Detección de puntos claves mediante SIFT paralelizado en GPU

Abstract

This paper presents an optimization method for detecting SIFT points (Scale-invariant feature transform), by using a GPU parallelization, taking advantage of multiple cores of it, to divide the processes using the API's CUDA. The goal is to accelerate the computation time, which is a critical variable for the entire process of key-point's detection. The strategy used is based on two assumptions, the load balance and distribution of calculation. Each thread will perform the operations required for calculating SIFT and obtain the necessary descriptors according to an appropriate threshold. Parallelizing the process of assignment of orientation, which consists of the accumulation of all the orientations concerning a region of a key-point, assigned to each pixel in the window a thread, centered on the location of a key point, was worked inside SIFT. The tests were performed with a Notebook with Core 2 Duo 2.2GHz, 3GB of RAM and a GeForce 8600GT VGA (32 Cores) 512MB. The results obtained show that performance is achieved in terms of speed of the order 42.5 millisecond on average, considering all tests and all resolutions worked (320×240, 480×360, 640×480, 800×600, 1024×768, 1280×960), where the parallelization of SIFT, shows no significant loss of key points, compared to the sequential version.