WGANVO: monocular visual odometry based on generative adversarial networks

Traditional Visual Odometry (VO) systems, direct or feature-based, are susceptible to matching errors between images. Furthermore, monocular configurations are only capable of estimating localization up to a scale factor, making impossible to use them out-of-the-box in robotics or virtual reality application. Recently, several Computer Vision problems have been successfully tackled by Deep Learning algorithms. In this paper we introduce a Deep Learning-based monocular Visual Odometry system called WGANVO. Specifically, we train a GAN-based neural network to regress a motion estimate. The resulting model receives a pair of images and estimates the relative motion between them. We train the neural network using a semi-supervised approach. In contrast to traditional geometry-based monocular systems, our Deep Learning-based method is able to estimate the absolute scale of the scene without extra information and prior knowledge. We evaluate WGANVO on the well-known KITTI dataset. We show that our system works in real time and the accuracy obtained encourages further development of Deep Learning-based localization systems.