Finding correspondences between local features is a fundamental building block of many computer vision applications.
Why it is Important
- Discretizing an image into a sparse set of points (local features) increases robustness and scalability of downstream applications.
- We can utilize local and global context to find an assignment matrix between points in two images.
- Point correspondences can be used for robust pose estimation and 3D reconstruction.
Feature Detection and Description: Local features need to be accurately and repeatably detected and described within images.
Feature Matching: After detection, features can be matched across views. These correspondences can be used for 2D/3D geometry estimation.
Applications: Local feature matches are widely used in many computer vision tasks, such as:
- Relative pose estimation
- 3D reconstruction
- Visual localization
Point features are at the core of 3D mapping and localization systems. They can be reliably detected and matched in many scenarios, and the inherent discretization allows mapping systems to scale to thousands of images.
- LightGlue: Local Feature matching at Light Speed (ICCV 2023) [Paper]
- GlueStick: Robust Image Matching by Sticking Points and Lines Together (ICCV 2023) [Project page]
- Online Invariance Selection for Local Feature Descriptors (ECCV 2020) [Project page]
- AdaLAM: Revisiting Handcrafted Outlier Detection (ECCV2020) [Paper]