acceptodds
Withdrawn from ICLR 2027

XGMap: Evidence-Grounded Active Mapping with Foundation Geometry

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Active mapping requires geometry that is useful for choosing future observations, not only accurate on a fixed image set. Feed-forward geometry models offer dense reconstruction from RGB inputs, yet integrating their predictions into a persistent planning state raises two distinct challenges: metric compatibility across causal inference windows and the interpretation of incomplete evidence. We introduce XGMap, an RGB-only active mapping framework built around causal, oracle-free geometry integration and evidence-grounded planning. The intended integration module reconciles predictions with known metric poses and past observations; the planning interface distinguishes observed surfaces, verified free space, and unknown regions while combining anticipated coverage with exploration value. To establish when these components help, our evaluation separates matched geometry and planner interventions from end-to-end baseline comparisons. It also distinguishes coverage measured from ground-truth observations at visited poses from the fidelity of the native reconstructed map.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D reconstruction, active mapping, neural geometry

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_utbQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

XGMap is an RGB-only active mapping framework that integrates foundation-model geometry into closed-loop exploration. It fuses repeated MapAnything depth predictions over overlapping windows, uses cross-window consistency to down-weight unreliable geometry, and maintains a tri-state occupied/free/unknown observation map to explicitly reward unexplored space. These signals are combined with MAGICIAN-style imagined-Gaussian coverage in a receding-horizon beam-search planner. Experiments on Macarons++ and Replica show strong coverage and reconstruction performance, including results competitive with several RGB-D baselines.

Critical strengths
  1. Clear problem formulation and a well-motivated system design.

The paper identifies a failure mode when directly using foundation-model geometry for closed-loop active mapping: predicted depth can be multi-view inconsistent and can incorrectly turn unobserved regions into free space. The separation between a learned structural prior and explicit observation evidence is intuitive and provides a clean way to combine geometric anticipation with exploration of genuinely unresolved space.


  1. Strong empirical performance with useful component ablations.

XGMap achieves strong coverage results on both Macarons++ and Replica, despite using predicted depth while several competing methods use simulator ground-truth depth. The ablations also show meaningful drops when removing the measured-uncertainty reward, the structured prior, or the tri-state representation, suggesting that the proposed components contribute to the overall system.

Critical weaknesses
  1. The key “Naive MapAnything” baseline is not controlled for planning horizon, making the main plug-and-play comparison difficult to interpret.

A central claim of the paper is that simply feeding MapAnything depth into MAGICIAN is insufficient, and Table 3 uses the “Naive MapAnything” baseline to support this claim. However, the full XGMap uses a search horizon of (S=5), whereas the direct MapAnything+MAGICIAN baseline retains MAGICIAN’s native (S=10) horizon. Importantly, the authors themselves state that they shorten the horizon to 5 specifically to reduce the accumulation of monocular prediction errors during rollouts.

Therefore, the observed improvement from the naive baseline to XGMap cannot be attributed solely to the proposed evidence-grounded fusion and planning mechanisms. A matched comparison with the same horizon is necessary, ideally together with a sensitivity study over horizons such as (S=3,5,10). This is consequential because the direct-integration baseline is one of the paper’s most important pieces of evidence that its additional machinery is necessary.

Questions for authors
  1. Could the authors provide a matched planning-horizon comparison and sensitivity analysis for XGMap and the Naive MapAnything + MAGICIAN baseline, e.g. using (S=3,5,10) for both methods? The current comparison uses (S=5) for XGMap but MAGICIAN’s native (S=10) for the baseline, while the paper itself notes that longer horizons can accumulate monocular prediction errors. This makes it difficult to determine how much of the observed gain comes from the proposed evidence-grounded design versus the shorter rollout horizon.
Reviewer_xL5MOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

XGMap addresses active mapping from RGB observations with known metric camera poses. It fuses repeated MapAnything depth predictions across causal sliding windows, maintains surface/free/unknown observation records, and combines exploration of unsupported volume with predicted surface coverage rendered from imagined Gaussians. Cross-window consistency weights the coverage reward, and a receding-horizon planner selects subsequent views. Experiments evaluate coverage and reconstruction on Macarons++ and Replica. The contribution is an integrated closed-loop mapping method for handling unreliable foundation-model geometry.

