Accelerating local feature extraction using OpenCL on heterogeneous platforms

Local feature extraction is one of the most important steps in image processing applications such as image matching and object recognition. The Scale Invariant Feature Transformation (SIFT) algorithm is one of the most robust as well as one of the most computation intensive algorithms to extract local features. Recent implementations of the algorithm focus on homogeneous processors like multi-core CPUs or many-core GPUs. In this paper, we introduce an OpenCL-based implementation, which can be used in homogeneous and heterogeneous CPU/GPU environments. We analyze possible coarse-grained and fine-grained parallelization solutions of the SIFT algorithm. Using a set of optimizations we implement a high-performance SIFT implementations for very different CPU/GPU architectures. The scalable implementation allows for a fast processing, more than 40 FPS for Full-HD images.