Geometric Verification Method of Best Score Increasing Subsequence for Object Instance Recognition

Abstract

Weighted Longest Increasing Subsequence (WLIS) and its improvement, Best Increasing Subsequence (BIS) are two methods that has been proposed for pair verification in object instance recognition using local features. Tested on the Stanford Mobile Visual Dataset (SMVS), the BIS achieves better performance than WLIS on most categories, except for the “video frames” category. In this paper we propose several modifications to BIS which resulted in a better overall performance compared to the WLIS and the basic BIS approaches. On average, the proposed Best Score Increasing Subsequence (BSIS) performs 4.53% better than the BIS and 9.43% better than the WLIS.