FlexGlue: Sparse-to-Dense Feature Matching and Visual Disambiguation

TL;DR

FlexGlue introduces a neural network for asymmetric sparse-to-dense feature matching that combines the speed of sparse matching with the accuracy of dense matching, while simultaneously rejecting non-matching image pairs. It achieves high precision in disambiguation by leveraging visual symmetry cues, improving relative pose estimation and visual localization.

Why It Matters

This approach significantly enhances Structure-from-Motion and visual localization by reducing false matches due to symmetry, which is a major challenge in real-world applications. Its efficiency and robustness make it suitable for large-scale 3D reconstruction and robotics systems.

Key Findings

• FlexGlue achieves sub-pixel accuracy in dense matching while maintaining sparse matching speed.
• It effectively rejects non-matching image pairs with high precision, addressing visual symmetry issues.
• It outperforms state-of-the-art methods on multiple benchmarks for relative pose estimation and visual localization.