high performance computing on graphics processing units: hgpu.org

Scale Invariant Feature Transform (SIFT) is a popular image feature extraction algorithm. SIFT’s features are invariant to many image related variables including scale and change in viewpoint. Despite its broad capabilities, it is computationally expensive. This characteristic makes it hard for researchers to use SIFT in their works especially in real time application. This is a common problem with many image-processing related algorithm. Utilizing graphical processing unit (GPU) through parallel programming is an affordable solution for this issue. In this paper we present a GPU-based implementation of SIFT using Compute Unified Device Architecture (CUDA) programming framework. We compare our CUDA-based implementation, namely siftCU, with CPU-based serial implementations of SIFT both in feature matching accuracy and time consumption. Results show our implementation can gain 4x speed up over serial CPU implementation even though we have used a low end graphic card while using a powerful CPU for test platform.