Performance and energy characterization of high-performance low-cost cornerness detection on GPUs and multicores

Feature detection and tracking is an important problem in Computer Vision. Corners in an image are a good indication of features to track. Original algorithms may be expensive even on multicore architectures because they require full convolutions to be performed. Although these can be performed in real time in modern GPUs and multicore CPUs, faster solutions are needed for embedded systems and complex algorithms, given that corner detections is just a step of the analysis process. In this paper we evaluate the performance and energy efficiency of the Harris corner detection algorithm as well as an approximation of it, in both desktop and mobile platforms. The purpose of this paper is three-fold: evaluate the performance gains of GPUs vs. CPUs for several mobile and desktop systems, evaluate whether the Harris approximation provides adequate performance gains to justify its use in mobile and desktop system configurations and, finally, determine which configurations provide real-time performance. According to our evaluation (a) the best GPU solution is 16.3 times faster than the best CPU solution for the desktop case while being 2.6 times more energy efficient and (b) the best GPU solution for the mobile case is 1.2 times faster while being 3.6 times more energy efficient than the respective CPU.