WovenAnchor Matcher: Specialized Intra- and Inter-Image Context Modeling for Feature Matching

Zhiyang Li; Ruijiang Jin; Thibaut Klenke; Yusuke Sekikawa; Nakamasa Inoue

cs.CVOct 22, 2026

WovenAnchor Matcher: Specialized Intra- and Inter-Image Context Modeling for Feature Matching

Authors: Zhiyang Li, Ruijiang Jin, Thibaut Klenke, Yusuke Sekikawa, Nakamasa Inoue

Organizations: Institute of Science Tokyo, Tokyo, Japan · Denso IT Laboratory, Tokyo, Japan

Abstract

Semi-dense local feature matching commonly aggregates contextual information before coarse-to-fine correspondence estimation. However, intra-image and inter-image aggregation serve different roles: the former should propagate spatial evidence within each image while retaining structured two-dimensional dependencies, whereas the latter should gather information from plausible corresponding regions while limiting noise from unrelated locations. We propose WovenAnchor Matcher (WAM), a semi-dense matching framework based on role-specialized context aggregation. WAM introduces two complementary operators. WovenMamba performs intra-image aggregation through horizontal-then-vertical state-space propagation, allowing vertical updates to operate on horizontally contextualized features and thereby encouraging a structured two-dimensional receptive-field bias. AnchorCrossAttention performs inter-image aggregation by using cross-attention to gather context from plausible corresponding regions across the image pair. It performs this exchange over local representative anchors that aggregate feature evidence within each window. On MegaDepth, WAM achieves 66.1 pose AUC@5 $^\circ$ , improving on the JamMa baseline by 2.0 AUC points under the same evaluation setup. Ablations show that removing or replacing either specialized component reduces accuracy, providing evidence that intra-image propagation and inter-image retrieval benefit from role-specific aggregation designs. Code and pretrained weights are available at https://github.com/mmaiLab/WovenAnchorMatcher.

Figures & tables

Figure 1: Relative pose AUC@5 ∘ vs. runtime on MegaDepth. Runtime is measured on a single NVIDIA H100 GPU.

Panel of Figure 2 — Figure 2: Context aggregation designs. LoFTR uses separate self-/cross-attention, JamMa uses a shared JointMamba design, and WAM explicitly separates intra-image and inter-image aggregation into WovenMamba and AnchorCrossAttention. The diagrams are schematic summaries of the coarse context aggregation blocks rather than layer-by-layer implementations.

Figure 3: WovenMamba. The input is split into propagation and gate branches. The propagation branch is processed by horizontal-then-vertical SSM propagation, then gated, projected, and added back through a residual connection. The arrows illustrate the dominant horizontal and vertical scan orders; the implementation uses the corresponding row-major and column-major flattenings.

Figure 4: Receptive-field intuition under simplified linear models. The red circle marks the output position $(i,j)$ , and color indicates the normalized influence from each source position $(p,q)$ . Rows compare the additive scan-and-sum baseline $W^{\mathrm{sum}}$ , the dominant-path approximation $W^{\mathrm{ill}}$ , and the exact finite-grid kernel $W^{\mathrm{exact}}$ of the simplified WovenMamba composition. Columns show different one-dimensional decay settings: Single State, Long Horizontal, Long Vertical, and Two States.

Figure 5: AnchorCrossAttention. Anchor construction pools local window features into anchors using soft clustering. After cross-image attention, anchor-to-feature projection distributes the anchor updates using the same assignment probabilities. Overlap fusion combines the window-specific updates on the dense grid through a weighted Fold using normalized overlap weights.

Model family	Matcher	Params	Runtime	Relative pose error AUC $\uparrow$
Model family	Matcher	(M)	(ms)	@5 ∘	@10 ∘	@20 ∘
Sparse	XFeat [ Potje et al., 2024 ]	0.7	12.5	44.2	58.2	69.2
	SP+SuperGlue [ Sarlin et al., 2020 ]	13.3	55.2	57.6	72.6	83.5
	SP+LightGlue [ Lindenberger et al., 2023 ]	13.2	47.2	58.8	73.6	84.1
Dense	DKM [ Edstedt et al., 2023 ]	72.3	315.9	67.3	79.7	88.1
Dense	RoMa [ Edstedt et al., 2024 ]	111.3	470.2	68.5	80.6	88.8

Table 1: MegaDepth-1500 relative pose estimation. We report parameters, runtime, and pose-error AUC. Best and second-best semi-dense results are highlighted. XFeat ∗ denotes the semi-dense variant with match refinement. WAM † reports the checkpoint whose AUC@5 ∘ equals the median over seven runs; full repeated-run statistics are in Appendix E .

