Image Feature Matching and Its Parallelization Using OpenMP

Parallel Computing has been gaining interest nowadays due to physical constraints preventing frequency scaling. Therefore, in order to achieve high performance on multicore systems, programmers need to focus on parallelizing their programs. Although there are many available parallelized APIs written by experts that should improve coding, they do not automatically guarantee good performance. This paper focuses on understanding factors that help to achieve better performance. This paper specifically focuses on SIFT based feature matching for various image collections. SIFT works by extracting information from images and then later compares this information for feature matching. Since SIFT performs CPU intensive computations, it requires parallelization in order to be feasible. This paper focuses on exploring a shared memory model based API known as Open multi-processing (OpenMP) to accelerate existing serial SIFT based image matching. In this paper, we were able to achieve a speedup of ~2× by using various OpenMP features. The paper later discusses factors like scalability and speedup and how multi-threading impacts them. We also discovered the importance of different levels of parallelism and their effects on performance.