Robustifying Direct VO to Large Baseline Motions

Abstract

While Direct Visual Odometry (VO) methods have been shown to outperform feature-based ones in terms of accuracy and processing time, their optimization is sensitive to the initialization pose typically seeded from heuristic motion models. In real-life applications, the motion of a hand-held or head-mounted camera is predominantly erratic, thereby violating the motion models used, causing large baselines between the initializing pose and the actual pose, which in turn negatively impacts the VO performance.