Fast robust large-scale mapping from video and internet photo collections

This paper presents a system approaching fully automatic 3D modeling of large-scale environments. Our system takes as input either a video stream or collection of photographs obtained from Internet photo sharing web-sites such as Flickr. The system achieves high computational performance through algorithmic optimizations for efficient robust estimation, the use of image-based recognition for efficient grouping of similar images, and two-stage stereo estimation for video streams that reduces the computational cost while maintaining competitive modeling results. In addition to algorithmic advances, we achieve a major improvement in computational speed through parallelization and execution on commodity graphics hardware. These improvements lead to real-time video processing and to reconstruction from tens of thousands of images within the span of a day on a single commodity computer. We demonstrate modeling results on a variety of real-world video sequences and photo collections.