high performance computing on graphics processing units: hgpu.org

Graph cuts have found many applications that address the problem of energy minimization, which occur frequently in computer vision and image processing. One of the most common applications is binary image segmentation, or silhouette extraction. Image segmentation is the process of applying a labeling to each pixel in an image to determine a list of boundaries of objects, areas of interest, or silhouettes of objects in the scene. The resulting pixel labeling enables higher level vision systems to perform complex actions such as gesture recognition for human-computer interfaces Ueda et al. (2003), real-time 3D volume reconstructions Laurentini (1994), and mixed reality applications Takaya et al. (2009). Without accurate and coherent segmentations, these higher end-to-end systems will fail to perform their required functions. This necessitates a method which can correctly identify and extract objects from a scene with high accuracy. However, accuracy usually comes with serious performance penalties, particularly in image processing which examines hundreds of thousands of pixels in each image. This is unacceptable for high-level vision systems, which generally require the processing to be completed in real time.