WovenAnchor Matcher splits intra/inter-image context modeling, topping feature matching with only 4.5M params
zhenjun_zhao · x · 2026-10-24
The WovenAnchor Matcher (WAM) paper explicitly separates feature-matching context aggregation into specialized intra-image (WovenMamba) and inter-image (AnchorCrossAttention) modules, contrasting with LoFTR's self/cross-attention and JamMa's shared Mamba design.
- Achieves 66.1 AUC@5° on MegaDepth-1500 relative pose estimation, beating semi-dense rivals like LoFTR, JamMa and EDM with only 4.5M parameters at 35.4ms runtime
- Also leads zero-shot transfer with 23.2/41.2 AUC@5°/10° on ScanNet
More from Research
- SpecTr-GBV unifies multi-draft generation and greedy block verification in speculative decoding — jiqizhixin · 2026-10-24
- Dev experiments with autoencoders over functions; low-dim search may be key to local FunSearch — cephaloform · 2026-10-24
- Jon Barron details his workflow: Gemini runs experiments, drafts the paper, then rewrites in his style — jon_barron · 2026-10-24
- Jon Barron: every softmax/cross-entropy implementation has two subtle numerical bugs — jon_barron · 2026-10-24
- Marin's 535B pretraining passes 80%, loss runs below scaling-ladder forecasts — dlwh · 2026-10-24
- Annotated example breaks down what each paragraph of a great paper's discussion section should do — Faheem_uh · 2026-10-24