MV-RoMa: Multi-View Dense Matching

• MV-RoMa is a multi-view dense matching model that jointly predicts source-to-target warps and confidence maps to form coherent 2D tracks.
• It employs a track-guided encoder and a pixel-aligned multi-view refiner to efficiently align features across images while enforcing geometric consistency.
• The approach enhances SfM performance by reducing errors from pairwise match chaining, leading to denser and more accurate reconstruction metrics.

MV-RoMa is a multi-view dense matching model for establishing dense, mutually consistent correspondences from one source image to multiple co-visible target images, with the explicit aim of turning pairwise image matching into multi-view track reconstruction for structure-from-motion (SfM) and related 3D vision pipelines. It jointly predicts source-to-target warps and confidence maps, then converts those correspondences into high-quality 2D tracks suitable for pose estimation, triangulation, and bundle adjustment. The method is designed to avoid the fragmentation and geometric inconsistency that arise when independently predicted pairwise matches are chained across views (Lee et al., 29 Mar 2026).

1. Definition, scope, and nomenclatural context

MV-RoMa takes as input a set of images $\mathcal{I}=\{I_0, I_1, \ldots, I_{V-1}\}$ that share co-visible scene content. Its output is a set of dense source-to-target warps $\{W_{0\to v}\}_{v=1}^{V-1}$ and confidence maps $\{p_{0\to v}\}_{v=1}^{V-1}$, all predicted jointly from a chosen source view $I_0$ to the target views. These dense, multi-view consistent correspondences naturally induce 2D tracks that can be exported into standard SfM pipelines for pose estimation, triangulation, and bundle adjustment (Lee et al., 29 Mar 2026).

The model is situated in dense correspondence estimation rather than optimization or generic multi-view transformers. Its immediate technical lineage is the pairwise dense matcher "RoMa: Robust Dense Feature Matching" (Edstedt et al., 2023), whose coarse-to-fine dense matching design is extended to the source-to-many setting. By contrast, the earlier paper "RoMA: Robust Model Adaptation for Offline Model-based Optimization" concerns offline black-box optimization with proxy models rather than image matching, and the term “MV-RoMa” does not appear in that work (Yu et al., 2021).

Name	Scope	Relation to MV-RoMa
"RoMA: Robust Model Adaptation for Offline Model-based Optimization" (Yu et al., 2021)	Offline model-based optimization	Unrelated research area; no MV-RoMa variant is defined
"RoMa: Robust Dense Feature Matching" (Edstedt et al., 2023)	Pairwise dense feature matching	Architectural base extended by MV-RoMa
"MV-RoMa: From Pairwise Matching into Multi-View Track Reconstruction" (Lee et al., 29 Mar 2026)	Multi-view dense matching and track reconstruction	The method under discussion

This distinction is methodologically important because the visual-matching RoMa and the offline-optimization RoMA share only a similar acronym. In the literature represented here, MV-RoMa is specifically the multi-view extension of the dense matching line, not a variant of robust model adaptation for offline optimization.

2. Problem formulation and motivation

The central problem is the construction of consistent correspondences across multiple views of the same scene. Pairwise matchers such as LoFTR, SuperGlue, and RoMa operate on two images at a time, after which tracks are typically built by chaining pairwise matches. According to the MV-RoMa formulation, this procedure often accumulates small errors and inconsistencies because it lacks joint reasoning across all views, has no mechanism to enforce multi-view consistency at matching time, and can resolve ambiguities differently for different image pairs, especially in textureless or repetitive regions (Lee et al., 29 Mar 2026).

MV-RoMa addresses this by performing joint source-to-many prediction. It does not require camera intrinsics or extrinsics at inference. During training, however, ground-truth depth and poses are used to derive supervision, and standard pipelines may use known or estimated intrinsics and extrinsics for evaluation, triangulation, and downstream SfM. A track is represented as a set of image-observation pairs referring to the same 3D point,

$\tau = \{(a, u_a), (b, u_b), \ldots\}.$

This setup shifts the modeling objective from isolated correspondences to consistent track formation. A plausible implication is that consistency is treated as a first-class property of the predictor rather than a post hoc filter. The paper makes this explicit: MV-RoMa jointly estimates dense correspondences from a source image to multiple co-visible targets and is designed so that its outputs already form coherent tracks, rather than requiring pairwise predictions to be reconciled later (Lee et al., 29 Mar 2026).

