MaskVO: Self-Supervised Visual Odometry with a Learnable Dynamic Mask

Visual odometry, one of the main research topics in robotics, is capable of offering pose information for mobile robotic systems. Recent advances in deep learning allow mobile robots to learn ego-motion and depth maps jointly in a self-supervised manner. However, the existing approaches are confounded by the problem of scale ambiguity and by environmental issues, preventing real-world applications. Our work aims to tackle these two problems by proposing a self-supervised visual odometry model that exploits the temporal dependencies of image sequences and produces scale-consistent motion transformations from a monocular camera. Our proposed framework is integrated with a novel mask network to provide learnable dynamic masks that reduce the influences of scene dynamics and illumination changes. We evaluated our framework against public benchmarks, showing that our MaskVO outperforms existing baselines. Moreover, we also investigated the effectiveness of our proposed dynamic mask network via a detailed ablation study.