Distributed Kernelized Locality-Sensitive Hashing for Faster Image Based Navigation

Content based image retrieval (CBIR) remains one of the most heavily researched areas in computer vision. Different image retrieval techniques and algorithms have been implemented and used in localization research, object recognition applications, and commercially by companies such as Facebook, Google, and Yahoo!. Current methods for image retrieval become problematic when implemented on image datasets that can easily reach billions of images. In order to process extremely large datasets, the computation must be distributed across a cluster of machines using software such as Apache Hadoop. There are many different algorithms for conducting content based image retrieval, but this research focuses on Kernelized Locality-Sensitive Hashing (KLSH). For the first time, a distributed implementation of the KLSH algorithm using the MapReduce programming paradigm performs CBIR and localization using an urban environment image dataset. This new distributed algorithm is shown to be 4.8 times faster than a brute force linear search while still maintaining localization accuracy within 8.5 meters