The method also frames efficiency as part of the problem definition. Full multi-view cross-attention over all pixels and views scales as $O((VHW)^2)$, which is computationally prohibitive. MV-RoMa therefore seeks multi-view interaction without incurring that cost, and this requirement shapes both the encoder and refiner design.

3. Architecture: track-guided encoder and multi-view refiner

The architecture has two principal components: a track-guided multi-view encoder and a multi-view matching refiner. The encoder uses a DINOv2 backbone, with track-guided attention modules inserted in the second half of the transformer blocks. Its purpose is to align features across views using a sparse multi-view geometric prior built from track tokens. Those tokens are derived from any off-the-shelf pairwise matcher; the default is UFM, although ALIKED+LightGlue is reported to work with similar performance (Lee et al., 29 Mar 2026).

Track tokens are constructed by aggregating pairwise matches across source and targets into preliminary multi-view tracks, partitioning tracks by visibility pattern, and applying k-means within each partition to select representative tracks. The default number of tokens is $T=512$. Each token $t_i=(u_i,m_i)$ stacks 2D coordinates across views in $u_i\in\mathbb{R}^{2V}$, with missing entries padded, and stores a visibility mask $m_i\in\{0,1\}^V$ indicating which views observe the point.

Feature exchange proceeds in three stages. Let $\{W_{0\to v}\}_{v=1}^{V-1}$0 be the grid tokens for view $\{W_{0\to v}\}_{v=1}^{V-1}$1, and let $\{W_{0\to v}\}_{v=1}^{V-1}$2 be the 2D track coordinates in that view. First, attentional sampling transfers information from the image grid to the sparse tracks:

$\{W_{0\to v}\}_{v=1}^{V-1}$3

with spatial bias

$\{W_{0\to v}\}_{v=1}^{V-1}$4

Second, a track transformer performs self-attention along the view axis on each track, using the visibility mask and no view-index embeddings, making the interaction view-order invariant. Third, attentional splatting returns the updated track information to the grid, producing multi-view aligned features $\{W_{0\to v}\}_{v=1}^{V-1}$5.

The second component is a coarse-to-fine refiner built on RoMa’s global matcher. At the coarse stage,

$\{W_{0\to v}\}_{v=1}^{V-1}$6

Refinement then uses VGG19 fine features $\{W_{0\to v}\}_{v=1}^{V-1}$7 at strides $\{W_{0\to v}\}_{v=1}^{V-1}$8. At refinement level $\{W_{0\to v}\}_{v=1}^{V-1}$9, each target feature map is warped into the source frame using the previous estimate $\{p_{0\to v}\}_{v=1}^{V-1}$0, after which multi-view fusion alternates between pixel-aligned attention and a lightweight ConvNeXt block. The key restriction is that attention is computed across views at the same source pixel after warping, not globally across all pixels. After $\{p_{0\to v}\}_{v=1}^{V-1}$1 iterations, the model regresses residuals $\{p_{0\to v}\}_{v=1}^{V-1}$2 and updates

$\{p_{0\to v}\}_{v=1}^{V-1}$3

$\{p_{0\to v}\}_{v=1}^{V-1}$4

This design preserves multi-view interaction while avoiding full multi-view cross-attention. Encoder sampling and splatting scale as $\{p_{0\to v}\}_{v=1}^{V-1}$5 per view, the track transformer scales as $\{p_{0\to v}\}_{v=1}^{V-1}$6, and the refiner’s pixel-aligned attention scales as $\{p_{0\to v}\}_{v=1}^{V-1}$7 rather than $\{p_{0\to v}\}_{v=1}^{V-1}$8 (Lee et al., 29 Mar 2026). This suggests that the model’s efficiency derives not from eliminating multi-view reasoning, but from constraining where and how it occurs.

4. Supervision, losses, and track reconstruction for SfM

Training supervision is derived from dense depth and known poses on MegaDepth and ScanNet. Ground-truth warps $\{p_{0\to v}\}_{v=1}^{V-1}$9 and co-visible masks $I_0$0 are computed by projecting 3D points to each view and verifying visibility. The matching objective uses the robust loss inherited from RoMa,

$I_0$1

with multi-scale supervision applied at the coarse stage and across refiner levels (Lee et al., 29 Mar 2026).

