Texture-less Object Recognition Using Contour Fragment-Based Features with Bisected Local Regions

Abstract

This paper presents a new texture-less object recognition algorithm using contour fragment-based features with bisected local regions. We propose a new feature descriptor which is computed on a bisected local region around each feature point. For each feature point, the local region is divided into two bisected regions according to the tangent line of the contour fragment, and then the proposed feature descriptor is computed on each of the bisected local regions. In our feature matching, we remove spurious potential matches by selecting top-k minimum cost image features for each model node, and the survived potential matches are refined by spectral matching algorithm using pairwise geometric interaction between features.