TwinVO: Unsupervised Learning of Monocular Visual Odometry using Bi-direction Twin Network

In recent years, more attention has been paid to the use of unsupervised deep learning approaches in Visual Odometry (VO).
In this paper, we present a novel unsupervised learning framework called TwinVO for the estimation of 6-DoF camera poses and monocular depths.
Taking account of the extreme imbalance between forward and backward camera motions in datasets, we provide an innovative twin module to predict bi-direction ego-motions simultaneously.
Meanwhile, motivated by the cooperative game theory, an Inversion Consistency Constraint is suggested to supervise the bi-direction motions so that a final win-win state is achieved.
Furthermore, more delicate structures are adopted in depth estimation network to gain about 37\% the number of parameters reduction as well as achieve better performance.
Extensive experiments on the KITTI dataset reveal that our scheme achieves superior performance and provides better results for both pose and depth estimation.