FlexGlue: sparse-to-dense matching that beats dense accuracy at a fraction of the cost
zhenjun_zhao · x · 2026-10-27
ETH Zurich researchers (Lindenberger, Sarlin, Miksik, Pollefeys) released FlexGlue (arXiv 2610.32976), a network for asymmetric sparse-to-dense feature matching with built-in disambiguation — tl;dr: LightGlue plus pair-level covisibility.
Problem: sparse matching is fast but limited by repeatability; dense matching is robust but costly; both fail to reject visually symmetric non-matching pairs, which matters for SfM.
Method: CNN-encoded dense tokens fused by a Transformer first classify whether the pair is matchable (early exit otherwise), then interpolate features, search dense correlations, and refine matches to sub-pixel accuracy.
Results:
- Pushes the Pareto frontier on MegaDepth-1500 and ScanNet-1500 — near-dense accuracy at 31ms extraction+matching vs 239ms for RoMa;
- Significantly outperforms prior work on SfM (including with distractors on IMC2025) and visual localization;
- Best doppelganger detection, +10 AUROC on LaMAR over dedicated methods, at higher throughput; top accuracy across ground-aerial and other domains.
Code is publicly available.
More from Research
- UK AISI: under $80 of fine-tuning fully restores filtered bio capabilities in open-weight models — connoraxiotes · 2026-10-27
- RFD4-Proteina runs on an RTX 3060: protein design from $0.001 per design — AllThingsApx · 2026-10-27
- University of Freiburg's Generative Intelligence Lab recruiting funded PhDs via ELLIS, apply by Oct 31 — yuyinzhou_cs · 2026-10-27
- Nature Physics: reprogrammable refractive index enables arbitrary multimode wave control for ML — MarioKrenn6240 · 2026-10-27
- Wakaru open-sources ontologize: labeling datasets from model activations — amplifiedamp · 2026-10-27
- EMNLP Oral paper traces multilingual representations in LLMs with cross-layer transcoders — ZhijingJin · 2026-10-27