acceptodds
Withdrawn from ICLR 2027

Shared Recoverable Evidence for Ground-Aerial 3D Gaussian Splatting Registration and Fusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Registration and fusion of independently reconstructed ground and aerial 3D Gaussian Splatting (3DGS) submaps are essential for large-scale scene reconstruction, while cross-platform viewpoint disparities make the task ill-posed. Existing methods overlook the view-dependent radiance of 3DGS primitives, yielding brittle geometric correspondences and severe photometric interference upon naive merging. We present a framework that couples Shared Recoverable Evidence (SRE)-guided registration with camera-conditioned fusion of ground-aerial submaps. SRE-guided registration derives submap-local support from associated training views of 3DGS submaps, then converts it into transform-consistent shared support for Sim(3) proposal refinement and selection. With the transformation fixed, camera-conditioned fusion learns per-Gaussian soft routing selection weights through cross-view map-teacher supervision from frozen submap renders, and modulates opacity to produce one global 3DGS map. Extensive evaluations on ground-aerial pairs from the TerraSky3D, Horizon-GS, and official ScanNet-GSReg benchmarks demonstrate that our approach outperforms existing state-of-the-art methods in both registration accuracy and cross-view rendering fidelity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Ground-Aerial 3D Gaussian Splatting Registration and Fusion

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

The paper introduces a method for combining aerial- and ground- generated 3D Gaussian Splatting maps. The problem is formulated as a registration problem and uses Shared Recoverable Evidence and camera conditioning. The method is tested on TerraSky3D, Horizon-GS, and ScanNet-GSReg benchmarks, and shows an improvement over the baselines.

Critical strengths
  • The paper presents a clear method for the ground- and -aerial acquisition map combination using Gaussians as the representation. The problem formulation inspired from point cloud registration is a sensible approach; supported by the experimental evaluation
Critical weaknesses
  • My main reservation pertains to limited discussion on the failure cases wrt the underlying assumptions (which are not specified but one needs to infer - this should be explicitly stated). In particular, the aerial and ground observations are rarely taken at the same time, neither from the problem formulation nor from the experiments I can see any discussion on the temporal mismatches across the views. So, I understand that the implicit assumption is that there are no geometrical changes present in the submaps. This has been a critical challenge in the co-registration community (both point cloud and Gaussian representations), as not only the geometrical matching challenges arise (different spatial coverage, occlusions, resolution) but also temporal mismatches. For example, pedestrians, cars but also buildings being rebuilt, streets change etc.
Questions for authors
  • I would like to ask for clarification on the failure case of the Reference-frame proposal (l. 194). How sensitive is this initialization? What is the impact on the refinement then? I would imagine that a proposal generated from e.g., a close-up image from a car may not align with an aerial view, also the overlap might be minimal or non-existent. How is this achieved? What happens in that kind of case? Could you provide some visual/or quantitative examples and elaborate?
  • Another question pertains to the underlying assumption, which is not mentioned in the paper. What happens if the acquisitions differ in timestamp and hence there is a temporal mismatch between the acquisitions? It can be a drastic example of aerial acquisition taking place in 2008 and the ground one in 2020, the trees, buildings and other elements may not be present. It can be even more practical – a drone flyers just a few minutes after the ground acquisition and the pedestrians, cars and other dynamic elements can play a role. Could you elaborate on that and present some quantitative and/or qualitative examples?
  • Line 319, mentions that in Tab 1 uses “each pipeline’s initializer”. Can it be that this has an impact on the reported performance of this method? The used VGGT-Ω is more powerful than DUSt3R, so the improvement may stem from just a better backbone. Could you please elaborate and maybe present an ablation for your method on that?
  • I was also wondering – the method setup somewhat resembles the traditional approach to point cloud registration - get a coarse registration (e.g., RANSAC) and then refine (e.g., ICP). This is also similar to the FGR-ICP baseline that is selected here. Could you please comment on the performance on assuming that the Reference-frame proposal is replaced with the classical RANSAC-like approach and the fine tuning is performed with the proposed method? If I missed something, please feel free to clarify here but I would imagine this as an interesting angle to the variability/generalisation of the method.
  • Another point for discussion is – does that paper need to be about the aerial to ground? It seems that it is about overall and more general registration of Gaussians, as in the tested ScanNet-GSReg. It seems that there is no underlying assumption in the method that is disqualified from registering any RGB-derived Gaussians (excluding the ones in different media than air, e.g., underwater etc.)? Could you elaborate? If yes, I would change the positioning of this paper to a more general problem.
Reviewer_v3SNOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses registration and fusion of independently reconstructed ground and aerial 3DGS submaps. It proposes recoverability-based correspondence weighting for Sim(3) registration and camera-conditioned opacity routing for fusion. The problem is relevant, and the overall pipeline is easy to follow, but I find a gap between the claimed contributions and what the experiments establish.

