acceptodds
Under review as a conference paper at ICLR 2027

MVLoMa: Multi-View Local Feature Matching with Any Geometry Foundation Model

Abstract

Local feature matching is typically formulated for image pairs, although structure-from-motion and visual localization commonly match one reference image against many retrieved or covisible neighbours. Applying a pairwise matcher independently to these centre–neighbour edges duplicates reference-side computation and prevents the matcher from exploiting complementary multi-view geometric evidence from the remaining observations. To address these limitations, we introduce MVLoMa, a sparse star-graph matcher that extends LoMa from image pairs to multiple views. MVLoMa first uses any frozen multi-view geometry model to jointly describe all images in a star, then injects the resulting cross-view geometric context into local descriptors. It then matches the centre image to all neighbours in a single forward pass through per-view self-attention and bidirectional centre–neighbour cross-attention. For supervision, we derive positive correspondences only from depth-valid, mutually visible dense warps and explicitly penalize high-confidence matches that violate epipolar geometry. On 25-view star graphs from HyperSim, ScanNet++, Mip-NeRF 360, and Tanks and Temples, MVLoMa achieves state-of-the-art average relative-pose AUC@ across the four benchmarks. At 25 views, its matcher is faster than sequential evaluation of the 24 centre–neighbour edges. Furthermore, MVLoMa can be instantiated with any frozen geometry foundation model—including DA3, , VGGT, and VGGT-—and all variants achieve strong performance. Controlled backbone, supervision, and view-count ablations demonstrate the benefits of geometry-aware descriptors, visibility-aware supervision, geometric negative supervision, and joint multi-view inference.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multi-view local feature matching, geometry foundation models, visual localization, sparse star-graph matching

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_PRc4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a novel framework called MVLoMa, a feature matcher that handles many images at once instead of one pair at a time. For structure-from-motion and visual localization, one uses multiple pairs of images which finally results in forming a graph of unique points having unique feature descriptors which the paper refers to as “star graph”. Standard methods do not process all the images at once which wastes computation and also discards evidence of ambiguity since the scope is limited to only two images vs many. The proposed approach MVLoMa mitigates that with Geometry Aware Descriptors, Star Graph Attention and Better Supervision. The paper also validates the claim with extensive experimentation and also presents comparison with other similar methods. Furthermore, the claims are also supported with insightful ablation study.

Critical strengths
  • The paper tackles a well motivated problem of using pairwise image matchers which put the image pairs in isolation from the other images which belong to the same multi view.
  • The paper acknowledges that RoMa still wins on Mip-NeRF 360 at AUC@1,that LoMa-L leads on Tanks and Temples, and that the view-count trend is non-monotonic. The limitations section addresses dependence on foundation-feature quality, the lack of neighbor - neighbor exchange, and the train/test view-count gap.
Critical weaknesses
  • While the paper presents a novel framework called MVLoMa which improves upon LoMa, still it have some major flaws
    • MVLoMa performance in the benchmark presented in the paper seems very incremental.
    • The paper heavily focused on showing the benchmarks which are incremental but fails to highlight any unique insight of using MVLoMa.
    • Even in ablation study, the difference between using one component vs using full network costs roughly around 1% accuracy. One result stands out in the paper when the overlap is > 0.05 in hypersim, DA3-G performs surprisingly better than DINOv2. This must imply that DAG-3 helps significantly when the overlap is very minimal among the images. This could be great insight if that would be the case with other datasets such as AerialMD. But this phenomenon is not observed in the AerialMD dataset.
  • Paper does not seem to be polished
    • There are several notation collisions
      • l is used for center and layer, N is used for Number of views and V is used for the same, K is used for intrinsics and keypoint count. τ\tau is used for softmax temperature and epipolar threshold.
      • Irregular switch of “MVLOMA” to “MVLoMa”
Questions for authors

Please refer to Weaknesses

Reviewer_y2iHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MVLoMa, a sparse multi-view feature matching framework that extends pairwise matching to star-structured image collections. The authors identify two limitations of conventional pairwise matchers: redundant reference-side computation and the inability to exploit complementary geometric evidence across multiple views. MVLoMa incorporates frozen multi-view geometry foundation model features into local descriptors and introduces a star-graph attention mechanism that jointly processes a center image and its neighboring views. It also uses visibility-filtered correspondence supervision and a geometric negative loss to penalize high-confidence matches that violate epipolar constraints. Experiments across multiple benchmarks show competitive relative-pose estimation accuracy and a 2.20× matcher-only speedup over sequential pairwise matching on 25-view star graphs.

Critical strengths
  1. The paper addresses a practical limitation of pairwise matching. Processing each image pair independently repeats computation on the reference image and makes it difficult to exploit complementary evidence from other views. MVLoMa's star-graph architecture offers a reasonable solution by sharing the center representation across multiple views while retaining the familiar pairwise correspondence output.

