Deep Visual Odometry with Self-Supervised Training

In recent years, deep learning-based methods for visual odometry have evolved significantly, and now they outperform classical algorithms on classical benchmarks. However, acquiring ground truth data for training supervised methods require special equipment and thus might be difficult and expensive. To overcome these limitations, a number of unsupervised visual odometry methods have been proposed. To train these methods, ground truth object poses are not required. Instead, they exploit known geometric relations between distance to the scene objects and the motion of the object to estimate the trajectory. However, there is still a large gap in accuracy between supervised and unsupervised methods. In this study, we propose a novel learnable visual odometry method which estimates ego-motion given optical flow. Moreover, we describe how to generate supervision for an optical flow-based visual odometry method. In self-supervised training, labeled data is created automatically without any manual labeling. In this work, we propose to generate labeled training samples using the approximate motion distribution of an object and dense depth maps. The labeled data is created as follows: for an arbitrary input image, the motion is randomly sampled from the motion distribution. Then using the depth map we calculate the optical flow corresponding to this sampled motion. The resulting optical flow serves as an input to the learnable visual odometry method while the ego-motion serves as the target. The experiments on the KITTI Visual Odometry dataset demonstrate that the novel learnable visual odometry method trained in the proposed self-supervised manner outperforms unsupervised visual odometry methods, thereby reducing the quality gap between the methods that do not require supervision and fully-supervised methods. Keywords: learnable visual odometry, self-supervised learning, optical flow, depth maps.