high performance computing on graphics processing units: hgpu.org

In this work we present a novel GPU based Visual Programming Language for Computer Vision and Image Processing systems. Many vision algorithms have been shown to perform better on GPUs. However, one of the current drawbacks is the need for considerable GPU programming expertise. We propose an abstraction over GPU implementation details by providing an intuitive graphical programming interface. Though several libraries have abstracted GPGPU programming such as CUDA, we believe that VisDom can combine some of these abstraction in a graphical form, thus making it even easier for developing vision based systems on GPUs. The basic building block of our system is a Cell, which encapsulates primitive and high-level vision algorithms. The concept of Cells and a novel programming environment makes it easier to design and view complex vision systems. In conclusion, we present two examples to illustrate VisDom in action – optical flow on a color segmented image and a face tracking application.