Critical strengths
  1. The distinction between missing observation support and predicted surface presence addresses a real active-perception problem. A point cloud alone cannot distinguish unexplored space from free space, and learned completion alone can direct exploration toward incorrect geometry. Combining observation memory with anticipated coverage is a meaningful design choice; Table 3 shows substantial degradation when either reward is removed.
  2. The method supplies a practical route from frozen foundation depth to closed-loop planning. Causal overlapping windows, delayed commitment, persistent evidence, and receding-horizon action selection form a coherent implementation. This integration is useful because reconstruction errors affect subsequent observations, rather than only an offline accuracy metric.
Critical weaknesses
  1. The claimed mechanism-level advance is not sufficiently isolated. The abstract and conclusion identify consistency-weighted planning as essential, yet Table 3 does not remove this weighting separately or independently test repeated depth fusion. The three-state control uses binary frustum marking, which discards ray-depth and occlusion semantics; it therefore cannot establish the benefit of irreversible surface precedence over a conventional occupancy update using the same observations. Since a single erroneous surface observation becomes permanent, the proposed protection against false clearing can also retain false geometry. The existing coverage ablations demonstrate that the complete system is useful, but not that these particular mechanisms resolve the stated failure modes.
  2. The direct-integration comparison and aggregate results leave the significance of the incremental contribution uncertain. Naive MapAnything uses MAGICIAN's native ten-step horizon, whereas XGMap uses five steps, alongside changes in fusion, rewards, and rollout novelty. This comparison does not attribute the gain to the proposed evidence handling. Outdoor AUC improves to 0.718 from 0.703 for MAGICIAN and 0.700 for MACARONS, but final coverage is below MAGICIAN, 0.882 versus 0.892. The reported within-scene dispersion does not quantify uncertainty on paired method differences. Given that most ingredients build on established mapping and exploration machinery, matched controls and a clearer account of the gains are important to establish sufficient new knowledge for acceptance.
Questions for authors
  1. With all planner settings fixed, what are the separate effects of cross-window depth fusion and consistency weighting? Can you provide independent removals of these components to substantiate the claim that consistency-weighted planning is essential?
  2. How does irreversible surface precedence compare with an occupancy update that can revise contradictory evidence under the same MapAnything inputs? Quantifying false clearing and persistent false occupancy would directly test the proposed trade-off.
  3. Does the improvement over Naive MapAnything persist at a matched search horizon? Please also provide paired trajectory-level differences or confidence intervals for the comparisons with MAGICIAN and MACARONS, so the robustness of the relatively modest outdoor AUC gains can be assessed.
Reviewer_ck9QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents XGMap, an RGB-only framework for active 3D mapping. It uses MapAnything to predict depth from overlapping image windows and combines these predictions into a point cloud based on their confidence. Depth consistency across windows is used to weight the new-surface coverage reward from imagined Gaussians. The method also tracks which regions remain unknown and rewards exploring them. It combines these two rewards and uses beam search to choose the next best view. Experiments on Macarons++ and Replica show better final coverage and reconstruction quality than the RGB-only baseline, with coverage efficiency comparable to that of sota RGB-D methods.

Critical strengths
  1. RGB-only active mapping is a meaningful problem, as depth cameras can be unreliable in real-world conditions.
  2. Using depth consistency to weight planning rewards allows the reliability of geometry predictions to guide subsequent planning.
  3. The main experiments are fairly comprehensive, covering both indoor and outdoor scenes across different benchmarks.
Critical weaknesses
  1. Limited contribution. The method mainly integrates these components: MapAnything, occupancy network, and MAGICIAN’s imagined-Gaussian beam search. The contribution seems primarily focused on component integration.
  2. Lack of ablation study. The paper lacks separate ablations of the consistency weight and sliding-window depth fusion. Moreover, removing G_meas yields an AUC of 0.646, below the baseline that directly feeds MapAnything depth into MAGICIAN (0.657). This raises the question of whether the gains mainly come from the unknown-space reward rather than fusion or consistency weighting.
  3. The lack of occlusion checks in G_meas may limit indoor exploration. This reward simply counts unknown points within the FOV. This could contribute to failures in indoor environments, such as Replica apartment_0.
Questions for authors
  1. The direct-integration baseline uses a search horizon of S=10S=10, while XGMap uses S=5S=5. Could the authors report both methods at S=5S=5 and S=10S=10 to see the effect of the search horizon?
  2. Incorrect surface evidence cannot be removed. Could the authors quantify how these errors affect exploration or evaluate robustness under different levels of depth noise?
  3. Could the authors provide a per-step runtime breakdown for depth prediction, map updates, and planning, along with a comparison against the baselines, to clarify the computational overhead of XGMap?
Reviewer_xUY2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper integrates a frozen geometry foundation model into active mapping. Assuming known metric poses and intrinsics, the method repeatedly predicts depth over sliding image windows and fuses the prediction after a fixed delay. A volumetric representation is recorded that distinguishes surface-supported, free-space-supported, and unknown proxy points. The planner combines unknown volume reward with imaged Gaussian coverage from a learned occupancy prior. Receding-horizon beam search selects actions in a closed loop for active data acquisition. The evaluation is conducted across 15 Macarons++ scenes and 18 Replica scenes with better results.

Critical strengths

The separation of anticipated surface coverage and observation support is well motivated and gains empirical support. Tab.3 indicates that the removal of unknown-space reward or the learned prior reward leads to deterioration.

Critical weaknesses

The experiments do not isolate the mechanisms most central to the claimed contribution: Neither Tab.3 nor the appendix reports a control with consistency weights fixed to one, although the abstract and conclusion state that consistency-weighted planning is essential. Besides, the naive MapAnything comparison additionally uses a 10-step search horizon, whereas XGMap uses five steps specifically to reduce prediction-error accumulation. The comparison changes the planner setting along with fusion and reward design.

Another issue is that the proposed evidence handling can permanently encode the same depth mistakes it aims to make planning robust against as shown in Eq.4. The surface-support event makes a proxy point occupied indefinitely, whereas the erroneous free-space event removes its unknown-space reward permanently unless later surface evidence changes it to occupied. The false occupied/free assignments are not quantified, and the effect on later action selection is not validated.

Questions for authors

The questions requested are clarified in the Weakness section.

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

Related papers

Discussion (0)

Sign in to comment.