Reconstructing urban scene 3D using VisualSfM

Abstract

New computer vision techniques use photo dataset to rapidly build detailed 3D models. Computer- vision researchers have explored many approaches to city-scale 3D reconstruc on. Among these systems stands out VisualSfM developed by the University of Washington & Google Inc. It is a open source GUI applica on of a Structure from Mo on (SfM) so! ware that uses a feature extractor called Si! GPU and the Mul core Bundle Adjustment. In addi on it embeds the CMVS/ PMVS2 able to reconstruct dense 3D point cloud. Our goal is to demonstrate the metric accuracy of VisualSfM+CMVS/PMVS2 and that to get run it, you can use an unstructured photo dataset but the result improves if you use a structured photo dataset. The approach has been tested on several large datasets with structured images.