Parallelizing Scale Invariant Feature Transform on a Distributed Memory Cluster

Scale Invariant Feature Transform (SIFT) is a computer vision algorithm that is widely-used to extract features from images. We explored accelerating an existing implementation of this algorithm with message passing in order to analyze large data sets. We successfully tested two approaches to data decomposition in order to parallelize SIFT on a distributed memory cluster. Introduction In certain domains, it is very useful to extract information about objects in images. A specific domain, geospatial sciences, is facing the problem of ever increasing high resolution data. streams of data from satellites, unmanned aerial vehicles, airplanes, and people need to be accurately georeferenced and registered. Using conventional methods, including desktop computers that run serial programs, to analyze this data takes too long or requires more resources than a single desktop contains. Parallel cluster computing provides more resources than a desktop and allows processing of different parts of the problem at the same time. Using parallel processing, it is possible to solve the problem of analyzing large sets of geospatial data. Manual time-consuming tasks like image mosaicking, stitching, alignment, and matching of geospatial data collected by multiple sensors can be made autonomous by the use of computer vision algorithms such as scale Invariant Feature transform (sIFt). these techniques are extensively used in geospatial sciences. specifically, there exists a need to take an input image from a user, analyze and describe it, and finally match the image to a known location that has been georeferenced. the work presented here is part of a larger project that is building a system that uses computer vision techniques, databases, and algorithms to quickly and autonomously solve certain geospatial science problems like georeferencing and registering new and existing Geospatial Information systems (GIs) data. the GIs data sets that motivate this parallel implementation are terabytes in size. A single image may be larger than the memory of a single node, hence the need to extract features and descriptors from an image in parallel. Also, as output of data from different sensors increases, the amount of data that needs to be processed in a timely manner will increase. this article describes ways to implement a distributed memory parallel version of a popular computer vision algorithm scale Invariant Feature transform (sIFt) using the Message Passing Interface (MPI) library in order to solve the problem of timely analysis of large GIs data sets for which the original implementation of sIFt was not designed. there have been successful prior parallel implementations of sIFt, but they are geared toward real-time processing of small data whereas this implementation emphasizes scalability and capacity computing.