The post-processing stage converts dense correspondences into robust 2D tracks. For scalability, the image set is partitioned into groups of one source plus up to $I_0$2 targets, with default $I_0$3. A two-stage sampler builds groups with strong overlap and reciprocity, and the group budget is set to approximately $I_0$4 per scene. When an ordered pair $I_0$5 appears in multiple groups, predictions are pooled by max-confidence selection: for each pixel $I_0$6 in $I_0$7, the selected prediction is the one from the group with highest confidence.

Reciprocity is enforced using a forward-backward cycle check. For a pair $I_0$8, a correspondence is retained only if

$I_0$9

and in post-processing the default is $\tau = \{(a, u_a), (b, u_b), \ldots\}.$0 px. Track seeds are then selected on each source image using a score

$\tau = \{(a, u_a), (b, u_b), \ldots\}.$1

where $\tau = \{(a, u_a), (b, u_b), \ldots\}.$2 is the number of targets whose confidence exceeds $\tau = \{(a, u_a), (b, u_b), \ldots\}.$3 and $\tau = \{(a, u_a), (b, u_b), \ldots\}.$4 is the mean confidence, with default $\tau = \{(a, u_a), (b, u_b), \ldots\}.$5. Non-maximum suppression with radius $\tau = \{(a, u_a), (b, u_b), \ldots\}.$6 px is applied to $\tau = \{(a, u_a), (b, u_b), \ldots\}.$7, and each selected source pixel yields a multi-view track

$\tau = \{(a, u_a), (b, u_b), \ldots\}.$8

These tracks are constructed to be compatible with COLMAP-based SfM. MV-RoMa itself does not estimate camera poses. Instead, its output supports downstream robust geometric verification, pose initialization, triangulation, and bundle adjustment. The standard bundle adjustment objective used by COLMAP is

$\tau = \{(a, u_a), (b, u_b), \ldots\}.$9

In this sense, MV-RoMa is best understood as a front-end for multi-view reconstruction: it improves the quality of the tracks that classical SfM back-ends optimize (Lee et al., 29 Mar 2026).

5. Empirical performance and ablation results

Across homography estimation, triangulation, and camera pose estimation, MV-RoMa is reported to produce more reliable correspondences and denser, more accurate reconstructions than pairwise sparse and dense matching baselines. On HPatches homography estimation, the DLT AUC at $O((VHW)^2)$0 px is $O((VHW)^2)$1 for MV-RoMa, compared with $O((VHW)^2)$2 for RoMa and $O((VHW)^2)$3 for LoFTR. Under RANSAC, MV-RoMa attains $O((VHW)^2)$4, compared with $O((VHW)^2)$5 for RoMa. The paper notes that the small DLT–RANSAC gap suggests high inlier quality (Lee et al., 29 Mar 2026).

On ETH3D triangulation, MV-RoMa improves both accuracy and completeness over pairwise baselines. In the pairwise-baseline setting, MV-RoMa reaches accuracy $O((VHW)^2)$6 and completeness $O((VHW)^2)$7 at $O((VHW)^2)$8 cm, compared with RoMa at accuracy $O((VHW)^2)$9 and completeness $T=512$0. Against full SfM pipelines that exclude two-view tracks, MV-RoMa achieves the highest reported accuracy, $T=512$1, while maintaining competitive completeness.

For multi-view camera pose estimation, the reported AUCs on Texture-Poor SfM are $T=512$2 at $T=512$3, compared with $T=512$4 for RoMa+Dense-SfM and $T=512$5 for LoFTR+DF-SfM. On IMC PhotoTourism, MV-RoMa attains $T=512$6, compared with $T=512$7 for RoMa+Dense-SfM.

Benchmark	MV-RoMa	Comparison
HPatches DLT AUC @ 1/3/5 px	46.1 / 71.9 / 80.1	RoMa: 41.0 / 67.9 / 76.9
ETH3D Accuracy @ 1/2/5 cm	81.91 / 90.45 / 96.93	RoMa: 75.58 / 86.25 / 94.95
Texture-Poor pose AUC @ 3°/5°/10°	51.79 / 66.77 / 81.74	RoMa+Dense-SfM: 49.94 / 66.23 / 81.41

The ablations isolate the contributions of the MV-Encoder and MV-Refiner. Starting from the RoMa baseline, adding the MV-Encoder improves ETH3D accuracy/completeness from $T=512$8 and $T=512$9 to $t_i=(u_i,m_i)$0 and $t_i=(u_i,m_i)$1, while adding the MV-Refiner further improves them to $t_i=(u_i,m_i)$2 and $t_i=(u_i,m_i)$3. The conclusion reported in the paper is that encoder and refiner are complementary.

