modest gains from measurement refresh (0.0026 m) and the full system (0.0023 m). It would be helpful to see ablations on other datasets to better understand how much measurement refresh adds beyond the already strong tracker, initialization, and intrinsic refinement.

  2. The benefit of measurement locking remains unclear, as the ablations do not separate its contribution from correspondence replacement itself.

Questions for authors
  1. Could the authors extend the ablations to TUM and KITTI, keeping intrinsic optimization and other backend settings identical? Given the strong performance of Basic Tracking on Replica, these comparisons would help clarify the additional benefit of measurement refresh.

  2. What is the empirical benefit of locking accepted measurements against subsequent high-frequency updates? A comparison with replacement without locking would help establish its contribution. Jointly retaining both measurement sources could provide another informative comparison.

  3. Could the authors report the runtime and peak GPU memory usage of Basic Tracking, the initialization-only variant, and the full system under identical settings? This would help clarify the accuracy–efficiency trade-off introduced by measurement refresh.

  4. Could the authors revisit the layout of Figure 2? The “Persistent Dense Factor Graph” block in particular appears crowded and is difficult to parse, while there is noticeable unused space on the right. Adjusting the layout would greatly improve readability.

Reviewer_Sa4QOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a visual SLAM system (without loop closure) built on two complementary neural models. A high-frequency model handles frame-to-frame tracking. A low-frequency model applies a feed-forward geometric foundation model to keyframes for wide-baseline pose estimation and reconstruction. The resulting depth, poses, and (optionally) intrinsics are then refined jointly in a dense bundle adjustment (BA). The paper has two key insights. First, the geometric foundation model is best used at a low frequency, where it can exploit a broader multi-view context. Second, its outputs should enter the subsequent optimization directly. Experiments on several benchmarks show competitive results, outperforming the baselines.

Critical strengths
  1. The figures with easy-to-follow explanations are of high quality.
  2. The method achieves state-of-the-art results on several public datasets, even without loop closure.
  3. The ablation studies are clear and informative.
  4. The complementary use of two models at different frequencies, together with a dense BA that consumes the foundation model’s outputs, is an interesting design.
Critical weaknesses
  1. The tracking weights and the foundation-model weights come from two different networks, so their scales are not naturally comparable. The paper appears to rely on a single fixed scalar (η_low = 0.25, used on all datasets) to set the relative weight of the foundation-model measurements, while the tracking weights are used as predicted, and their range is not reported. The paper provides sensitivity analyses for other hyperparameters (τ_geo, target motion, window size), but I could not find one for η_low. I encourage the authors to report how performance varies with this parameter and to show the distribution of the two kinds of weights.
  2. The paper contrasts F²SLAM with existing systems that treat feed-forward geometry as an external state to be aligned or fused, and states that it instead converts feed-forward outputs into optimization-native target–weight measurements. From my reading, the foundation model’s contribution to the optimization is a dense displacement target with a confidence-derived weight (plus pose and depth for initialization), admitted through a consistency check. To my understanding, systems such as MASt3R-SLAM also use foundation-model correspondences and confidences inside their optimization, so I would ask the authors to clarify precisely how their formulation differs, and to support the claimed benefit with a controlled comparison using the same foundation model. It would also help if the introduction or the beginning of Sec. 3.1 stated concretely which outputs of the foundation model are used and how they enter the optimization.
Questions for authors
  1. It would be interesting to see peak GPU memory on a larger-scale dataset such as KITTI.
  2. This paper presents an effective engineering contribution. It combines the high tracking accuracy of DROID-SLAM (see TUM, Replica, and 7-Scenes) with Flow3r’s ability to provide longer-range constraints between distant keyframes (see KITTI). The proposed framework seems extensible and easy to interpret. Would it still work if the tracking and feed-forward components were replaced with other methods, for example ORB-SLAM for tracking and VGGT for the feed-forward part? Experiments of this kind would make the work considerably more convincing.
Reviewer_JZmcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

F²SLAM combines high-frequency DROID-style tracking with lower-frequency multi-view predictions from Flow3r. Accepted feed-forward correspondences replace measurements on existing graph edges and remain protected from tracking updates during their active lifetime. A shared bundle-adjustment objective optimizes camera poses, dense inverse depths, and optionally intrinsics. Experiments cover four datasets, calibrated and uncalibrated settings, trajectory accuracy, reconstruction, and computational cost.

Critical strengths
  1. The refresh-and-lock mechanism gives accepted multi-view correspondences a sustained influence on subsequent BA updates. Joint refinement of poses, dense depths, and shared intrinsics can then propagate this information through the active graph. This is a useful integration design for incorporating broader visual context into an ongoing SLAM estimate.

  2. The reported results show practical value when camera calibration is unavailable. On TUM, F2SLAM achieves a mean ATE of 0.027 m, compared with 0.039 m for SLAM-Former and 0.041 m for VGGT-SLAM 2.0 (Table 1). This supports the system’s usefulness for monocular SLAM without supplied calibration.

Critical weaknesses
  1. The public Flow3r implementation exposes per-image confidence, but the manuscript does not explain its conversion into pair-specific correspondence weights or their scale relative to DROID weights. Active/inactive-factor handling is also incomplete. More fundamentally, Eqs. (23)/(37) require an explicit scale constraint: scaling translations by s and inverse depths by 1/s preserves reprojections while reducing inverse-depth smoothness by 1/s². With positive smoothness weight and no additional anchor, the displayed objective favors increasing scale. Initialization alone does not resolve this issue.
  2. Characterizing previous systems as one-off external corrections overlooks repeated graph optimization in MASt3R-SLAM and VGGT-based systems. DROID-SLAM already provides weighted correspondence-based dense BA; the distinction should therefore center on measurement refresh and its demonstrated benefit. In Table 7, Basic Tracking already achieves 0.0027 m on Replica versus 0.0023 m for Full, so the much larger cross-system gap cannot be attributed entirely to the proposed additions. Several averages in Tables 2, 4, and 5 also disagree with the displayed sequence values beyond rounding tolerance under equal weighting.
Questions for authors
  1. How exactly are pair-specific Flow3r confidence and BA weights constructed, including coordinate-grid conversion and cross-stream scaling? What constraint fixes scale in Eqs. (23)/(37)?
  2. What differs between Basic Tracking and a matched DROID baseline, and which measured gains specifically establish the contribution of persistent measurement refresh?
  3. What aggregation rules produce the reported averages? For example, Table 4's uncalibrated F2SLAM entries average 30.54 m, whereas the table reports 30.18 m. Please reconcile this and the discrepancies in Tables 2 and 5.

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

Related papers

Discussion (0)

Sign in to comment.