Accelerated multi-view stereo using parallel processing capababilities of the GPUS

This paper presents an accelerated implementation of a multi-view stereo pipeline using parallel processing capabilities of the GPUs. Our system takes as input a set of calibrated photographs and produces a textured 3D mesh of the scene. The pipeline is divided into three parts: dense stereo matching, multi-view correspondence linking and 3D model generation. First, we use a combined vertical aggregation and dynamic programming (DP) scheme to produce disparity maps between pairs of photographs. Then, the depth maps are computed using a multi-view correspondence linking algorithm. Finally, we use a Delaunay triangulation algorithm and texture mapping to produce the 3D model of the scene.