High Efficient Local Feature Matching

Local features have proven successful in image matching. The goal of this paper is to find a way to match the local feature efficiently. We firstly propose a novel local feature descriptor and then a new index structure for searching in the high-dimensional feature space. The proposed descriptor is based on the gradient distance and orientation histogram (GDOH). GDOH has only half the dimensional size of SIFT descriptor, yet still maintain distinctness and robustness as much as SIFT. Secondly, we present a new index structure on dimensions of the feature vector for KNN(k nearest neighbor) search, called iDs(indexing the dimensions). iDs outperforms BBF(best bin first) algorithm both in accuracy and speed in high-dimensional feature vector search. The experimental results demonstrate that our scheme, jointly use GDOH and iDs, can result in high efficiency in local feature matching.