UnDeepVO: Monocular Visual Odometry through Unsupervised Deep Learning.

We propose a novel monocular visual odometry (VO) system called UnDeepVO inthis paper. UnDeepVO is able to estimate the 6-DoF pose of a monocular cameraand the depth of its view by using deep neural networks. There are two salientfeatures of the proposed UnDeepVO: one is the unsupervised deep learningscheme, and the other is the absolute scale recovery. Specifically, we trainUnDeepVO by using stereo image pairs to recover the scale but test it by usingconsecutive monocular images. Thus, UnDeepVO is a monocular system. The lossfunction defined for training the networks is based on spatial and temporaldense information. A system overview is shown in Fig. 1. The experiments onKITTI dataset show our UnDeepVO achieves good performance in terms of poseaccuracy.