Matcher	ScanNet relative pose error AUC $\uparrow$			HPatches homography error AUC $\uparrow$
Matcher	@5 ∘	@10 ∘	@20 ∘	@3px	@5px	@10px
SP+SuperGlue [ Sarlin et al., 2020 ]	16.5	32.2	48.8	53.9	68.3	81.7
SP+LightGlue [ Lindenberger et al., 2023 ]	15.9	32.2	49.7	54.2	68.3	81.5
LoFTR [ Sun et al., 2021 ]	20.7	37.7	53.4	65.9	75.6	84.6
ASpanFormer [ Chen et al., 2022 ]	21.9	40.6	57.4	67.4	76.9	85.6
JamMa [ Lu and Du, 2025 ]	18.5	35.6	51.3	68.1	77.0	85.4

Table 2: Zero-shot transfer using MegaDepth-trained weights. ScanNet reports relative pose AUC at angular error thresholds; HPatches reports homography AUC at corner-error thresholds in pixels.

Variant	Runtime	Relative pose error AUC $\uparrow$
Variant	(ms)	@5 ∘	@10 ∘	@20 ∘
WovenAnchorModule (Default)	35.4	66.1	78.9	87.7
w/o WovenMamba	30.5	64.6	77.7	86.6
w/o AnchorCrossAttention	27.4	64.0	77.0	86.1
w/o both	20.9	62.0	75.0	84.7

Table 3: Component ablations on MegaDepth-1500. We report pose-error AUC under WAM component removals. Runtime is measured for the full matcher.

Table 9

Intra-image	Inter-image	AUC@5 ∘	AUC@10 ∘	AUC@20 ∘	Runtime (ms)
WovenMamba	ACA (default)	66.1	78.9	87.7	35.4
ASA	ACA	64.7	77.8	86.7	38.6
WovenMamba	CWM	65.4	78.3	87.4	33.9
ASA	CWM	64.9	77.7	86.7	36.0

Table 6: Operator-to-role assignment on MegaDepth-1500. ASA: AnchorSelfAttention; CWM: CrossWovenMamba. Runtime is measured for the full matcher.

Figure 6: Qualitative matching examples. We compare LoFTR, JamMa, and WAM. Green and red lines indicate geometrically consistent and inconsistent matches, respectively.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Variant	MegaDepth-1500			ScanNet-1500
Variant	AUC@5 ∘	AUC@10 ∘	AUC@20 ∘	AUC@5 ∘	AUC@10 ∘	AUC@20 ∘
w/o separators	66.0	78.8	87.7	22.5	40.6	56.5
with separators	66.1	78.9	87.7	23.2	41.2	57.2

Appendix

Table 7: Effect of WovenMamba separator tokens.

Component	WAM (ms)	JamMa (ms)
Backbone	5.6	5.9
Context module	14.0	2.9
Coarse matching	4.7	20.9
Fine matching	9.7	10.2
End-to-end total	35.4	43.8

Appendix

Table 8: Runtime breakdown on MegaDepth (ms). Stage timings exclude some overhead and therefore do not sum exactly to the end-to-end totals.

Stage	Representation	Spatial sampling
Before WovenAnchorModule	$104\times 104$ coarse features	Dense $1/8$ grid
WovenMamba	$104\times 104$ coarse features	Dense $1/8$ grid
ACA anchor construction	$8\times 8$ coarse cells/window	Dense $1/8$ grid
ACA cross-attention	$25\times 25$ windows, 4 anchors each	$1/32$ center spacing
ACA anchor-to-feature projection and overlap fusion	$104\times 104$ coarse features	Dense $1/8$ grid
Coarse matching	$104\times 104$ coarse features	Dense $1/8$ grid

Appendix

Table 9: Spatial representations in WAM. Grid sizes illustrate a padded

832\times 832

input with the default window and anchor settings.

Panel of Figure 7 — Figure 7: Local ERFs of trained intra-image context aggregation blocks. WovenMamba-1 (top) and OrthMamba-1 (bottom) for image pair 1. Brighter colors indicate greater relative sensitivity to the block input.

Panel of Figure 8 — Figure 8: Local ERFs of trained intra-image context aggregation blocks. WovenMamba-2 (top) and OrthMamba-2 (bottom) for image pair 1. Panel layout and color normalization follow the protocol in Appendix C .

Panel of Figure 9 — Figure 9: Local ERFs of trained intra-image context aggregation blocks. WovenMamba-1 (top) and OrthMamba-1 (bottom) for image pair 2. Panel layout and color normalization follow the protocol in Appendix C .

Panel of Figure 10 — Figure 10: Local ERFs of trained intra-image context aggregation blocks. WovenMamba-2 (top) and OrthMamba-2 (bottom) for image pair 2. Panel layout and color normalization follow the protocol in Appendix C .

