Pose Refinement: Bridging The Gap Between Unsupervised Learning And Geometric Methods For Visual Odometry

Unsupervised Learning based monocular visual odometry (VO) has lately drawn significant attention owing to its potential in label-free leaning ability and robustness to camera parameters and environmental variations. However, due to the lack of pose optimization or drift correction, there still exist a huge gap between unsupervised learning and geometric approaches. We propose a hybrid VO system which combines an unsupervised monocular VO with a pose refinement back-end, performing optimization in real time, leading to boosted performance. Experiments on the KITTI odometry dataset show that our framework can outperform all unsupervised learning methods and show favorable overall translational accuracy compared to established monocular SLAM systems.