InstanceVO: Self-Supervised Semantic Visual Odometry by Using Metric Learning to Incorporate Geometrical Priors in Instance Objects

Visual odometry is one of the key technologies for unmanned ground vehicles. To improve the robustness of the systems and enable intelligent tasks, researchers introduced learning-based recognition modules into visual odometry systems, but didn't realize tight coupling between visual odometry systems and recognition modules. This letter proposes a self-supervised semantic visual odometry method, which can complete the tasks of ego-motion estimation, depth prediction, and instance segmentation with a shared encoder. The potential dynamic regions are removed and the image reconstruction loss is rectified by instance detection results. Moreover, the instance-guided triplet loss and cross-task self-attention modules are devised to learn the geometrical relationships among pixels that are implied in instance object priors. The proposed method is validated on KITTI and ComplexUrban datasets. The experimental results show that our method has superiority to baseline models in both pose estimation and depth prediction. We also discuss the efficacy of evaluation metrics for pose estimation, and consider the accumulation errors of trajectories.