FlexGlue paper: sparse-to-dense matching beats LoMa on SfM, hits 74.62 on IMC2025
ducha_aiki · x · 2026-11-05
Researchers including Philipp Lindenberger, Paul-Edouard Sarlin and Marc Pollefeys released FlexGlue, an open-source network for asymmetric sparse-to-dense feature matching and visual disambiguation.
- A Transformer fuses dense DINOv3 tokens from two CNN-encoded images; queries (grid or keypoints) are matched against dense pixels, with an overlap classifier rejecting non-matching pairs — critical for SfM, where visual symmetries fool existing matchers.
- Combines sub-pixel accuracy of dense matching with sparse-matching speed; trained with negative pairs.
- Notable details: softmax replaced by 1.5-entmax with a sparse entmax loss for exact-zero probabilities and better sub-pixel accuracy; a small MLP predicts sub-cell offsets, shifting matches only 0.6px on average and scaling linearly with queries.
- Results: pushes the Pareto frontier in relative pose estimation; on IMC2025 SfM with distractors it scores 66.04 (4k keypoints) and 74.62 (10k), far above all LoMa variants — LoMa-B at 10k times out on Kaggle.
Related event: LightGlue Team Releases FlexGlue, Topping IMC2025 Benchmarks(5 posts)→
More from Research
- MIT's UQ-DAAAM helps robots spot and fix unreliable object memories in real time — MacrinePhD · 2026-11-05
- Fixed seeds don't make scientific ML reliable: MAE swings 0.10–0.15 eV across splits — bravo_abad · 2026-11-05
- Venomics AI raises $2.7M to mine venom biodiversity for drug discovery — adityaag · 2026-11-05
- UniRestorer++ scales universal image restoration with multi-granularity degradation modeling, beating dense and MoE baselines — Jingbo Lin · 2026-11-05
- GaussianRenderFormer: Hebrew U team renders 3D Gaussian splats with a pure transformer, no rasterizer, hitting 38.08 dB — HUJI-IL · 2026-11-05
- Yacine probes mingru hidden states to check if domain randomization captured the real world — yacineMTB · 2026-11-05