Critical strengths
  1. Integrating independently trained ground and aerial maps is a useful and challenging problem.
  2. The complete registration pipeline reports strong outdoor results.
  3. The paper provides useful controlled comparisons and openly discusses limitations and negative results.
Critical weaknesses
  1. Suppressing the opposite submap can minimize the native-teacher photometric objective. Tables 5 and 15 further show that the central claim of preserving complementary information largely unsupported.
  2. The main comparison uses different initializers, including VGGT-Ω for the proposed method. The matched-initialization controls are helpful, but SRE+LESA does not consistently improve over ICP. So it does not directly settle the full-pipeline comparison, but the added value beyond initialization remains unclear.
  3. The rendering evaluation focuses on native acquisition views, with aerial cameras transformed together with the aerial map using the estimated registration. Selecting the corresponding source can therefore preserve its rendering quality despite inter-map misalignment. These scores support native-view preservation, but provide limited evidence of a coherent, complementary global reconstruction.
  4. The specific SRE construction are reasonable, but their relationship to reliable cross-map geometry is not established clearly. In Table 12, visibility-only evidence slightly outperforms the full evidence field on both TerraSky3D metrics.
Questions for authors
  1. Does the complete registration method outperform an ICP-based replacement using the same selection procedure?
Reviewer_cXvKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses registration and fusion of independently reconstructed ground and aerial 3DGS submaps. For registration, it obtains initial Sim(3) proposals from VGGT-Ω on submap renderings and refines them with weighted Umeyama/LESA iterations. The correspondence weights are a per-Gaussian "Shared Recoverable Evidence" (SRE) score computed from training views. For fusion, a camera-conditioned router and per-Gaussian routing weights modulate opacity to produce one global 3DGS map. Experiments are on TerraSky3D, Horizon-GS, and ScanNet-GSReg.

Critical strengths
  • Post-hoc merging of ground and aerial 3DGS submaps is a practical problem, and point-cloud registration baselines clearly fail in this setting.
  • The appendix is unusually candid and includes controlled interventions and negative results.
Critical weaknesses
  1. Limited novelty. The pipeline combines existing components: foundation-model initialization (VGGT-Ω), classical weighted Umeyama/ICP-style refinement, and view-dependent selection between submaps. The paper's own controls show that the proposed parts contribute little beyond these components:
    • Registration. With the same initialization on TerraSky3D, plain ICP is more accurate than SRE+LESA: 0.518° vs. 0.746° RRE, with SRE+LESA worse on 12 of 15 pairs (Tab. 14). Uniform evidence also performs no worse than full SRE (Tab. 10, Tab. 12). The large margin over baselines in Tab. 1 therefore appears to come mainly from the VGGT-Ω initialization.
    • Fusion. A non-learned nearest-camera selection gives the same numbers as the final method on AG_0018, Park, and Road (26.052 / 22.226 / 19.983; Tab. 6 vs. Tab. 2). Zero, constant, shuffled, and learned per-Gaussian admission give identical results (Tab. 15). The main text nonetheless presents the fusion as preserving complementary evidence (Eq. 4, Fig. 1, contribution 2).
Questions for authors
  1. Under the same VGGT-Ω initialization and candidate selection, does SRE+LESA outperform ICP/GICP end to end?
  2. Is there any test view where opposite-submap Gaussians contribute measurably to the rendering? If not, should the fusion be described as camera-conditioned submap selection?

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

Related papers

Discussion (0)

Sign in to comment.