UC San Diego

Visual odometry has become an important tool given the new popularity of mobile robotics. Camera pose estimation is a key part of visual odometry and has traditionally been computed by hand-engineered algorithms. Given the recent explosion of deep learning, learned networks are making headway in replacing such algorithms. They have proven to be very successful, and in certain instances are already better traditional methods like ORB-SLAM. However, such methods are usually trained and tested on a relatively homogeneous dataset that contains little variation in lighting, motion, or other cues that can prove challenging. In this work, a novel dataset is curated to specifically introduce more realistic and diverse trajectories. This dataset is a combination of three other datasets: Advio, Aqualoc and Euroc MAV. Using this dataset, it is shown how traditional methods like ORB-SLAM and DSO still out-perform learned, structure from motion methods, and certain failure cases are identified as a ways to improve the learned algorithms. Such a dataset can be used in the future to truly ascertain the strength of a network’s performance.