  2. Using cross-view features from frozen geometry foundation models is a well-motivated design choice. Rather than relying solely on generic 2D visual descriptors, MVLoMa incorporates representations from models such as DA3, VGGT, and π³ into a specialized local matching framework. This provides a practical way to exploit pretrained geometric representations without fine-tuning the large foundation encoder.

Critical weaknesses
  1. The improvements in matching accuracy are positive but relatively modest. MVLoMa achieves the best macro-average relative-pose AUC across the evaluated benchmarks, but its advantage over strong baselines is limited. The gains are particularly small at the less restrictive thresholds, and the method does not consistently outperform the strongest baseline on every dataset. Given the additional multi-view architecture and the use of a large geometry foundation model, the overall accuracy improvement is not particularly substantial.

  2. The reported speedup does not reflect the cost of the complete pipeline. The 2.20× improvement is measured for the matcher alone and excludes the computation required to extract features from the frozen geometry foundation model. Since MVLoMa relies on a large multi-view encoder to generate its geometry-aware descriptors, the end-to-end efficiency advantage over conventional pairwise matching remains unclear. The star-graph matcher also uses more peak memory than sequential pairwise evaluation. Reporting the full computational cost would help clarify the practical efficiency benefit.

Questions for authors
  1. Could the authors further clarify how much of the improvement comes from the proposed star-graph attention mechanism, rather than the use of stronger multi-view geometry representations? The current results make it difficult to distinguish the contributions of these two components.

  2. The reported 2.20× speedup excludes the geometry foundation model. Could the authors clarify the end-to-end inference time and memory consumption, including geometry feature extraction? Under what deployment conditions does MVLoMa offer an overall efficiency advantage over conventional pairwise matching?

Reviewer_1JzeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MVLoMa adapts frozen multi-view geometry features to sparse local matching. It combines these features with local refinement and shares one centre-image state across several neighbours, then predicts sparse matches for each edge. Visibility filtering and epipolar hard negatives support training. Experiments show useful geometry-feature gains, compatibility with four encoders, competitive relative-pose accuracy, and lower matcher cost than sequential pairwise inference.

Critical strengths
  1. Useful representation evidence. Table 3 holds the matcher and training setup fixed and compares DA3-G against both DINOv2-G and DINOv2-L. Against DINOv2-G, Hypersim AUC@1 rises from 49.70 to 61.14 at overlap >0.05. This is stronger evidence than a model-size comparison. Table 4 also demonstrates that the adapter works with four geometry encoders.
  2. A sparse design with lower matcher cost. The method combines frozen geometry features with sparse per-edge assignments without requiring precomputed tracks. Sharing the centre computation reduces Transformer FLOPs by 37% at 25 views. Table 6 also reports 203.09 versus 446.96 ms, explicitly for matcher-only timing against sequential pairwise inference with 2048 keypoints.
Critical weaknesses
  1. The added benefit of joint matching is not well established. Table 1 improves macro AUC@1/5/10 over RoMa v2 [21 by 1.70/0.54/0.02 points, with uneven gains across datasets. In Table 5, increasing views from 2 to 25 adds 0.71/0.82/0.83 points, but changes both the geometry encoder's context and the matcher's context. This does not isolate the new head's contribution. The paper also does not break down pairwise-versus-multi-view gains by occlusion, repetitive texture or overlap, and Appendix B lacks side-by-side examples of errors corrected by additional views. Thus, it remains unclear when and why multi-view context helps.
  2. The full inference cost and practical trade-off remain unclear. Table 6's 2.20x speedup is measured against sequential pairwise matching, excludes geometry encoding and feature extraction, and uses 2048 keypoints rather than the 4096 used for the main accuracy results. The final geometry features depend on the input view group, so changing groups may require repeated cross-view computation; no caching or incremental scheme is evaluated. Without complete timings and a batched baseline, it is unclear whether the modest accuracy gains justify the total cost.

[1] Johan Edstedt, David Nordström, Yushan Zhang, Georg Bökman, Jonathan Astermark, Viktor Larsson, Anders Heyden, Fredrik Kahl, Mårten Wadenbäck, and Michael Felsberg. RoMa v2: Harder Better Faster Denser Feature Matching. ECCV, 2026.

Questions for authors
  1. With the same frozen features computed from all 25 views, how do independent pairwise matching and the shared-star head compare on identical edges? Which conditions benefit, such as low overlap, occlusion or repetitive texture? Please provide a breakdown or matched examples showing errors corrected by the joint head.
  2. At the main 4096-keypoint accuracy setting, what are the geometry-encoding, feature-extraction and matching times, total latency versus batched pairwise baselines? For overlapping or changing view groups, which features are actually reused and which must be recomputed?
  3. Which tables use the 100k-step schedule in Section 4.1 or the 200k-step schedule in Appendix A, and which checkpoints are evaluated? Please provide scene-level training/evaluation splits and counts, and state known overlap with frozen-encoder pretraining.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.