Research on Fast Matching of Large Scale Binary Descriptions
Abstract：Binary descriptor is becoming a hot research topic in related areas due to its potential in large scale visual computing owing to its low memory footprint and efficient Hamming distance computation. However, there are no satisfactory solutions to fast nearest neighbor matching of large scale binary descriptors up to now. The baseline linear search method is extremely slow for large database, limiting its potential for large scale applications. To deal with this problem, this proposal aims to conduct research on two different kinds of fast nearest neighbor matching methods respectively, e.g., fast approximate nearest neighbor matching and exact nearest neighbor matching. More specifically, for approximate nearest neighbor matching, we research on methods for hash key selection and multiple hash tables learning, so as to select only a few positions from the original binary feature to generate a hash table, based on which an approximate nearest neighbor matching method could has high accuracy and high matching speed. Moreover, we study how to learn multiple hash tables with complementary properties to further improve performance. For exact nearest neighbor matching, we research on data
structures of binary descriptors with efficient store and traverse characteristic as well as machine learning based multi-index hash methods to reduce the access times of unnecessary data, hence improve matching speed.
Keywords: image matching; image localization
Contact:
FAN Bin
E-mail: bfan@nlpr.ia.ac.cn