MVLoMa: Multi-View Local Feature Matching with Any Geometry Foundation Model
Local feature matching underpins SfM and visual localization, but standard matchers process every image pair independently, which is inefficient and ignores multi-view cues available in star-graph neighbourhoods. MVLoMa addresses this by jointly matching a centre image against all neighbours in a single forward pass, reusing frozen geometry-aware tokens from a multi-view foundation model such as DA3-Giant and adding star-structured cross-attention so the centre aggregates evidence from all neighbours. It preserves the conventional sparse pairwise outputs required by RANSAC while improving strict-threshold pose AUC and cutting matcher compute by roughly $2.2\times$ at $25$ views.
MVLoMa is a careful, well-scoped architectural contribution that improves sparse star-graph matching by reusing frozen multi-view foundation features and joint attention. The experiments are honest about boundaries: the paper explicitly notes that dense two-view RoMa still wins on Mip-NeRF 360 at the strictest threshold ($AUC@1^{\circ}$) and that additional views beyond a small number do not monotonically improve results. The gains over the strong LoMa-L baseline are modest but consistent, and the $2.20\times$ matcher-only speedup is a practical benefit for star-shaped subproblems. Overall, the work makes a solid case for treating the star graph as a native matching unit rather than a batch of independent pairs.
The star-graph attention design is principled and specific: neighbours never attend to one another, so the output remains a set of pairwise score matrices that downstream RANSAC can consume directly. The reuse of frozen foundation tokens is a timely and pragmatic choice—it avoids fine-tuning a massive geometry encoder while injecting cross-view context into descriptors before assignment. The geometric negative loss is well-motivated, targeting high-confidence false matches that survive mutual selection and disproportionately harm pose estimation.
The absolute improvements over LoMa-L are small—only $1.91$ and $0.76$ $AUC@1^{\circ}$ points on HyperSim and ScanNet++ respectively—and the method loses to dense RoMa on Mip-NeRF 360. The $2$–$8$ view training range is far smaller than the $25$-view evaluation star, and the non-monotonic view-count trend (HyperSim@$1^{\circ}$ drops from $54.47$ at $13$ views to $53.90$ at $25$ views) suggests the joint attention does not yet reliably exploit larger contexts.
The $2.20\times$ speedup figure is matcher-only and ignores the cost of the frozen multi-view foundation encoder, which is the dominant runtime; similarly, peak matcher memory grows to $979$ MiB versus $176$ MiB for sequential pairs. No complete SfM or localization experiment is shown, so the practical impact beyond relative pose AUC remains unvalidated.
The paper compares methods fairly by holding the set of evaluated centre–neighbour edges constant, so differences stem from joint versus pairwise reasoning rather than easier pair selection. The backbone ablation cleanly isolates the frozen feature source, showing DA3-G outperforms DINOv2-G and DINOv2-L by a meaningful margin. However, the multi-view baseline MV-RoMa is forced to $13$ views due to out-of-memory at $25$ views, weakening the head-to-head multi-view comparison. The paper is transparent that dense two-view RoMa still leads on Mip-NeRF 360, but it offers little diagnosis for why geometry-aware features do not help there, leaving a gap in understanding the method's domain sensitivity.
Training hyperparameters, data mixture weights, and architecture details are given in Appendix A, and the authors promise code release upon publication. The model builds on publicly available LoMa components and frozen foundation encoders, which aids reproduction. Barriers include the need for a high-end GPU (A800), 200k iterations, and a complex training mixture involving several datasets with custom overlap sampling and recomputed ground truth. The paper does not report variance or confidence intervals for the AUC metrics, and it omits end-to-end wall-clock timings that include the foundation encoder, which is critical for assessing real-world deployment.
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 a 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 obtains the best macro-average relative-pose AUC among the compared methods at all three thresholds (58.65/85.41/90.99). Its matcher is 2.20 times faster than sequential evaluation of the 24 centre-neighbour edges. The same design supports four tested frozen geometry encoders -- DA3, $\pi^3$, VGGT, and VGGT-$\Omega$ -- with DA3-Giant performing best on all reported encoder-ablation metrics. Controlled ablations further isolate the effects of geometry-aware descriptors, visibility filtering, geometric negative supervision, and joint multi-view inference.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.