Toward Free-Form Local Feature Matching.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Existing feature matching methods are limited by fixed position priors (e.g., keypoints, grids).
• These priors restrict matching point distribution and introduce limitations like reliance on keypoint repeatability or lack of texture precision.

Purpose of the Study:

• To develop a novel free-form feature matching paradigm (RCM+) that decouples from position priors.
• To enable flexible, zero-shot matching of arbitrary input positions without retraining.

Main Methods:

• Introduced RCM+, a free-form matching paradigm using a position-agnostic encoder and parameter-free decoder.
• Developed the Balancer to reconcile multiple position priors for improved point distribution.
• Enhanced existing view switcher and conflict-free matching layers from prior work (RCM).

Main Results:

• RCM+ demonstrates exceptional flexibility, matching diverse input types (keypoints, lines, grids, etc.) in a zero-shot manner.
• The Balancer improves point distribution for downstream tasks.
• Experiments confirm RCM+'s excellent performance, efficiency, and adaptability.

Conclusions:

• RCM+ overcomes limitations of prior methods by decoupling from position priors, offering unprecedented flexibility.
• The free-form matching paradigm allows users to leverage various priors without retraining, adapting to specific scene properties.
• RCM+ shows significant promise for diverse computer vision applications requiring robust and flexible feature matching.