high performance computing on graphics processing units: hgpu.org

In this paper, we present a novel parallel implementation of HARRIS and DOG detector on GPU for feature extractions of Patch-based Multi-View Stereo (PMVS) algorithm in image sequence. With the Compute Unified Device Architecture(CUDA)-enabled GPU, the acceleration is significant and it obtains a 34 times performance boost comparing to a CPU implementation. We adopt the hardware built-in bilinear interpolation of texture to shorten the time for image resampling while doing image transformations. While lots of time can be saved by using our improved PMVS Algorithm with GPU-based feature extraction, experimental results also show that our implementation can obtain fine details for building accurate object and scene models.