Large-Scale Dense 3D Reconstruction from Stereo Imagery

In this paper we propose a novel method for largescale dense 3D reconstruction from stereo imagery. Assuming that stereo camera calibration and camera motion are known, our method is able to reconstruct accurately dense 3D models of urban environments in the form of point clouds. We take advantage of recent stereo matching techniques that are able to build dense and accurate disparity maps from two rectified images. Then, we fuse the information from multiple disparity maps into a global model by using an efficient data association technique that takes into account stereo uncertainty and performs geometric and photometric consistency validation in a multi-view setup. Finally, we use efficient voxel grid filtering techniques to deal with storage requirements in large-scale environments. In addition, our method automatically discards possible moving obstacles in the scene. We show experimental results on real video large-scale sequences and compare our approach with respect to other state-of-the-art methods such as PMVS and StereoScan.