A Local Sparse Model for Matching Problem

Abstract

Feature matching problem that incorporates pairwise constraints is usually formulated as a quadratic assignment problem (QAP). Since it is NP-hard, relaxation models are required. In this paper, we first formulate the QAP from the match selection point of view; and then propose a local sparse model for matching problem. Our local sparse matching (LSM) method has the following advantages: (1) It is parameter-free; (2) It generates a local sparse solution which is closer to a discrete matrix than most other continuous relaxation methods for the matching problem. (3) The one-to-one matching constraints are better maintained in LSM solution. Promising experimental results show the effectiveness of the Proposed LSM method.