high performance computing on graphics processing units: hgpu.org

In recent years, we have seen several approaches to implement hardware-accelerated multi-view stereo (MVS) algorithms employing the graphics processing unit (GPU) for fast and parallel computation. To our knowledge, all of them resort to various rendering passes to perform their computations. In contrast, modern GPU compute frameworks give access to the massively parallel compute capability of a GPU without forcing users to express their computations as rendering passes. This allows for a broader class of algorithms to be executed on a GPU and improves flexibility and portability. We implemented a region-growing MVS approach for NVIDIA’s CUDA framework, tested it on several data sets, and compared its computation time and reconstruction quality to an existing CPU version. For comparable reconstruction quality we often achieve a moderate speedup which is mainly limited by scarce register and memory resources of the current GPU generation.