Improving Performances of MSER Features in Matching and Retrieval Tasks

Abstract

MSER features are redefined to improve their performances in matching and retrieval tasks. The proposed SIMSER features (i.e. scale-insensitive MSERs) are the extremal regions which are maximally stable not only under the threshold changes (like MSERs) but, additionally, under image rescaling (smoothing). Theoretical advantages of such a modification are discussed. It is also preliminarily verified experimentally that such a modification preserves the fundamental properties of MSERs, i.e. the average numbers of features, repeatability, and computational complexity (which is only multiplicatively increased by the number of scales used), while performances (measured by typical CBVIR metrics) can be significantly improved. In particular, results on benchmark datasets indicate significant increments in recall values, both for descriptor-based matching and word-based matching. In general, SIMSERs seem particularly suitable for a usage with large visual vocabularies, e.g. they can be prospectively applied to improve quality of BoW pre-retrieval operations in large-scale databases.