Massively Parallel Implementation of a Multi-View Stereo Algorithm

This thesis deals with the implementation of a multi-view stereo algorithm in a massively parallel way. The nature of multi-view stereo algorithms favors a parallel approach, as these algorithms operate on multiple images and in these images on multiple pixels. This work focuses on the matching optimization, of a region growing technique, in the multiview stereo approach by Goesele et al. which is the most computationally intensive step of the algorithm. Starting from a sparse 3D point cloud reconstructed using structure-from-motion methods, a nonlinear optimization is employed using a photoconsistency measure. Building upon an existing CPU implementation, an implementation for the matching optimization is written for NVIDIA graphics hardware which is essentially a massively parallel computing device. To access the hardware we make use of the Compute Unified Device Architecture (CUDA), a technology for General-Purpose computation on GPUs (GPGPU). Subsequently the GPU implementation is compared with the CPU version with regard to reconstruction quality and processing time. Furthermore different optimizations for exploiting the GPU architecture (and therefore reducing the processing time) are discussed and evaluated