Variant	AUC@5 ∘	AUC@10 ∘	AUC@20 ∘	Runtime (ms)
Default ACA	66.1	78.9	87.7	35.4
Without geometry features	65.8	78.7	87.6	35.1
Joint attention across anchor IDs	66.3	79.1	87.8	36.4
Uniform anchor assignments	65.7	78.7	87.4	34.7
Uniform overlap weights	65.7	78.8	87.7	34.8

Appendix

Table 10: ACA component ablations on MegaDepth-1500. Runtime is measured for the full matcher.

	ACA-1		ACA-2
Anchor $k$	Mean	Variance	Mean	Variance
0	0.251	0.008	0.217	0.012
1	0.249	0.007	0.267	0.042
2	0.253	0.008	0.232	0.020
3	0.247	0.006	0.284	0.047

Appendix

Table 11: Anchor assignment statistics on MegaDepth-1500. Mean and population variance of assignment probabilities in the two ACA blocks.

(a) Anchor count, with $w=8$
Setting	AUC@5 ∘	AUC@10 ∘	AUC@20 ∘	Runtime (ms)
$K=2$	65.5	78.4	87.3	32.9
$K=3$	65.7	78.6	87.4	33.8
$K=4$ (default)	66.1	78.9	87.7	35.4
$K=5$	65.7	78.8	87.6	36.1
$K=6$	65.7	78.6	87.5	37.1

Appendix

Table 12: Sensitivity to anchor count and window size on MegaDepth-1500. Runtime is measured for the full matcher.

Panel of Figure 11 — Figure 11: Learned anchor assignments in AnchorCrossAttention. ACA-1 (top) and ACA-2 (bottom) for image pair 1, window offset $(0,0)$ . Brighter heatmap colors indicate higher assignment probabilities.

Panel of Figure 12 — Figure 12: Learned anchor assignments in AnchorCrossAttention. ACA-1 (top) and ACA-2 (bottom) for image pair 1, window offset $(0,4)$ . See Appendix D.1 for visualization details.

Panel of Figure 13 — Figure 13: Learned anchor assignments in AnchorCrossAttention. ACA-1 (top) and ACA-2 (bottom) for image pair 1, window offset $(4,0)$ . See Appendix D.1 for visualization details.

Panel of Figure 14 — Figure 14: Learned anchor assignments in AnchorCrossAttention. ACA-1 (top) and ACA-2 (bottom) for image pair 1, window offset $(4,4)$ . See Appendix D.1 for visualization details.

Panel of Figure 15 — Figure 15: Learned anchor assignments in AnchorCrossAttention. ACA-1 (top) and ACA-2 (bottom) for image pair 2, window offset $(0,0)$ . See Appendix D.1 for visualization details.

Panel of Figure 16 — Figure 16: Learned anchor assignments in AnchorCrossAttention. ACA-1 (top) and ACA-2 (bottom) for image pair 2, window offset $(0,4)$ . See Appendix D.1 for visualization details.

Panel of Figure 17 — Figure 17: Learned anchor assignments in AnchorCrossAttention. ACA-1 (top) and ACA-2 (bottom) for image pair 2, window offset $(4,0)$ . See Appendix D.1 for visualization details.

Panel of Figure 18 — Figure 18: Learned anchor assignments in AnchorCrossAttention. ACA-1 (top) and ACA-2 (bottom) for image pair 2, window offset $(4,4)$ . See Appendix D.1 for visualization details.

Run	AUC@5 ∘	AUC@10 ∘	AUC@20 ∘	Runtime (ms)
A	66.4	78.9	87.5	34.9
B	66.3	79.0	87.7	35.0
C	66.2	78.9	87.6	35.0
D (default)	66.1	78.9	87.7	35.4
E	66.1	78.8	87.7	34.8
F	65.9	78.5	87.4	35.0

Appendix

Table 13: Repeated-run variability of WAM on MegaDepth-1500. Runtime is measured on a single NVIDIA H100 GPU under the protocol in Appendix A .

Input	Coarse	Method	Pose AUC $\uparrow$			Matcher time	Context time
size	size		@5 ∘	@10 ∘	@20 ∘	(ms)	(ms)
480	$60\times 60$	WAM	55.9	71.2	82.4	15.4	6.7
480	$60\times 60$	JamMa	53.3	68.2	79.2	16.0	2.8
640	$80\times 80$	WAM	63.7	76.7	86.3	22.0	8.8
640	$80\times 80$	JamMa	60.3	73.9	83.8	25.0	2.8
832	$104\times 104$	WAM	66.1	78.9	87.7	35.4	14.0

Appendix

Table 14: Input-resolution scalability on MegaDepth. We report pose AUC and runtime when varying the input resize-and-pad resolution. Context time measures WovenAnchorModule for WAM and JointMamba Module for JamMa.

