Investigating large-scale feature matching using the Intel® Xeon Phi™ coprocessor

Many computer vision applications are entering the 'big data' era: it is straightforward to acquire very large datasets that need to be processed. Our current research targets a large-scale structure-from-motion application, in which 3D models are formed from large collections of digital photographs. There have also been many recent technological developments suitable for speeding up the data processing for these computer vision applications. However many of the emerging technologies have very different costs in terms of developer time and experience. We have previously implemented our system on multicore CPUs, clusters of such multicore machines, and GPUs. The Intel (R) Xeon Phi (TM) coprocessor aims to provide highly efficient processing of massively parallel workloads. The Phi tries to strike a pragmatic balance between the vector processing power of GPUs, and the ease of programming provided by deploying to CPUs. Very recently, some Phi coprocessors have been made available through the New Zealand eScience Infrastructure (NeSI) facilities. This paper reports on our initial findings porting and running part of our processing pipeline on the Intel (R) Xeon Phi (TM)