Unsupervised Learning of Visual Odometry with Depth Warp Constraints

Visual Odometry (VO) is one of the important components of Visual SLAM system. Some impressive work on the end-to-end deep neural networks for 6-DoF VO has appeared. We propose two-part cascade network structure to learn depth from binocular image and to infer ego-motion from consecutive frames. We propose depth warp constraints to make the Network learning more geometrically information. A lot of experiments on KITTI data set show that our model is superior to previous unsupervised methods and has comparable results with the supervised method, verifying that such a depth warp constraints perform successfully in the unsupervised deep method which is an important complement to the geometric method.