high performance computing on graphics processing units: hgpu.org

Face recognition is a specialized case of object recognition, and has broad applications in security, surveillance, identity management, law enforcement, human-computer interaction, and automatic photo and video indexing. Because human faces occupy a narrow portion of the total image space, specialized methods are required to identify faces based on subtle differences. One such method is the Eigenfaces classifier, a holistic statistical method which significantly reduces the dimensionality of the search space. Despite the dimensionality reduction, the k-nearest neighbor (kNN) search remains a bottleneck in this application as well as most others that use it. The kNN search is both computationally complex and highly data-parallel, making it a candidate for acceleration on graphics hardware. I present an implementation of an Eigenfaces classifier with a portion of the kNN search accelerated on an Nvidia Tesla K20X GPU with 2688 cores.