high performance computing on graphics processing units: hgpu.org

Since 2011, University of Michigan-Shanghai Jiao Tong University Joint Institute (JI) has established 122 corporate-sponsored Capstone Design Projects (CDPs) with world leading companies such as Covidien, General Electric, Hewlett Packard, Intel, and Siemens. Of these corporations, Intel was the first sponsor, having funded 21 projects and mentored 105 students over four consecutive years. This paper is a longitudinal case study following three Intel-sponsored multi-view real-time face detection CDP teams with 15 undergraduate students during 2013 and 2014. On the technical side, the system design of face detection is based on Intel High Definition (HD) 4000 graphics and OpenCL. With numerous techniques, including accelerated pipeline over CPU and GPU, image decomposition, two-dimensional (2D) task allocation, and a combination of Viola-Jones algorithm and continuously adaptive mean-shift (Camshift) algorithm, a speed of 32 fps was attained for real-time multi-view face detection. In addition, a frontal view detection accuracy of 81% was achieved in Phase I and a multi-view detection accuracy of 95% in Phase III. Furthermore, an innovative application called face-detection game controller (FDGC) was developed. On the other side, this research also addresses benefitsof stakeholders. After graduation, a third(5) of these students worked in multinational corporations (MNCs) and two thirds (10) of these students entered top American graduate schools. At the time of this writing, five of them have finished their master’s degrees and are currently working for famous companies, such as Amazon, Facebook, and Google.