Method	Day	Night
LoFTR [ Sun et al., 2021 ]	88.7 / 95.8 / 98.8	76.4 / 91.1 / 97.9
JamMa [ Lu and Du, 2025 ]	86.4 / 94.5 / 97.6	72.8 / 90.1 / 98.4
WAM (Ours)	87.5 / 94.5 / 98.2	74.3 / 91.1 / 97.9

Appendix

Table 15: Visual localization results on Aachen Day-Night v1.1. We report localization recall at

(0.25\,\mathrm{m},2^{\circ})

(0.5\,\mathrm{m},5^{\circ})

(5.0\,\mathrm{m},10^{\circ})

. All methods are evaluated under the same HLoc-based pipeline.

Asset	Use	License / terms	Citation
MegaDepth	Training and MegaDepth-1500 relative pose evaluation	CC BY 4.0 for released depth maps, SfM models, and related derived assets; original Internet images retain their original licenses	[ Li and Snavely, 2018 ]
ScanNet	Zero-shot ScanNet-1500 relative pose evaluation	ScanNet Terms of Use for data; MIT license for code	[ Dai et al., 2017 ]
HPatches	Zero-shot homography evaluation	MIT license for the HPatches dataset repository; BSD-2-Clause license for the HPatches benchmark/evaluation code; no dataset assets are redistributed	[ Balntas et al., 2017 ]
Aachen Day-Night v1.1	Visual localization evaluation	Official dataset/benchmark terms as specified in the accompanying README files; no dataset assets are redistributed	[ Sattler et al., 2018 , Zhang et al., 2021 ]
HLoc	Visual localization pipeline	Apache-2.0 license for the HLoc codebase; third-party components follow their own licenses	[ Sarlin et al., 2019 ]
JamMa codebase	Base implementation adapted for WAM	JamMa code is released under MIT license. Its repository notes that parts of the code are based on upstream LoFTR-family repositories under their own licenses, terms, and conditions; we preserve applicable notices and do not redistribute third-party checkpoints	[ Lu and Du, 2025 , Sun et al., 2021 ]

Appendix

Table 16: Existing assets and licenses/terms.

Local Feature Matching

Local feature matching finds corresponding points or regions across images to support geometric estimation, localization, reconstruction, and recognition.

See all 36 papers

RoMa v2: Harder Better Faster Denser Feature Matching
Nov 19, 2025Johan Edstedt, David Nordström, Yushan Zhang +7Image Matching
Linköping University · Chalmers University of Technology · University of Amsterdam +1
SAMatcher: Co-Visibility Modeling with Segment Anything for Robust Feature Matching
Jun 2, 2026Xu Pan, Qiyuan Ma, Mingyue Dong +3Multi-View Geometry Image Matching
State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing (LIESMARS), Wuhan University, Wuhan, P. R. China · National Key Laboratory of Space Target Awareness, Space Engineering University, Beijing, P. R. China
REDI-Match: Rotation-Equivariant Distillation for Efficient and Robust Dense Matching
Jun 23, 2026Yinji Ge, Guixu Zheng, Wulong Guo +5Image Matching Vision Foundation Model Adaptation
1Tsinghua University · 2Southern University of Science and Technology · 3Beihang University +1

Nov 19, 2025J/K move · Enter open · S save

RoMa v2: Harder Better Faster Denser Feature Matching

Johan Edstedt, David Nordström, Yushan Zhang, Georg Bökman+6

Linköping University · Chalmers University of Technology · University of Amsterdam+1

Image Matching

Open paper AskPDF arXiv

Dense feature matching aims to estimate all correspondences between two images of a 3D scene and has recently been established as the gold standard due to its high accuracy and robustness. However, existing dense matchers still fail or perform poorly for many hard real-world scenarios, and high-precision models are often slow, limiting their applicability. In this paper, we attack these weaknesses on a wide front through a series of systematic improvements that together yield a significantly better model. In particular, we construct a novel matching architecture and loss, which, combined with a curated diverse training distribution, enables our model to solve many complex matching tasks. We further make training faster through a decoupled two-stage matching-then-refinement pipeline, and at the same time, significantly reduce refinement memory usage through a custom CUDA kernel. Finally, we leverage the recent DINOv3 foundation model along with multiple other insights to make the model more robust and unbiased. In our extensive set of experiments, we show that the resulting novel matcher sets a new state-of-the-art, being significantly more accurate than its predecessors. Code is available at https://github.com/Parskatt/romav2

Abstract

Local Feature Matching

Explore similar work

RoMa v2: Harder Better Faster Denser Feature Matching

SAMatcher: Co-Visibility Modeling with Segment Anything for Robust Feature Matching

REDI-Match: Rotation-Equivariant Distillation for Efficient and Robust Dense Matching