Research map / Detection and segmentation / CNN features
WovenAnchor Matcher: Specialized Intra- and Inter-Image Context Modeling for Feature Matching
Zhiyang Li, Ruijiang Jin, Thibaut Klenke, Yusuke Sekikawa, Nakamasa Inoue
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WovenAnchor Matcher, a semi-dense feature matcher that separates intra-image context modeling (preserving 2D structure) from inter-image aggregation (gathering from plausible corresponding regions while limiting noise).
Problem
Semi-dense local feature matching aggregates context before coarse-to-fine correspondence estimation, but intra-image and inter-image aggregation serve different roles and are commonly treated uniformly.
Approach
Specialize the context modeling per role: intra-image aggregation propagates spatial evidence while retaining structured 2D dependencies, and inter-image aggregation gathers information only from plausible corresponding regions to limit noise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add WovenMamba intra-image aggregation (ablation) | WAM · MegaDepth-1500 | pose AUC@5deg: 64.6 -> 66.1 | Improved |
| add AnchorCrossAttention inter-image aggregation (ablation) | WAM · MegaDepth-1500 | pose AUC@5deg: 64.0 -> 66.1 | Improved |
| use both specialized components instead of neither (ablation) | WAM · MegaDepth-1500 | pose AUC@5deg: 62.0 -> 66.1 | Improved |
| use horizontal-then-vertical WovenMamba instead of independent directional scans (FourMamba) (ablation) | WAM intra-image module · MegaDepth-1500 | pose AUC@5deg: 65.3 -> 66.1 | Improved |
| use WovenMamba instead of self-attention for intra-image (ablation) | WAM · MegaDepth-1500 | pose AUC@5deg: 64.8 -> 66.1 | Improved |
| use anchor cross-attention instead of dense cross-attention (ablation) | WAM inter-image module · MegaDepth-1500 | pose AUC@5deg: 65.0 -> 66.1 | Improved |
| assign WovenMamba to intra and ACA to inter roles instead of reversed or shared roles (ablation) | WAM · MegaDepth-1500 | pose AUC@5deg: 64.9 -> 66.1 | Improved |
| use learned instead of uniform anchor assignments (ablation) | ACA · MegaDepth-1500 | pose AUC@5deg: 65.7 -> 66.1 | Improved |
| add anchor geometry features (ablation) | ACA · MegaDepth-1500 | pose AUC@5deg: 65.8 -> 66.1 | Improved |
| attend jointly across all anchor IDs (ablation) | ACA · MegaDepth-1500 | pose AUC@5deg: 66.1 -> 66.3 | Improved |
| increase anchors per window beyond 4 (ablation) | ACA · MegaDepth-1500 | pose AUC@5deg: 66.1 -> 65.7 | No clear effect |
| add WovenMamba separator tokens (ablation) | WAM · ScanNet-1500 zero-shot | pose AUC@5deg: 22.5 -> 23.2 | Improved |
Closest papers by meaning
- Unifying Feature and Cost Aggregation with Transformers for Semantic and Visual Correspondence (ICLR 2024 · similarity 0.87)
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.86)
- MATCHA: Towards Matching Anything (CVPR 2025 · similarity 0.85)
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · similarity 0.85)
- Context-I2W: Mapping Images to Context-Dependent Words for Accurate Zero-Shot Composed Image Retrieval (AAAI 2024 · similarity 0.85)
- GIM: Learning Generalizable Image Matcher From Internet Videos (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
