Fast Structure from Motion for Sequential and Wide Area Motion Imagery

Abstract

We present a fast and efficient Structure-from-Motion (SfM) pipeline for refinement of camera parameters and 3D scene reconstruction given initial noisy camera metadata measurements. Examples include aerial Wide Area Motion Imagery (WAMI) which is typically acquired in a circular trajectory and other sequentially ordered multiview stereo imagery like Middlebury, Fountain or body-worn videos. Image frames are assumed (partially) ordered with approximate camera position and orientation information available from (imprecise) IMU and GPS sensors. In the proposed BA4S pipeline the noisy camera parameters or poses are directly used in a fast Bundle Adjustment (BA) optimization. Since the sequential ordering of the cameras is known, consecutive frame-to-frame matching is used to find a set of feature correspondences for the triangulation step of SfM. These putative correspondences are directly used in the BA optimization without any early-stage filtering (i.e. no RANSAC) using a statistical robust error function based on co-visibility, to deal with outliers (mismatches), which significantly speeds up our SfM pipeline by more than 100 times compared to VisualSfM.