Tokenization and interaction ablations further support the design. With $t_i=(u_i,m_i)$4 track tokens, HPatches DLT AUC is $t_i=(u_i,m_i)$5, whereas with $t_i=(u_i,m_i)$6 it falls to $t_i=(u_i,m_i)$7, indicating that many well-distributed tokens help feature exchange. For number of interacting views, $t_i=(u_i,m_i)$8 yields $t_i=(u_i,m_i)$9, slightly above the pairwise $u_i\in\mathbb{R}^{2V}$0 setting at $u_i\in\mathbb{R}^{2V}$1, showing a measurable benefit from joint multi-view reasoning. Clustering-based token sampling outperforms random sampling and improves spatial coverage (Lee et al., 29 Mar 2026).

Runtime measurements also reflect the efficiency goal. On an RTX 6000 Ada at $u_i\in\mathbb{R}^{2V}$2, RoMa takes $u_i\in\mathbb{R}^{2V}$3 ms per $u_i\in\mathbb{R}^{2V}$4 pair and $u_i\in\mathbb{R}^{2V}$5 ms for five pairs batched, whereas MV-RoMa takes $u_i\in\mathbb{R}^{2V}$6 ms for $u_i\in\mathbb{R}^{2V}$7 jointly. For end-to-end SfM on a 20-image Texture-Poor scene, DF-SfM + RoMa requires $u_i\in\mathbb{R}^{2V}$8 min total, MV-RoMa Full-Budget requires $u_i\in\mathbb{R}^{2V}$9 min, and MV-RoMa Half-Budget requires $m_i\in\{0,1\}^V$0 min, with no per-track post-refinement cost and negligible selection/filtering cost.

6. Relation to pairwise RoMa, limitations, and research significance

MV-RoMa is explicitly built on the pairwise dense matcher RoMa rather than replacing it wholesale. The earlier RoMa combines frozen DINOv2 coarse features, VGG19 fine features, a coordinate-agnostic transformer decoder with anchor probabilities, and a training objective that uses regression-by-classification at coarse scales and robust regression at fine scales (Edstedt et al., 2023). MV-RoMa inherits the coarse-to-fine dense matching paradigm but adds three capabilities absent from the pairwise model: a track-guided multi-view encoder, a multi-view pixel-aligned attention refiner, and one-pass production of multi-view consistent outputs instead of post hoc chaining (Lee et al., 29 Mar 2026).

The contrast with pairwise matchers is substantive. RoMa, LoFTR, and SuperGlue operate on two images at a time and do not perform joint multi-view feature exchange. MV-RoMa instead estimates source-to-many correspondences simultaneously, with encoder-level multi-view interaction guided by sparse geometric priors and refiner-level pixel-wise multi-view attention after warping. The paper also positions it against multi-view transformers with full attention, such as VGGT and Pi3, emphasizing that MV-RoMa avoids full $m_i\in\{0,1\}^V$1 cross-attention through sparse track-token exchange and pixel-aligned mixing.

The reported limitations are specific. The method assumes co-visibility; very sparse overlap or extreme baselines can reduce the quality of the prior and the effectiveness of attention. Its encoder exchange depends on prior pairwise matches, so very noisy or spatially concentrated priors weaken performance, although clustering and coverage-aware sampling mitigate this. Textureless and dynamic regions remain challenging, even if multi-view evidence improves robustness relative to pairwise methods. Scalability with very large $m_i\in\{0,1\}^V$2 is also limited: although per-pixel across-view attention is linear in $m_i\in\{0,1\}^V$3, memory and time still grow with the number of targets, and practical group sizes around $m_i\in\{0,1\}^V$4–$m_i\in\{0,1\}^V$5 are recommended (Lee et al., 29 Mar 2026).

These limitations clarify the method’s intended operating regime. MV-RoMa is not a generic all-to-all multi-view correspondence transformer; it is a source-to-many dense matcher optimized for moderate group sizes and SfM-compatible track formation. This suggests that its main significance lies in reframing dense matching as track reconstruction under explicit multi-view consistency constraints. Within that framing, the method provides a computationally tractable bridge between modern dense correspondence networks and classical geometric reconstruction pipelines.