Active Image-based Modeling

摘要

We seek to automate the data capturing process in image-based modeling, whichis often tedious and time consuming now. At the heart of our system is aniterative linear method to solve the multi-view stereo (MVS) problem quickly.Unlike conventional MVS algorithms that solve a per-pixel depth at each inputimage, we represent the depth map at each image as a piecewise planar trianglemesh and solve it by an iterative linear method. The edges of the triangle meshare snapped to image edges to better capture scene structures. Our fast MVSalgorithm enables online model reconstruction and quality assessment todetermine the next-best-views (NBVs) for modeling. The NBVs are searched in aplane above unreconstructed shapes. In this way, our path planning can use theresult from 3D reconstruction to guarantee obstacle avoidance. We test thissystem with an unmanned aerial vehicle (UAV) in a simulator, an indoor motioncapture system (Vicon) room, and outdoor open spaces.