Visual Odometry based on Semantic Supervision

Traditional Visual Odometry (VO) relies on geometric features, such as points and lines, without considering the high-level semantic information. However, the quality of these features suffers from the changeable external environment. In this paper, we propose a new method to capture significant and stable visual features, aiming to obtaining highly-accurate and robust VO. First, since statistics suggest that landmarks with different semantic labels contribute differently to pose estimation, we incorporate semantic understanding of the scenes to feature selection and rank features by significance. Second, a strategy for feature selection based on the attention region from prior is designed. Last, our proposed algorithm can be easily integrated into existing Visual Odom-etry frameworks. Experiments on KITTI dataset indicate that our method achieves better performance on localization.