An Open-Source Benchmark for Scale-Aware Visual Odometry Algorithms

This paper introduces an open-source benchmark for scale-aware visual odometry (SAVO) algorithms, such as stereo visual odometry (VO) and visual-inertial odometry (VIO). The latest open-source VO algorithms are collected and evaluated with the EuRoC MAV and TUM VI datasets. Although there have been a number of benchmarks for VO, we were the first to make the evaluation system containing algorithm sources publicly available. The algorithms are ORB SLAM2 with stereo inputs, ROVIOLI, VINS-fusion, and SVO2, and the latter two algorithms have variations with different sensor configurations. The evaluation results suggest that ORB-SLAM2 makes the best tracking performance with smooth motion, ROVIOLI is robust to highly dynamic motions, and VINS-fusion and SVO2 have the merits of short processing time. Our benchmark system is available at: https://github.com/goodgodgd/docker-vo-bench.